FIELD OF THE INVENTION

[0001] The present disclosure relates to methods, apparatus and software for determining a geographic location of an electronic device configured to communicate over a mobile telecommunications network. The present disclosure further relates to apparatus, methods and software for training a prediction model for determining a geographic location of an electronic device configured to communicate over a mobile telecommunications network.

[0002] BACKGROUND

[0003] Mobile telecommunications networks, such as cellular networks, are typically capable of providing network connectivity to a wide range of different electronic devices. Devices capable of communication over a mobile telecommunications network may include useroperated devices such as mobile telephones (including smartphones), tablets, personal computers etc. and may also include connected vehicles (which might include land-borne and/or air-borne vehicles), Machine to Machine (M2M) devices and/or Internet of Things (loT) devices.

[0004] Many electronic devices capable of communicating over a mobile telecommunications network may be portable and thus may be operable at a number of different geographic locations. It is often desirable to determine a geographic location of an electronic device. For example, the determination of a geographic location of a device may be used to track the location of the device, may be used to provide a user of the device with an indication of their current location and/or may be used to facilitate providing navigation instructions to the device or a user of the device.

[0005] One method of determining a location of an electronic device is to use a Global Navigation Satellite System (GNSS) such as the Global Positioning System (GPS). However, in some situations it may be desirable to determine the location of a device without relying on a GNSS. For example, GNSS based location determination may be prone to jamming and or spoofing which could affect the reliability of a location determination.

[0006] As an alternative, as a backup to, or as a supplement to GNSS based location determination, a geographic location of an electronic device may be determined by analysis of measurements of signals (e.g. radio frequency signals) transmitted over a mobile telecommunications network. For example, techniques have been developed which may be referred to as a Radio Positioning System (RPS) in which measurements of radio frequency signals transmitted over a mobile telecommunications network and received by a device are used to derive the location of the device. Such techniques may use a database of previous measurements of radio conditions at known locations to derive the current geographic location of a device based on corresponding measurements of radio signals taken at the current location.

[0007] RPS based techniques may be used for devices situated at ground based locations. Additionally or alternatively, RPS location techniques may be used for devices operating above ground level. For example, RPS location techniques may be used to derive the location of network connected drones which fly at altitude (i.e. above ground level). A network connected drone may be able to fly freely and without the need to maintain radio communication with a single control device. For example, unlike traditional drones, it may not be necessary to maintain visual line of sight between a network connected drone and a ground based control device, since control and communication may be provided through connection to a mobile telecommunications network which operates over a wider area.

[0008] Network connected drones are a particular application in which an alternative to GNSS based location may be desirable. For example, it may be particularly desirable to maintain an accurate record of a drone’s location to ensure that it does not significantly vary from an agreed flight path or does not enter areas in which it is not permitted to fly. As explained above, GNSS based location determination may be vulnerable to spoofing and/or jamming and may not therefore be entirely reliable. RPS based location techniques may be used as an alternative and more secure way to determine the location of a network connected drone. An RPS determined location might for example, be compared to a GNSS based location determination as a check that the GNSS based location determination is not being spoofed.

[0009] It is in this context that the subject matter contained in the present application has been devised.

SUMMARY OF THE INVENTION

[0010] It has been found that the accuracy of the determination of the geographic location of electronic devices can be improved by training one or more prediction models using training data. The accuracy of the determination of the geographic location of electronic devices can be improved through one or more improvements to methods of training a prediction model.

[0011] According to a first aspect of the present disclosure there is provided a computer implemented method of training a prediction model for determining the geographic location of an electronic device configured to communicate over a mobile telecommunications network comprising a plurality of cells, the method comprising: receiving data representative of coverage of the mobile telecommunications network at a plurality of training locations, wherein for each of the plurality of training locations, the data representative of coverage of the mobile telecommunications network at that training location includes an indication of a serving cell of the mobile telecommunications network for that training location, such that each training location is associated with a serving cell; determining location information indicative of a location of the one or more serving cells included in the data representative of coverage of the mobile telecommunications network; forming training data comprising a plurality of training data records, each training data record being associated with a training location of the plurality of training locations and comprising: the training location with which the training data record is associated and the received data representative of the coverage of the mobile telecommunications network at that training location, wherein the training data further comprises the determined location information; and training the prediction model for determining the geographic location of an electronic device using the training data.

[0012] The determining location information indicative of a location of the one or more serving cells included in the data representative of coverage of the mobile telecommunications network may comprise: for each serving cell included in the data representative of coverage of the mobile telecommunications network at the plurality of training locations, determining a serving cell location associated with the serving cell.

[0013] Each training data record may further comprise the determined serving cell location associated with the serving cell for the training location with which the training data record is associated.

[0014] For at least one of the serving cells included in the data representative of coverage of the mobile telecommunications network, the determining a serving cell location associated with the serving cell may comprise grouping the training locations associated with that serving cell into a plurality of sub-groups and determining a serving cell sub-group location for each sub-group of training locations. For at least some of the training locations, the determined serving cell location included in the training data record associated with that training location may comprise a determined serving cell sub-group location for a sub-group of training locations into which that training location is grouped.

[0015] The determining location information indicative of a location of the one or more serving cells included in the data representative of coverage of the mobile telecommunications network may comprise: for each of the plurality of training locations determining, in dependence on the received data, probabilities that each of a plurality of cells are the serving cell for that training location. The location information may comprise an indication of the determined probabilities and the serving cells associated with each probability.

[0016] Determining probabilities that each of a plurality of cells are the serving cell for that training location may comprise: for each of a plurality of reference regions, determining, in dependence on the received data, probabilities that each of a plurality of cells are the serving cell for locations within that reference region, and determining a reference region of the plurality of reference regions within which the training location is situated. The probabilities that each of a plurality of cells are the serving cell for that training location may comprise the determined probabilities that each of a plurality of cells are the serving cell for locations within the determined reference region.

[0017] The plurality of training locations may be situated within the plurality of reference regions.

[0018] The plurality of reference regions may be arranged having substantially uniform separation between centres of adjacent reference regions.

[0019] Each of the formed training data records may be associated with a training location which is associated with the same serving cell.

[0020] The forming training data may comprise determining first training locations of the plurality of training locations which are each associated with the same serving cell and forming training data records for the first training locations.

[0021 ] According to a second aspect of the present disclosure, there is provided a computer implemented method of training prediction models for determining the geographic location of an electronic device configured to communicate over a mobile telecommunications network comprising a plurality of cells, the method comprising: receiving data representative of coverage of the mobile telecommunications network at a plurality of training locations, wherein for each of the plurality of training locations, the data representative of coverage of the mobile telecommunications network at that training location includes an indication of a serving cell of the mobile telecommunications network for that training location; forming training data comprising a plurality of training data records, each training data record being associated with a training location of the plurality of training locations and a serving cell of the mobile telecommunications network for that training location, wherein each training data record comprises: the training location with which the training data record is associated and the received data representative of the coverage of the mobile telecommunications network at that training location; and training a plurality of prediction models for determining the geographic location of an electronic device, wherein training each of the plurality of prediction models comprises: selecting a subset of the training data records associated with the same serving cell of the mobile telecommunications network; and training the prediction model using the selected subset of the training data records, wherein the trained prediction model is associated with the serving cell with which the selected subset of training data records is associated. [0022] Selecting a subset of the training data records associated with the same serving cell of the mobile telecommunications network may comprise: grouping the training data records associated with the same serving cell into a plurality of sub-groups of training data records and selecting a first sub-group of the sub-groups as the selected subset of the training data records associated with the same serving cell, and training the prediction model using the selected subset of the training data records comprises training a first prediction model using the first sub-group of the sub-groups of training data.

[0023] The method may further comprise selecting a second sub-group of the sub-groups of training data records as the subset of the training data records associated with the same serving cell; and training a second prediction model using the second sub-group of the subgroups of training data.

[0024] According to a third aspect of the present disclosure there is provided a computer implemented method of training a prediction model for determining the geographic location of an electronic device configured to communicate over a mobile telecommunications network comprising a plurality of cells, the method comprising: receiving first data representative of coverage of the mobile telecommunications network at a first plurality of measurement locations, wherein the first data is based on measurements made at the first plurality of measurement locations; generating second data representative of coverage of the mobile telecommunications network at a second plurality of training locations, wherein the second data is based on the first data and wherein the second plurality of training locations are more evenly geographically distributed than the first plurality of measurement locations; forming training data comprising a plurality of training data records, each training data record being associated with a training location of the second plurality of training locations and comprising: the training location with which the training data record is associated and the generated second data representative of the coverage of the mobile telecommunications network at that training location; and training the prediction model for determining the geographic location of an electronic device using the plurality of training data records.

[0025] The generating second data may comprise: interpolating the first data representative of coverage of the network at the first plurality of measurement locations to determine second data representative of coverage of the network at one or more training locations of the second plurality of training locations.

[0026] The generating second data may comprise: including first data representative of coverage of the network at a first measurement location of the first plurality of measurement locations a plurality of times in the second data at a training location corresponding to the first measurement location. [0027] The first measurement location may be situated in a region for which a spatial density of measurement locations included in the first plurality of measurement locations is low relative to other regions covered by the first plurality of measurement locations.

[0028] The generating second data may comprise: omitting first data representative of coverage of the network at a second measurement location of the first plurality of measurement locations from the second data.

[0029] The second measurement location may be situated in a region for which a spatial density of measurement locations included in the first plurality of measurement locations is high relative to other regions covered by the first plurality of measurement locations.

[0030] According to a fourth aspect of the present disclosure there is provided a computer implemented method of training a prediction model for determining the geographic location of an electronic device configured to communicate over a mobile telecommunications network comprising a plurality of cells, the method comprising: receiving data representative of coverage of the mobile telecommunications network at a plurality of training locations, wherein for each of the plurality of training locations, the data representative of coverage of the mobile telecommunications network at that training location includes a measure of a propagation time of signals exchanged between the training location and a base station of the mobile telecommunications network; determining a subset of the plurality training locations for which the measure of a propagation time is less than a threshold propagation time measure; forming training data comprising a plurality of training data records, each training data record being associated with a training location of the determined subset of plurality of training locations and comprising: the training location with which the training data record is associated and the received data representative of the coverage of the mobile telecommunications network at that training location; and training the prediction model for determining the geographic location of an electronic device using the plurality of training data records.

[0031] For each of the plurality of training locations, the received data representative of coverage of the mobile telecommunications network at that training location may include an indication of a serving cell of the mobile telecommunications network for that training location, such that each training location is associated with a serving cell. Each of the formed training data records may be associated with a training location which is associated with the same serving cell.

[0032] The forming training data may comprise determining first training locations of the plurality of training locations which are each associated with the same serving cell and forming training data records for the first training locations.

[0033] For each of the plurality of training locations, the received data representative of coverage of the mobile telecommunications network at that training location may include an indication of a serving cell of the mobile telecommunications network for that training location, such that each training location is associated with a serving cell. The method may further comprise: determining location information indicative of a location of the one or more serving cells included in the data representative of coverage of the mobile telecommunications network, wherein the training data further comprises the determined location information.

[0034] For each of the plurality of training locations, the data representative of coverage of the mobile telecommunications network at that training location may include a measure of a propagation time of signals exchanged between the training location and a base station of the mobile telecommunications network. The forming training data may comprise determining a subset of the plurality of training locations for which the measure of propagation time is less than a threshold propagation time and forming training data records for the determined subset of training locations.

[0035] A method as described above may further comprise: receiving first data representative of coverage of the mobile telecommunications network at a first plurality of measurement locations, wherein the first data is based on measurements made at the first plurality of measurement locations; and generating second data representative of coverage of the mobile telecommunications network at a second plurality of training locations, wherein the second data is based on the first data and wherein the second plurality of training locations are more evenly geographically distributed than the first plurality of measurement locations. The receiving data representative of coverage of the mobile telecommunications network at a plurality of training locations may comprise receiving the generated second data representative of coverage of the mobile telecommunications network at the second plurality of training locations.

[0036] The received data representative of coverage of the mobile telecommunications network at a plurality of training locations may comprise one or more coverage properties determined for each of the plurality of training locations. The one or more coverage properties may comprise at least one of a received signal power, a received signal quality and/or a measure of a propagation time of signals exchanged between the training location and a base station of the mobile telecommunications network.

[0037] The one or more coverage properties for each of the plurality of training locations may include one or more coverage properties determined for a serving cell at each of the plurality of training locations.

[0038] The one or more coverage properties for each of the plurality of training locations may include one or more coverage properties determined for one or more neighbouring cells at each of the plurality of training locations.

[0039] The plurality of training locations may include one or more training locations situated above ground. [0040] According to a fifth aspect of the present disclosure there is provided a computer implemented method of determining the geographic location of an electronic device configured to communicate over a mobile telecommunications network comprising a plurality of cells, the method comprising: obtaining data representative of coverage of the mobile telecommunications network at the location of the electronic device, wherein the data representative of coverage of the mobile telecommunications network is based on measurements made by the electronic device at the location of the electronic device and wherein the data includes an indication of a serving cell of the mobile telecommunications network at the location of the electronic device; providing the obtained data as an input to a prediction model, configured through training to determine a location of an electronic device in dependence on data representative of coverage of the mobile telecommunications network; and implementing the prediction model to determine the geographic location of the electronic device in dependence on the provided inputs, wherein the prediction model is configured through training based on training data comprising a plurality of training data records, wherein each training data record is associated with a training location of a plurality of training locations and comprises: the training location with which the training data record is associated and data representative of the coverage of the mobile telecommunications network at that training location, and wherein the training data further comprises location information indicative of a location of a serving cell for that training location.

[0041] According to a sixth aspect of the present disclosure there is provided a computer implemented method of determining the geographic location of an electronic device configured to communicate over a mobile telecommunications network comprising a plurality of cells, the method comprising: obtaining data representative of coverage of the mobile telecommunications network at the location of the electronic device, wherein the data representative of coverage of the mobile telecommunications network is based on measurements made by the electronic device at the location of the electronic device and wherein the data includes an indication of a serving cell of the mobile telecommunications network at the location of the electronic device; selecting a prediction model from a plurality of prediction models configured through training to determine a location of an electronic device in dependence on data representative of coverage of the mobile telecommunications network, wherein each of the plurality of prediction models is associated with a serving cell of the mobile telecommunications network and wherein selecting the prediction model comprises selecting a prediction model which is associated with the indicated serving cell of the mobile telecommunications network at the location of the electronic device; providing the obtained data as an input to the selected prediction model; and implementing the prediction model to determine the geographic location of the electronic device in dependence on the provided inputs. [0042] According to seventh aspect of the present disclosure there is provided a computer implemented method of determining the geographic location of an electronic device configured to communicate over a mobile telecommunications network comprising a plurality of cells, the method comprising: obtaining data representative of coverage of the mobile telecommunications network at the location of the electronic device, wherein the data representative of coverage of the mobile telecommunications network is based on measurements made by the electronic device at the location of the electronic device; providing the obtained data as an input to a prediction model, wherein the prediction model is configured through training according to a method according to any of the first to fourth aspects; and implementing the prediction model to determine the geographic location of the electronic device in dependence on the provided inputs.

[0043] According to an eighth aspect of the present disclosure there is provided a computer program comprising instructions which, when executed, cause the method of any of the first to seventh aspects to be implemented.

[0044] According to a ninth aspect of the present disclosure there is provided apparatus for training a prediction model for training a prediction model for determining the geographic location of an electronic device configured to communicate over a mobile telecommunications network comprising a plurality of cells, the apparatus comprising: one or more processors; and memory storing instructions which, when executed by the one or more processors, cause the apparatus to: receive data representative of coverage of the mobile telecommunications network at a plurality of training locations, wherein for each of the plurality of training locations, the data representative of coverage of the mobile telecommunications network at that training location includes an indication of a serving cell of the mobile telecommunications network for that training location, such that each training location is associated with a serving cell; determine location information indicative of a location of the one or more serving cells included in the data representative of coverage of the mobile telecommunications network; form training data comprising a plurality of training data records, each training data record being associated with a training location of the plurality of training locations and comprising: the training location with which the training data record is associated and the received data representative of the coverage of the mobile telecommunications network at that training location, wherein the training data further comprises the determined location information; and train the prediction model for determining the geographic location of an electronic device using the training data. [0045] According to a tenth aspect of the present disclosure there is provided apparatus for training prediction models for determining the geographic location of an electronic device configured to communicate over a mobile telecommunications network comprising a plurality of cells, the apparatus comprising: one or more processors; and memory storing instructions which, when executed by the one or more processors, cause the apparatus to: receive data representative of coverage of the mobile telecommunications network at a plurality of training locations, wherein for each of the plurality of training locations, the data representative of coverage of the mobile telecommunications network at that training location includes an indication of a serving cell of the mobile telecommunications network for that training location; form training data comprising a plurality of training data records, each training data record being associated with a training location of the plurality of training locations and a serving cell of the mobile telecommunications network for that training location, wherein each training data record comprises: the training location with which the training data record is associated and the received data representative of the coverage of the mobile telecommunications network at that training location; and train a plurality of prediction models for determining the geographic location of an electronic device, wherein training each of the plurality of prediction models comprises: selecting a subset of the training data records associated with the same serving cell of the mobile telecommunications network; and training the prediction model using the selected subset of the training data records, wherein the trained prediction model is associated with the serving cell with which the selected subset of training data records is associated.

[0046] According to an eleventh aspect of the present disclosure there is provided apparatus for training a prediction model for determining the geographic location of an electronic device configured to communicate over a mobile telecommunications network comprising a plurality of cells, the apparatus comprising: one or more processors; and memory storing instructions which, when executed by the one or more processors, cause the apparatus to: receive first data representative of coverage of the mobile telecommunications network at a first plurality of measurement locations, wherein the first data is based on measurements made at the first plurality of measurement locations; generate second data representative of coverage of the mobile telecommunications network at a second plurality of training locations, wherein the second data is based on the first data and wherein the second plurality of training locations are more evenly geographically distributed than the first plurality of measurement locations; form training data comprising a plurality of training data records, each training data record being associated with a training location of the second plurality of training locations and comprising: the training location with which the training data record is associated and the generated second data representative of the coverage of the mobile telecommunications network at that training location; and train the prediction model for determining the geographic location of an electronic device using the plurality of training data records.

[0047] According to a twelfth aspect of the present disclosure there is provided apparatus for training a prediction model for determining the geographic location of an electronic device configured to communicate over a mobile telecommunications network comprising a plurality of cells, the apparatus comprising: one or more processors; and memory storing instructions which, when executed by the one or more processors, cause the apparatus to: receive data representative of coverage of the mobile telecommunications network at a plurality of training locations, wherein for each of the plurality of training locations, the data representative of coverage of the mobile telecommunications network at that training location includes a measure of a propagation time of signals exchanged between the training location and a base station of the mobile telecommunications network; determine a subset of the plurality training locations for which the measure of a propagation time is less than a threshold propagation time measure; form training data comprising a plurality of training data records, each training data record being associated with a training location of the determined subset of plurality of training locations and comprising: the training location with which the training data record is associated and the received data representative of the coverage of the mobile telecommunications network at that training location; and train the prediction model for determining the geographic location of an electronic device using the plurality of training data records.

[0048] According to a thirteenth aspect of the present disclosure there is provided apparatus for determining the geographic location of an electronic device configured to communicate over a mobile telecommunications network comprising a plurality of cells, the apparatus comprising: one or more processors; and memory storing instructions which, when executed by the one or more processors, cause the apparatus to: obtain data representative of coverage of the mobile telecommunications network at the location of the electronic device, wherein the data representative of coverage of the mobile telecommunications network is based on measurements made by the electronic device at the location of the electronic device and wherein the data includes an indication of a serving cell of the mobile telecommunications network at the location of the electronic device; provide the obtained data as an input to a prediction model, configured through training to determine a location of an electronic device in dependence on data representative of coverage of the mobile telecommunications network; and implement the prediction model to determine the geographic location of the electronic device in dependence on the provided inputs. The prediction model is configured through training based on training data comprising a plurality of training data records, wherein each training data record is associated with a training location of a plurality of training locations and comprises: the training location with which the training data record is associated and data representative of the coverage of the mobile telecommunications network at that training location, and wherein the training data further comprises location information indicative of a location of a serving cell for that training location.

[0049] According to a fourteenth aspect of the present disclosure there is provided apparatus for determining the geographic location of an electronic device configured to communicate over a mobile telecommunications network comprising a plurality of cells, the apparatus comprising: one or more processors; and memory storing instructions which, when executed by the one or more processors, cause the apparatus to: obtain data representative of coverage of the mobile telecommunications network at the location of the electronic device, wherein the data representative of coverage of the mobile telecommunications network is based on measurements made by the electronic device at the location of the electronic device and wherein the data includes an indication of a serving cell of the mobile telecommunications network at the location of the electronic device; select a prediction model from a plurality of prediction models configured through training to determine a location of an electronic device in dependence on data representative of coverage of the mobile telecommunications network, wherein each of the plurality of prediction models is associated with a serving cell of the mobile telecommunications network and wherein selecting the prediction model comprises selecting a prediction model which is associated with the indicated serving cell of the mobile telecommunications network at the location of the electronic device; provide the obtained data as an input to the selected prediction model; and implement the prediction model to determine the geographic location of the electronic device in dependence on the provided inputs.

[0050] According to a fifteenth aspect of the present disclosure there is provided apparatus for determining the geographic location of an electronic device configured to communicate over a mobile telecommunications network comprising a plurality of cells, the apparatus comprising: one or more processors; and memory storing instructions which, when executed by the one or more processors, cause the apparatus to: obtain data representative of coverage of the mobile telecommunications network at the location of the electronic device, wherein the data representative of coverage of the mobile telecommunications network is based on measurements made by the electronic device at the location of the electronic device; provide the obtained data as an input to a prediction model, wherein the prediction model is configured through training according to a method according to any of the first to fourth aspects; and implement the prediction model to determine the geographic location of the electronic device in dependence on the provided inputs.

[0051] Within the scope of this application it is expressly intended that the various aspects, embodiments, examples and alternatives set out in the preceding paragraphs, in the claims and/or in the following description and drawings, and in particular the individual features thereof, may be taken independently or in any combination. That is, all examples and/or features of any example can be combined in any way and/or combination, unless such features are incompatible. The applicant reserves the right to change any originally filed claim or file any new claim accordingly, including the right to amend any originally filed claim to depend from and/or incorporate any feature of any other claim although not originally claimed in that manner. BRIEF DESCRIPTION OF FIGURES

[0052] One or more embodiments of the invention are shown schematically, by way of example only, in the accompanying drawings, in which:

Figure 1 is a schematic illustration of a section of an environment in which a mobile telecommunications network may operate;

Figure 2 is a flow chart of an example method for training a prediction model for determining the geographic location of a terminal based on data representative of coverage of a mobile telecommunications network;

Figure 3 is a schematic illustration of geographic locations in a section of an environment in which a mobile telecommunications network may operate;

Figure 4 is a table showing example properties which may be included in a plurality of training data records used in a method for training a prediction model;

Figure 5 is a flow chart of an example method for determining a geographic location of an electronic device using a trained prediction model;

Figure 6 is a table showing an example of the properties which may be included in data obtained according to the method of Figure 5;

Figure 7 is a flow chart of a further example method for training a prediction model for determining above ground coverage of a mobile telecommunications network;

Figure 8 is a schematic illustration of training locations in a section of an environment in which a mobile telecommunications network may operate;

Figure 9 is a table showing example properties which may be included in a plurality of training data records formed according to the method of Figure 7;

Figure 10 is a table showing further example properties which may be included in a plurality of training data records according to methods described herein;

Figure 11 is a schematic illustration of a further section of an environment in which a mobile telecommunications network may operate;

Figure 12 is a table showing example properties which may be included in two training data records formed according the method of Figure 7;

Figure 13 is a table showing an example of the properties which may be included in the data obtained for providing inputs to a trained prediction model for determining the geographic location of a device; Figure 14 is a flow chart of an example method for training a prediction model for determining the geographic location of a device based on data representative of coverage of a mobile telecommunications network;

Figure 15 is a flow chart of an example method for determining a geographic location of an electronic device using a trained prediction model selected from a plurality of trained prediction models;

Figure 16 is a flow chart of an example method for training a prediction model for determining the geographic location of a terminal based on data representative of coverage of a mobile telecommunications network;

Figure 17 is a schematic depiction of an example distribution of training locations in a cell of a mobile telecommunications network;

Figure 18 is a flow chart of an example method for training a prediction model for determining the geographic location of a terminal based on data representative of coverage of a mobile telecommunications network; and

Figure 19 is a schematic illustration of an example electronic device which may be used to implement all or part of any method described herein.

DETAILED DESCRIPTION

[0053] Before particular examples of the present invention are described, it is to be understood that the present disclosure is not limited to the particular examples described herein. It is also to be understood that the terminology used herein is used for describing particular examples only and is not intended to limit the scope of the claims.

[0054] In describing and claiming the apparatus and methods of the present invention, the following terminology will be used: the singular forms "a", "an", and "the" include plural forms unless the context clearly dictates otherwise. Thus, for example, reference to "a terminal" includes reference to one or more of such elements.

[0055] Figure 1 is a schematic illustration of a section of an environment in which a mobile telecommunications network may operate. The mobile telecommunications network includes a plurality of base stations 101 including a first base station 101 a, a second base station 101b, a third base station 101c and a fourth base stations 101d. The base stations 101 are configured to transmit and receive communication signals over an air interface. For example, each base station 101 may comprise at least one antenna configured to exchange communications (e.g. radio frequency signals) with devices (e.g. terminals) situated within a geographical coverage area 102 (which may be referred to as a cell) serviced by the base station 101 over an air interface. [0056] Each base station 101 may exchange communications by transmitting and/or receiving communications in one or more frequency bands assigned to a Radio Access Technology (RAT) used by the base station 101 and utilising communication protocols specified for the RAT (e.g. standardised communication protocols for the RAT). Suitable RATs may include, for example, the Global System for Mobile Communications (GSM), the Universal Mobile Telecommunications System (UMTS), Long Term Evolution (LTE) and/or 5G New Radio (NR). The base stations 101 may take any suitable form and may, for example, comprise a GMS and/or UMTS compatible base station such as a Node B, an Evolved NodeB (eNB) and/or a 5G NR gNodeB. The base stations 101 typically have a backhaul connection with one or more core networks (not shown) with which users of the telecommunications network are registered.

[0057] Each base station 101 may have at least one geographical coverage area 102 over which it can reliably communicate with terminals 104 situated within the geographical coverage area 102. Such a geographical coverage area may be referred to as a cell 102. In the depiction shown in Figure 1 , a first cell 102a is associated with the first base station 101a, a second cell 102b is associated with the second base station 101 b, a third cell 102c is associated with the third base station 101c and a fourth cell 102d is associated with the fourth base station 101 d. In the simple depiction shown in Figure 1 , each base station 101 provides coverage to a single cell 102. Whilst not shown in Figure 1 , in at least some examples, a single base station 101 may transmit and receive in a plurality of cells. For example, a base station 101 may simultaneously operate a plurality of antennas which serve different geographical coverage areas. Such a base station 101 may be considered to operate a plurality of different cells 102.

[0058] A cell 102 associated with a base station 101 may be geographically separate from a cell 102 associated with other neighbouring base stations 101 and/or another cell operated by the same base station 101 . For example, in the simplified depiction shown in Figure 1 the geographical extent of each cell does not overlap with any other neighbouring cell 102. Alternatively there may be some geographic overlap between different cells 102 operated by the same or different base stations 101. A given terminal 104 may be situated within the geographic coverage of a single cell, multiple cells or may be situated in an area where no network coverage is provided (i.e. the terminal is not situated within the coverage area of any cell). In the depiction shown in Figure 1 , a first terminal 104a is situated within the third cell 102c and a second terminal 104b is situated within the second cell 102b.

[0059] The term terminal is used herein to refer to any suitable electronic device capable of connecting to or otherwise communicating over a mobile telecommunications network. The terms terminal and device may be used interchangeably herein. Suitable examples of a terminal 104 as referred to herein may include User Equipment devices (UEs) such as mobile telephones, tablets, personal computers etc. and/or other forms of terminal device which may not be directly used by a user. For example, terminals 104 which connect to and communicate over the mobile telecommunications network may or may not include a user interface which allows for direct user interaction with the terminal 104. In some examples, a terminal 104 may be included in or otherwise attached to a vehicle. The vehicle may be a ground based vehicle such as an automobile and/or may be an airborne vehicle such as an unmanned aerial vehicle (UAV) which is commonly referred to as a drone. A drone including a terminal 104 for communication over a mobile telecommunications network is referred to herein as a network connected drone.

[0060] As will be explained in further detail herein, terminals 104 which are in communication with a telecommunications network may make measurements which are representative of the coverage provided by the telecommunications network. Such measurements may be used to train a prediction model for determining the geographic location of a terminal 104. Furthermore, such measurements may be used to determine the geographic location of a terminal 104 based on the measurements.

[0061] During normal operation of a mobile telecommunications network, base stations 101 may routinely transmit reference signals for the purpose of measurement of the reference signal by a terminal 104. A terminal 104 may make measurements of the reference signal and may, for example, determine one or more variables indicative of measurement of the reference signal. For example, a terminal 104 may determine a measure of the power of one or more reference signals received at the terminal 104. A typical example of such a measure is the reference signal received power (RSRP), which may, for example, be determined by terminals operating according to LTE protocols. More specifically, the RSRP may be taken as an average power per resource element that a terminal 104 is receiving on. Additionally or alternatively, a terminal 104 may determine a measure of the quality of one or more reference signals received at the terminal 104. A typical example of such a measure is the reference signal received quality (RSRQ), which may, for example, be determined by terminals operating according to LTE protocols. More specifically, the RSRQ may be taken as a signal-to- interference plus noise ratio of one or more received reference signals.

[0062] Additionally or alternatively, a base station 101 and/or a terminal 104 may determine a measure of a propagation time associated with signals exchanged between the base station 101 and/or the terminal 104. Such a propagation time is generally at least a function of the distance between the terminal 104 and the base station 101. A typical measure of a propagation time between a base station 101 and a terminal 104 is a timing advance. A timing advance associated with a base station 101 and a terminal 104 may be determined and used to transmit a signal from one communicating party (e.g. a terminal 104 or a base station 103) in advance of a timeslot allocated to reception of the signal at the other communicating party (e.g. the other of the terminal 104 or the base station 101 ). Similarly to the received signal power and/or received signal quality, a measure of propagation time (e.g. timing advance) associated with a terminal 104 may be a measure which is routinely determined during operation of a terminal 104 in a mobile telecommunications network. A measure of a propagation time (e.g. timing advance) determined for a terminal 101 may be associated with a particular base station 103 and/or cell 102. In some examples, for a given terminal 104 a plurality of different measures of propagation time (e.g. timing advance), each associated with a different base station 101 and/or cell, may be determined.

[0063] Measurements such as a received signal power (e.g. RSRP), a received signal quality (e.g. RSRQ) and/or a measure of propagation time (e.g. a timing advance) may be utilised by a terminal 104 for a number of different purposes such as cell selection, resource allocation, determining a power with which to transmit signals and/or synchronisation between a terminal 104 and a base station 101. For example, a terminal 104 may receive reference signals transmitted over a plurality of different cells. The terminal may measure received reference signals and determine one or more properties such as received signal power (e.g. RSRP) and/or received signal quality (e.g. RSRQ) of reference signals of a plurality of different cells. Such properties may be used, during routine operation of a terminal, to select a cell over which to communicate with the network. For example, a cell having an RSRP and/or RSRQ exceeding given thresholds may be selected by a terminal as the terminal’s serving cell. Determined properties such as a measure of propagation time (e.g. a timing advance) may be used to determine times at which to transmit and/or receive signals for communication between a base station 101 and a terminal 104 to ensure synchronisation between the base station 101 and the terminal 104. Additionally or alternatively a determined property such as a measure of propagation time (e.g. a timing advance) may be used to determine a power with which to transmit signals between a base station 101 and a terminal 104.

[0064] A terminal 104 operating in the network may select a cell with which its main connection to the network is established. Such a cell may be considered to be a terminal’s serving cell. In some examples, a terminal 104 may have a plurality of serving cells. For example, where a terminal 104 is configured for carrier aggregation, a terminal’s serving cells may include a primary cell and/or one or more secondary cells.

[0065] A received signal power (e.g. RSRP), a received signal quality (RSRQ) and/or a measure of propagation time (e.g. a timing advance) determined for a terminal 104 may be associated with a particular base station 101 and/or cell 102. In some examples, for a given terminal 104 a plurality of different received signal powers, received signal qualities and/or measures of propagation time may be determined. Each determined received signal power, received signal quality and/or propagation time may be associated with a different cell. For example, a terminal 104 may determine a first received signal power and/or a first received signal quality based on a reference signal transmitted over a first cell (e.g. the first cell 102a shown in Figure 1 ) and may determine a second received signal power and/or a second received signal quality based on a reference signal transmitted over a second cell (e.g. the second cell 102b shown in Figure 1). Additionally or alternatively, a first measure of propagation time (e.g. a timing advance) may be determined for a terminal 104 and the first cell 102a and a second measure of propagation time (e.g. a timing advance) may be determined for the terminal 104 and the second cell 102b. However, in some examples a measure of propagation time (e.g. a timing advance) may only be available for a serving cell. The first cell 102a and the second cell 102b may be operated by the same or different base stations 101 (e.g. the first cell 102a may be operated by a first base station 101a and the second cell 102b may be operated by a second base station 101 b as shown in Figure 1 ).

[0066] Properties such as a received signal power (e.g. RSRP), a received signal quality (e.g. RSRQ) and/or a measure of propagation time (e.g. a timing advance) are examples of properties representative of the coverage of a mobile telecommunications network. Such properties may be referred to herein as coverage properties.

[0067] As was described above, a terminal 104 and/or a base station 101 operating in a mobile telecommunications network may measure or otherwise determine one or more properties (such as received signal power, received signal quality and/or propagation time) associated with the coverage provided by the network. Such determined properties may be associated with a geographic location at which the terminal 104 is situated. For example, measured coverage properties such as a received signal power, received signal quality and/or measure or propagation time for a given terminal 104 may vary as a function of location if the terminal changes its geographic location.

[0068] As further described above, measured coverage properties such as a received signal power, received signal quality and/or propagation time for a given terminal 104 are typically specific to a given cell 102. Measured coverage properties (such as a received signal power, received signal quality and/or propagation time) for a given terminal may be determined for a plurality of cells 102. For example, the first terminal 104a shown in Figure 1 is situated within the geographic coverage area of the third cell 102c and the third cell 102c may act as the serving cell for the first terminal 104a in its depicted location. Coverage properties such as a received signal power, received signal quality and/or propagation time associated with the first terminal and the third cell 101c may be measured or otherwise determined. For example, the first terminal 104a may measure reference signals transmitted by the third base station 101c over the third cell 102c and may determine coverage properties (e.g. received signal power, and/or received signal quality) associated with the third cell 102c based on the measurements. Additionally or alternatively measurements of signals exchanged between the first terminal 104a and third base station 101 c may be used to determine a measure of propagation time associated with the third cell 102c and the first terminal.

[0069] Coverage properties such as a received signal power, received signal quality and/or propagation time associated with the first terminal may additionally or alternatively be measured or otherwise determined for one or more neighbouring cells, such as the first cell 102a, the second cell 102b and/or the fourth cell 102d (and/or other neighbouring cells not shown in Figure 1). Coverage properties (e.g. received signal power, received signal quality and/or propagation time) associated with a plurality of cells, which may include a serving cell 102c and one or more neighbouring cells 102a, 102b, 102d may therefore be determined for the first terminal 104a.

[0070] Similarly, coverage properties (e.g. received signal power, received signal quality and/or propagation time) associated with a plurality of cells may be determined for the second terminal 104b. The plurality of cells may include the second cell 102b acting as a serving cell for the second terminal 104b and one or more neighbouring cells, which may include the first cell 102a, the third cell 102c, the fourth cell 102d and/or one or more other neighbouring cells not shown in Figure 1 .

[0071] Figure 2 is a flow chart of an example method for training a prediction model for determining the geographic location of a terminal based on data representative of coverage of a mobile telecommunications network. The method may be implemented on any suitable computing device. In some examples, each method step may be implemented on the same computing device. In other examples, different parts of the method may be implemented on different computing devices which may be in communication with each other.

[0072] Figure 3 is a schematic illustration of geographic locations in a section of an environment in which a mobile telecommunications network may operate, where the geographic locations may be used in an example of the method of Figure 2. The geographic locations 105 depicted in Figure 3 may be referred to as training locations. The section of the environment shown in Figure 3 is the same as that depicted in Figure 1 and the same components are labelled with the same reference numerals in Figures 1 and 3. No detailed description of the base stations 101 and cells 102 included in the depicted section of the environment will be provided with reference to Figure 3.

[0073] In addition to the features depicted in Figure 1 , Figure 3 also includes a depiction of a plurality of locations 105 at which data representative of coverage of the mobile telecommunications network may be available. For ease of illustration only some of the locations indicated by black circles in Figure s are labelled 105. However, it will be appreciated that each black circle shown in Figure 3 represents an example location 105 at which data representative of coverage of the mobile telecommunications network may be available.

[0074] At step 201 of the method of Figure 2, data representative of coverage of a mobile telecommunications network at a plurality of training locations is received. An example of a plurality of different training locations 105 is depicted in Figure 3.

[0075] As was explained above with reference to Figure 1 , data representative of coverage of a mobile telecommunications network may be measured or otherwise determined by terminals 104 operating in the network. For example, terminals 104 may determine one or more properties such as a received signal power (e.g. RSRP), a received signal quality (e.g. RSRQ) and/or a measure of propagation time (e.g. a timing advance). The plurality of training locations 105 may represent geographic locations at which coverage properties (e,g, received signal power, received signal quality and/or a measure of propagation time) have been determined. Terminals 104 may report such determined properties associated with the network coverage at the plurality of training locations 105 to the network, for example, to a base station 101 and/or a core network (e.g. via a base station and a backhaul connection). The network (or a node of the network) may therefore receive properties associated with network coverage at a plurality of different training locations 105. Receiving the data in step 201 may take any suitable form. For example, receiving the data may comprise reading the data from memory and/or receiving the data from another device at which the data is stored.

[0076] The training locations 105 at which coverage properties are available may represent locations at which the coverage related properties (e.g. received signal power, received signal quality and/or propagation time) have been measured directly by one or more terminals 104 whilst situated at that location. Additionally or alternatively, the locations 105 may include one or more locations at which coverage related properties are derived based on measurements made at other locations (e.g. at a base station receiving signals transmitted from a location 105). Coverage properties at the plurality of training locations 105 may be determined by different terminals 104 situated at different locations. In some examples, a given terminal 104 may move between different locations 105 and may determine one or more properties indicative of network coverage at a plurality of different training locations 105.

[0077] The training locations 105 are shown in Figure 2 in only two dimensions, which might for example represent different latitudes or longitudes. Whilst not shown in Figure 2 at least some of the training locations 105 may be situated at different altitudes. For example, at least some of the training locations 105 may represent above ground locations which may be accessed by a terminal situated on an airborne platform such as a drone (which may form a network connected drone). For example, one or more drones including a terminal device may be flown for a flight dedicated to measuring coverage properties (e.g. received signal power, received signal quality and/or propagation time) at one or more above ground locations 105. Additionally or alternatively, at least some of the training locations 105 may represent ground based locations. Coverage properties may be collected at ground based locations during normal operation of one or more terminals at ground based locations.

[0078] The data received at step 201 of Figure 2 may include one or more properties representative of coverage of the mobile telecommunications network (coverage properties) at each of the plurality of training locations 105. As was described above, the one or more coverage properties may include (but are not limited to) properties such as a received signal power (e.g. RSRP), a received signal quality (e.g. RSRQ) and/or a measure of propagation time (e.g. a timing advance). The one or more coverage properties may be associated with a particular cell 102 and/or base station 101. In some examples, one or more of the coverage properties may be determined for a plurality of different cells 102 or base stations 101 . The determined coverage properties may be associated in the received data with an identifier of the cell or base station with which it is associated. For example, a determined property (e.g. a received signal power, a received signal quality and/or a propagation time) may be associated with a Physical Cell Identifier (PCI) of the cell with which the property is associated.

[0079] The data representative of coverage of a mobile telecommunications network at a plurality of training locations 105 may, for at least some of the training locations 105, include data representative of network coverage provided by a plurality of cells 102. For example, for at least some training locations 105, the data may include one or more properties indicative of network coverage provided by a serving or primary cell. For at least some training locations 105 the data may include one or more properties indicative of network coverage provided by one or more additional cells 102, which may be neighbouring cells (e.g. to a serving or primary cell).

[0080] At step 202 of Figure 2, training data comprising a plurality of training records is formed. Each training data record is associated with a training location 105. Each training data record may comprise at least the received data representative of the coverage of the mobile telecommunications network (e.g. the data received at step 201 ) at the training location 105 with which the training record is associated and the training location 105 itself. For example, a given training data record for a training location 105 may include coverage properties such as a received signal power (e.g. RSRP), a received signal quality (e.g. RSRQ) and/or a measure of propagation time (e.g. a timing advance) associated with a serving cell at the training location 105 and an identifier (e.g. PCI) of the serving cell. The given training data record may further include coverage properties such as a received signal power (e.g. RSRP), a received signal quality (e.g. RSRQ) and/or a measure of propagation time (e.g. a timing advance) associated with one or more neighbouring cells and identifiers (e.g. PCIs) of the one or more neighbouring cells. The given training data record may further include the geographic location of the training location 105, for example, in the form of the latitude, longitude and/or altitude of the training location 105.

[0081] Figure 4 is a table showing example properties which may be included in a plurality of training data records. Each row in the table of Figure 4 represents a different training data record associated with a different training location 105. The training data records may include M training data records associated with M training locations 105. Each column in the table of Figure 4 represents a different field included in a training data record. In the conventions used in Figure 4 the field Lat_m represents the latitude of the training location 105 associated with the mth training data record, where m is an index running from 1 to M. Long_m represents the longitude of the training location 105 associated with the mth training data record. Alt_m represents the altitude of the training location 105 associated with the mth training data record. PCI_sc represents the PCI of the serving cell at the training location 105 associated with the training data record. RSRP_sc is the RSRP associated with the serving cell at the training location associated with the training data record. RSRQ_sc is the RSRQ associated with the serving cell at the training location associated with the training data record. TA_sc is the timing advance associated with the serving cell at the training location associated with the training data record. PCI_ncx represents the PCI of the xth neighbouring cell at the training location 105 associated with the training data record, where x is an index running from 1 to X. RSRP_ncx is the RSRP associated with the xth neighbouring cell at the training location associated with the training data record. RSRQ_ncx is the RSRQ associated with the xth neighbouring cell at the training location associated with the training data record. In some examples, the training data records may also include a timing advance associated with one or more neighbouring cells (not shown in Figure 4). However, in at least some examples, a timing advance may only be available for a serving cell (TA_sc). The number X of neighbouring cells for which coverage properties are included may be the same for each training location or may be different for at least some of the training locations.

[0082] The training data records may be considered to comprise at least one input field and at least one output field. The at least one output field represents the desired output of a model trained using the training data. The at least one input field represents inputs to be provided to a trained model in order to determine the output of the model. In accordance with examples contemplated herein, the at least one output field of each training data record comprises at least one field (e.g. latitude, longitude and/or altitude) associated with the geographic location of the training location 105 with which the training data record is associated. The at least one input field of each training data record comprises the data representative of network coverage at the training location 105 with which the training data record is associated.

[0083] Returning again to Figure 2, at step 203 of Figure 2, a prediction model is trained using the training data formed at step 202. The prediction model is trained for determining a geographic location of an electronic device. The prediction model may comprise a machine learning model. The training of the prediction model may comprise applying a supervised machine learning training algorithm to train the machine learning model.

[0084] As will be appreciated by those of ordinary skill in the art, supervised learning of a prediction model involves training the model to map an input to an output based on training data records. In this instance, the input to the prediction model comprises data representative of network coverage at a plurality of training locations and the output comprises the geographic locations of the training locations. The training data records formed in step 202 forms the training data used in a supervised learning of the prediction model. Supervised training of the prediction model may comprise determining parameters of the prediction model which map the input fields of the training data records to the output fields of the training data records to a desired accuracy (e.g. which minimise a cost function).

[0085] Any suitable prediction model and training algorithm may be used. In some examples, the prediction model comprises a regression model. For example, the output of the regression model may comprise one or more numerical values belonging to a continuous range of values.

[0086] Examples of suitable algorithms may include a K-nearest neighbour algorithm, a linear regression algorithm, a support vector machine (e.g. a support-vector regression algorithm), a decision tree algorithm, a random forest algorithm, an adaptive boosting (AdaBoost) algorithm, a gradient boosting algorithm, an extreme gradient boosting algorithm (e.g. XGBoost), a voting algorithm and/or a stacking algorithm. In some examples, a deep learning algorithm may be used to train an artificial neural network.

[0087] The output of the training process typically comprises a plurality of determined parameters of the prediction model which best matches the input fields of the training data to the output fields of the training data. As will be described in further detail below, the determined parameters of the prediction model may be used to implement the prediction model to generate an output in dependence on inputs provided to the prediction model.

[0088] In at least some examples, the trained prediction model may be evaluated for accuracy. For example, a first subset of the available training records may be used to train the prediction model. A second subset of the available training records may then be used to evaluate the trained prediction model for accuracy. The evaluation of the trained prediction model may comprise providing the input fields of the second subset of the training records as inputs to the trained prediction model and implementing the trained prediction model to generate an output dependent on the inputs. The output of the prediction model may then be compared to output fields of the second subset of the training records.

[0089] If the trained prediction model had a perfect accuracy then the output of the implemented prediction model would match the output fields of the training records used to provide inputs to the prediction model. However, in practice no model has perfect accuracy and there will be some error difference between the model output and the output fields of the training records. The model error may be analysed and used to assess the accuracy of the trained prediction model. The trained model may then be used to determine the geographic location of an electronic device (at an unknown location) based on data representative of network coverage at the unknown location of the electronic device.

[0090] Figure 5 is a flow chart of an example method for determining a geographic location of an electronic device (e.g. a terminal) using a trained prediction model. The trained prediction model may, for example, be trained according to the method described above with reference to Figure 2. The method may be implemented on any suitable computing device. In some examples, each method step may be implemented on the same computing device. In other examples, different parts of the method may be implemented on different computing devices which may be in communication with each other.

[0091] The method of Figure 5 may, for example, be implemented for an electronic device situated at an unknown location. The electronic device, which may be referred to as a terminal, is configured to communicate over a mobile telecommunications network and is operable to make measurements indicative of the coverage of the mobile telecommunications network at the location of the terminal. By way of illustrative example, a location of a first terminal 104a and a location of a second terminal 104b are shown in Figure 3 for which the method of Figure 5 may be implemented. The location of the first terminal 104a and/or the second terminal 104b may not be known. Alternatively, the location of the first terminal 104a and/or second terminal 104b may have been determined using an alternative location determining technique, for example, using a GNSS. In such a situation it may be desirable to separately determine the location of the first terminal 104a and/or second terminal 104b using an alternative method so as to verify the GNSS based location.

[0092] At step 501 of Figure 5 data representative of coverage of the mobile telecommunications network at the location of the terminal is obtained. The data may correspond to, or at least be based on, one or more measurements made by the terminal at its current location. For example, as was described in detail above a terminal may measure one or more reference signals transmitted by one or more base stations 101 . Measurements of the one or more reference signals may be used to determine properties such as a received signal power (e.g. RSRP) and/or a received signal quality (e.g. RSRQ). Additionally or alternatively, a terminal 104 and/or a base station 101 may determine a measure of a propagation time (e.g. a timing advance) of signals exchanged between the terminal 104 and a base station 101. The coverage properties (e.g. received signal power, received signal quality and/or measure of propagation time) obtained in step 501 of Figure 5 may be similar to the data representative of coverage of the mobile telecommunications network which are received for a plurality of training locations in step 201 of Figure 2. Any of the features described above in connection with data representative of network coverage with reference to the method of Figure 2 may therefore also apply to the obtaining of data in step 501 of Figure 5.

[0093] As was explained above with reference to figure 2, one or more determined coverage properties (e.g. received signal power, received signal quality and/or measure of propagation time) may be associated with a particular cell 102 and/or base station 101. In some examples, one or more of the determined coverage properties (e.g. a received signal power, a received signal quality and/or a propagation time) may be determined for a plurality of different cells 102 or base stations 101 . The determined properties may be associated with an identifier of the cell 102 or base station 101 with which it is associated. For example, a determined property (e.g. a received signal power, a received signal quality and/or a propagation time) may be associated with a Physical Cell Identifier (PCI) of the cell with which the property is associated.

[0094] In some examples, coverage properties (e.g. a received signal power, a received signal quality and/or a propagation time) may be determined for each of the terminal’s serving cells (which may be a single serving cell or a plurality of serving cells). Additionally or alternatively coverage properties may be determined for one or more of the terminal’s neighbouring cells which are not a serving cell of the terminal 104. For example, in the example location of the first terminal 104a illustrated in Figure 3, the third cell 102c may be the serving cell for the first terminal 104a. In such a location, coverage properties (e.g. received signal power, received signal quality and/or propagation time) may be determined for the first terminal 104a and the third cell 102c which acts as the first terminal’s 104 serving cell. Additionally, coverage properties may be determined for the first terminal 104a and one or more neighbouring cells which might, for example, include the first cell 102a, the second cell 102b, the third cell 102d and/or one or more other neighbouring cells not shown in Figure 3. [0095] Figure 6 is a table showing an example of the properties which may be included in the data obtained at step 501 of Figure 5. Each column in the table of Figure 6 represents a different field included in the obtained data. The same conventions are used in Figure 6 as those described above with reference to Figure 4. For example, PCI_sc represents the PCI of the serving cell at the location of the terminal. RSRP sc is the RSRP associated with the serving cell at the location of the terminal. RSRQ_sc is the RSRQ associated with the serving cell at the location of the terminal. TA_sc is the timing advance associated with the serving cell at the location of the terminal. PCI_ncx represents the PCI of the xth neighbouring cell at the location of the terminal, where x is an index running from 1 to X. RSRP_ncx is the RSRP associated with the xth neighbouring cell at the location of the terminal. RSRQ_ncx is the RSRQ associated with the xth neighbouring cell at location of the terminal.

[0096] The data fields which are indicated in the table of Figure 6 correspond to the input data fields included in each training data record indicated in the table of Figure 4.

[0097] The data representative of coverage of the mobile telecommunications network which is obtained at step 501 of Figure 5 may comprise measurements made by a terminal 104 (e.g. the first terminal 104a indicated in Figure 3) and/or a base station 101. Additionally or alternatively the data may comprise properties which are determined in dependence on measurements made at terminal and/or a base station 101 . Obtaining the data in step 501 may take any suitable form. For example, obtaining the data may comprise carrying out measurements, determining one or more properties based on one or more measurements, reading the data from memory and/or receiving the data from another device at which the data is stored or determined.

[0098] At step 502 of Figure 5, inputs are provided to a trained prediction model. The inputs include the data obtained at step 501 . The prediction model is configured through training to determine a geographic location of an electronic device in dependence on data representative of the network coverage at the location of the electronic device. The prediction model may comprise a model trained using any of the methods described above with reference to Figure 2.

[0099] In some examples, the prediction model comprises a regression model. For example, the output of the regression model may comprise one or more numerical values belonging to a continuous range of values. The prediction model may comprise a machine learning model. The prediction model may have been trained by applying a supervised machine learning training algorithm to train the machine learning model. Any suitable prediction model and training algorithm may be used. Examples of suitable algorithms may include a K-nearest neighbour algorithm, a linear prediction algorithm, a support vector machine (e.g. a support- vector regression algorithm), a decision tree algorithm, a random forest algorithm, an adaptive boosting (AdaBoost) algorithm, a gradient boosting algorithm, an extreme gradient boosting algorithm (e.g. XGBoost) a voting algorithm and/or a stacking algorithm. In some examples, a deep learning algorithm may be used to train an artificial neural network.

[00100] The inputs to the prediction model may generally correspond to input fields of the training data records used to train the prediction model. For example, the inputs to the prediction model may comprise any of the data fields indicated in the table of Figure 6. Corresponding properties to any of the properties described above as forming part of an input field of a training data record (e.g. those described above with reference to the method of Figure 2) may be included in inputs provided to the prediction model.

[00101] At step 503 of Figure 5 the prediction model is implemented to generate an output representative of the geographic location of the electronic device. The output of the prediction model is dependent on the inputs provided at step 502.

[00102] The output of the prediction model may generally correspond to output fields of the training data records used to train the prediction model. Corresponding properties to any of the properties described above as forming part of an output field of a training data record may therefore be included in outputs provided by the prediction model. For example, the output of the prediction model may include one or more of a latitude, longitude and/or altitude of the electronic device.

[00103] Methods have been described above with reference to Figures 1 -4 for training a prediction model for determining a geographic location of an electronic device based on data representative of coverage of a mobile telecommunications network at a plurality of training locations. Methods have also been described above with reference to Figures 5 and 6 for implementing a trained predicted model to determine the geographic location of an electronic device based on data representative of coverage of a mobile telecommunications network at the location of the electronic device. As was described above, such prediction models, their training and implementation may find utility in a number of applications. Such applications may, for example, include determining the location of an electronic device which is situated at an above ground location (i.e. at a location which is only accessible by an airborne device). For example, the methods and prediction models described herein may find application in determining the geographic location of a network connected drone.

[00104] The determination of a geographic location of a device situated at an above ground location may be complicated by a number of different factors (for example, when compared to determining the location of a device situated at a ground based location). For example, the availability of training data at above ground locations may be limited when compared to ground based locations. Typically, the vast majority of devices connecting to a mobile telecommunications network may be situated at ground based locations. Consequently data representative of coverage of a mobile telecommunications network may be relatively abundant at ground based locations. However, typically there are far fewer devices operating in a network and situated in above ground locations than there are devices situated at ground based locations. Consequently, there may be significantly less measured data available which is representative of network coverage at above ground locations than ground based locations. Accordingly the availability of training data at above ground locations and suitable for training a prediction model for determining a geographic location of a device at above ground locations may be limited.

[00105] Furthermore, the data which is available at above ground locations is often collected through dedicated flights of airborne devices (e.g. a drone including a terminal device) in order to obtain measurements at a plurality of different above ground locations. Such flights have been carried out and have been used to collect data for use in training a prediction model for determining the location of a device situated above ground. However, the distribution of training locations for which training data is collected is limited by the geographic extent of the flights carried out. Given the limited geographic extent of many flights of airborne devices, the distribution of training locations for which training data is collected may be limited in extent and in some cases non-uniform.

[00106] Another factor which may complicate the determination of a position of a device at above ground locations (e.g. when compared to determining the location of a device situated at ground based locations) is that an airborne device (e.g. a network connected drone) may be situated at a plurality of different altitudes. At a given latitude and longitude there may only be a limited number of ground based locations at which a device may be situated. For example, for a given latitude and longitude a ground based device may only be situated substantially at the local ground level at that latitude and longitude. Determining a location of a ground based device may therefore only require the determination of two variables (such as a latitude and longitude). However, for a given latitude and longitude there may be a plurality of different above ground locations corresponding to different altitudes at which an airborne device (e.g. a network connected drone) may be situated. Determining a location of an airborne device may therefore require the determination of three variables such as a latitude, longitude and altitude.

[00107] The factors described above and/or other factors described herein may adversely affect the accuracy of determinations of the location of airborne devices using the methods of the type described above. It has been realised that the accuracy of the determination of the geographic location of electronic devices can be improved through one or more improvements to the training and/or implementation methods described above. Such improvements will be described below with reference to Figures 7-18.

[00108] Figure 7 is a flow chart of a further example method for training a prediction model for determining above ground coverage of a mobile telecommunications network. The method may be implemented on any suitable computing device. In some examples, each method step may be implemented on the same computing device. In other examples, different parts of the method may be implemented on different computing devices which may be in communication with each other.

[00109] The method of Figure 7 is similar to the method of Figure 2. Any of the steps or features described above with reference to the method of Figure 2 may also apply to the method of Figure 7. A detailed description of the same or corresponding method steps with reference to Figure 7 may be omitted and only the differences between the methods of Figure 2 and 7 will be described in detail.

[00110] At step 701 , data representative of coverage of a mobile telecommunications network at a plurality of training locations is received. The data and the receiving of the data may correspond to the data and the receiving of the data which was described above with reference to step 201 of Figure 2. For example, the data may include coverage properties such as a received signal power (e.g. RSRP), a received signal quality (e.g. RSRQ) and/or a measure of propagation time (e.g. a timing advance) at each training location. The data may include, for each training location, a plurality of different sets of coverage properties (e.g. received signal power, received signal quality and/or measure of propagation time) each set of coverage properties being associated with one of a plurality of different cells. For example, for a given training location, the data may include coverage properties associated with a serving cell at that training location and coverage properties associated with one or more neighbouring cells.

[00111] According to the method Figure 7, the data received at step 701 includes an indication of a serving cell at each training location. For example, the data may include an identifier, such as a PCI, of at least one serving cell at each training location. The data may further include an indication of one or more neighbouring cells at one or more of the training locations. For example, the data may include an identifier, such as a PCI, of one or more neighbouring cells at one or more of the training locations.

[00112] The method of Figure 7 includes at additional step 702 compared to the method of Figure 2. At step 702, location information is determined which is indicative of a location of the serving cells included in the received data. For example, for each serving cell included in the data received at step 701 a serving cell location associated with the serving cell may be determined. The location information may comprise the determined serving cell locations for each serving cell included in the data received at step 701 . The serving cell location for each serving cell may be any suitable geographic location associated with that serving cell. For example, the serving cell location for each serving cell may comprise a geographic location (e.g. latitude and longitude) of a base station 101 which operates that cell. In such examples, the geographic location of base stations 101 in the network may be stored in memory (e.g. on a server device and/or on a device implementing the method of Figure 7). Determining a serving cell location associated with a serving cell may comprise reading the location of a base station 101 which operates that serving cell from memory and/or querying a device which stores the location of the base station 101 .

[00113] Additionally or alternatively, the serving cell location for each serving cell may be determined in dependence on the training locations for which that cell acts as the serving cell. Figure 8 is a schematic illustration of training locations 105a, 105b in a section of an environment in which a mobile telecommunications network may operate. The section of the environment shown in Figure 8 includes a first cell 102a operated by a first base station 101 a and a second cell 102b operated by a second base station 101b. Also depicted in Figure 8 are a plurality of first training locations 105a indicated by black circles and a plurality of second training locations 105b indicated by black triangles. The first training locations 105a represent training locations for which data representative of network coverage is available and at which the first cell 102a acts as the serving cell at that training location. The second training locations 105b represent training locations for which data representative of network coverage is available and at which the second cell 102b acts as the serving cell at that location.

[00114] The first 105a and second 105b training locations may, for example, represent locations at which a network connected drone (or other suitable device) has been positioned (e.g. during one or more test flights) and data representative of network coverage at that location has been determined. The first training locations 105a may represent locations at which the network connected drone (or other suitable device) connected to the first cell 102a as its serving cell. The second training locations 105b may represent locations at which the network connected drone (or other suitable device) connected to the second cell 102b as its serving cell.

[00115] In at least some examples, a serving cell location for the first cell 102a may be determined in dependence on the first training locations 105a (at which the first cell acted as the serving cell). Similarly, the serving cell location for the second cell 102b may be determined in dependence on the second training locations 105b (at which the second cell acted as the serving cell). [00116] According to at least some examples, a serving cell location for a given serving cell may be taken as an average location of the training locations 105 associated with that serving cell. For example, a serving cell location for the first cell 102a may be taken as an average of the first training locations 105a at which the first cell 102a acts as the serving cell. An example, of such an average is shown as a first serving cell location 106a in Figure 8. Similarly, a serving cell location for the second cell 102b may be taken as an average of the second training locations 105b at which the second cell 102b acts as the serving cell. An example of such an average is shown as a second serving cell location 106b in Figure 8. Determining an average of a plurality of training locations may, for example, comprise determining an average longitude of the training locations and an average latitude of the training locations. Any suitable average may be determined such as a mean, a median and/or a mode.

[00117] For example, the first serving cell location 106a may be determined by determining an average (e.g. a mean, median and/or mode) of the longitudes of the first training locations 105a and determining an average (e.g. a mean, median and/or mode) of the latitudes of the first training locations 105a. Similarly, the second serving cell location 106b may be determined by determining an average (e.g. a mean, median and/or mode) of the longitudes of the second training locations 105b and determining an average (e.g. a mean, median and/or mode) of the latitudes of the second training locations 105b.

[00118] At step 703 of Figure 7, training data is formed comprising a plurality of training data records. Each training data record is associated with a training location 105. The training data records and the forming of the training data records may correspond to the forming of training data records which was described above with reference to step 202 of Figure 2. For example, each training data record may comprise at least the received data representative of the coverage of the mobile telecommunications network (e.g. the data received at step 701) at the training location 105 with which the training data record is associated and the training location 105 with which the training data record is associated. For example, a given training data record for a training location 105 may include coverage properties such as a received signal power (e.g. RSRP), a received signal quality (e.g. RSRQ) and/or a measure of propagation time (e.g. a timing advance) associated with a serving cell at the training location 105 and an identifier (e.g. PCI) of the serving cell. The given training data record may further include coverage properties such as a received signal power (e.g. RSRP), a received signal quality (e.g. RSRQ) and/or a measure of propagation time (e.g. a timing advance) associated with one or more neighbouring cells and identifiers (e.g. PCIs) of the one or more neighbouring cells. The given training data record may further include the geographic location of the training location 105, for example, in the form of the latitude, longitude and/or altitude of the training location 105. [00119] The training data further comprises the location information determined at step 702. As was explained above, the determined location information may comprise determined serving cell locations for each serving cell included in the data received at step 701 . In such examples, in addition to the fields included in the training data records described above and described with reference to Figures 2 and 4, the training data records may further include a serving cell location associated with the serving cell for each training location. For example, each training data record may additionally include the determined serving cell location (i.e. the serving cell location determined at step 702) associated with the serving cell for that training location.

[00120] Figure 9 is a table showing example properties which may be included in a plurality of training data records formed according to step 703 of the method of Figure 7. Most of the data fields included in the data training records indicated in Figure 9 are the same as the data fields included in the data training records indicated in Figure 4 and the same conventions are used to label the data fields and training data records in Figures 4 and 9. No further detailed description of the same data fields will therefore be provided with reference to Figure 9.

[00121] In addition to the data fields included in the training data records of Figure 4, the training data records of Figure 9 additionally include fields labelled Lat_sc and Long sc. The field Lat_sc represents the latitude of the serving cell location determined for the serving cell associated with the training data record (i.e. the serving cell indicated by the field PCI_sc in each training data record). The field Long sc represents the longitude of the serving cell location determined for the serving cell associated with the training data record (i.e. the serving cell indicated by the field PCI_sc in each training data record). The additional fields Lat_sc and Lon_sc shown in Figure 9 may be considered to comprise additional input fields of the training data records.

[00122] At step 704 of Figure 7, a prediction model is trained using the training data formed at step 703. The prediction model is trained for determining a geographic location of an electronic device. The prediction model and the training of the prediction model may correspond to the prediction model and the training of the prediction model which was described above with reference to step 203 of Figure 2. For example, the prediction model may comprise a machine learning model. The training of the prediction model may comprise applying a supervised machine learning training algorithm to train the machine learning model.

[00123] An example was described above in which the location information indicative of a location of the serving cells included in the received data, which is determined at step 702 of Figure 7, comprises determined serving cell locations for each serving cell. However, in other examples, the location information determined at step 702 may take alternative forms. For example, determining the location information may comprise determining probabilities that each of a plurality of cells are the serving cell for a given training location. For example, for each of the plurality of training locations probabilities may be determined which are each associated with a given cell and represent the probability that the given cell is the serving cell for that training location.

[00124] In some examples, probabilities for different serving cells may be determined for each of a plurality of reference regions. The plurality of training locations may be situated within the plurality of reference regions. For example, each training location may be situated within at least one reference region. The reference regions may for example, comprise a plurality of regions arranged on a periodic basis (e.g. a grid-like arrangement) such that the centres of the reference regions have a substantially uniform separation between adjacent reference region centres.

[00125] For each reference region, a serving cell may be determined which is most likely to be the serving cell within that reference region. That is, a serving cell may be determined which is most likely to be the serving cell at positions within the reference region. Such a determination is made in dependence on the data received at step 701. As was explained above, the data received at step 701 includes an indication of a serving cell at each training location. The indications of the serving cell at each training location may be used to determine a most likely serving cell for each reference region. For a given reference region, the most likely serving cell at one or more training locations situated within the reference region may be used to determine a most likely serving cell for the given reference region. .

[00126] In some examples, for each reference region a probability of one or more serving cells being the serving cell for that reference region (e.g. at location within the reference region) may be determined. For example, for a given reference region a probability may be determined that the serving cell for that region is a first cell. Additionally, one or more further probabilities may be determined that the serving cell for that reference region is one or more further cells. For example, probabilities may be determined for a second cell, a third cell and/or a fourth cell etc. Such a determination is made in dependence on the data received at step 701. For example, for each reference region, the serving cells at training locations situated within the reference region may be used to determine probabilities associated with each serving cell for the reference region as a whole. For example, a given reference region may include a first number of training locations for which the serving cell is a first cell, a second number of training locations for which the serving cell is a second cell and a third number of training locations for which the serving cell is a third cell etc. The first, second and third numbers may be used to determine probabilities that a serving cell for the given reference region is the first, second and third cells. [00127] In such examples, the location information may comprise one or more probabilities associated with one or more serving cells for each training location. For example, for each training location a reference region may be determined within which the training location is situated. The determined reference region may be used to determine probabilities that each of one or more serving cells are the serving cell for that training location. For example, as was explained above, probabilities associated with one or more serving cells may be determined for each reference region. The probabilities associated with the reference region within which a given training location is situated may be used to determine the serving cell probabilities for the given training location. That is, the probabilities that each of a plurality of cells are the serving cell for a given training location may comprise the determined probabilities that each of a plurality of cells are the serving cell for a reference region within which the training location is situated.

[00128] In examples in which probabilities associated with serving cells are determined, location information included in the training data may comprise probabilities that each of a plurality of cells are the serving cell for each of the plurality of training locations. For example, in addition to or as an alternative to the serving cell location fields (Lat_sc, Long_sc) indicated in the training data records shown in the table of Figure 9, training data records may include fields indicating probabilities associated with different serving cells for each training location. For example, for a given training data record the input fields may include a field indicating a first serving cell (e.g. a PCI of the first serving cell PCI_sc1 ) and a field indicating a probability (e.g. Prob_sc1) that the first serving cell (PCI_sc1 ) is the serving cell for the training location with which the training data record is associated. The training data record may further include input fields identifying one or more further serving cells (e.g. PCI_sc2, PCI_sc3, PCI_sc4 etc.) and indicating probabilities (e.g. Prob_sc2, Prob_sc3, Prob_sc4 etc.) that the one or more further serving cells is the serving cell for the training location.

[00129] In examples in which probabilities associated with a plurality of serving cells are determined and included in the training data records, the training data records may not include additional explicit locations (e.g. the Lat_sc, Long_sc fields described above with reference to Figure). However, the probabilities associated with a plurality of serving cells are indicative of a location of one or more serving cells and are therefore considered as examples of location information indicative of a location of the one or more serving cells included in the data representative of coverage of the mobile telecommunications network as determined at step 702 of the method of Figure 7. This is because the probabilities are determined in dependence on the received data which indicates a serving cell at a plurality of different training locations. The probabilities determined in dependence on this data and for a plurality of different training locations is therefore location information indicative of a location of the serving cells for which probabilities are determined. Including such information in the training data will allow a prediction model which is trained based on the training data to account for the geographic distribution of the serving cells in a similar way to explicitly including a location associated with a serving cell in the training data as was described above.

[00130] A prediction model trained according to the method of Figure 7 may be implemented in order to determine the geographic location of an electronic device in dependence on data representative of coverage of the mobile telecommunications network at the location of the electronic device. For example, the method described above with reference to Figure 5 may be implemented using a prediction model trained according to the method of Figure 7. The obtaining data representative of coverage of the mobile telecommunications network at the location of the electronic device at step 501 of Figure 5 may include obtaining an indication of a serving cell at the location of the electronic device. For example, an identifier (e.g. a PCI) of the serving cell at the location of the electronic device may be obtained. The input to the trained prediction model at step 502 of Figure 5 may include the obtained indication of a serving cell at the location of the electronic device. For example, an identifier (e.g. a PCI) of the serving cell at the location of the electronic device may be included in the input provided to the prediction model. The input to the trained prediction model may, for example, include one or more of the data fields shown in the table of Figure 6 which includes the field PCI_sc corresponding to the PCI of the serving cell.

[00131] It has been found that by including location information indicative of a location of one or more serving cells in the training data used to train a prediction model, as was described above with reference to Figure 7, the accuracy of location determinations made using the trained prediction model may be substantially improved. The location information provides additional location related data which improves a machine learning training process in order to more accurately capture a relationship between network coverage properties and location. Consequently the accuracy of the determination of a location of an electronic device made using such a trained prediction model may be substantially improved.

[00132] According to some examples of the methods described above with reference to Figures 7-9, serving cell locations associated with a serving cell at each training location are added to the training data records. Additionally or alternatively, corresponding neighbour cell locations may be determined and added to the training data records. Figure 10 is a table showing further example properties which may be included in a plurality of training data records according to methods described herein. The training data records indicated in Figure 10 additionally include (when compared to the training data records indicated in Figure 9) fields labelled as Lat_ncx and Lon_ncx. The field Lat_ncx represents the latitude of a neighbour cell location determined for the xth neighbouring cell associated with the training data record (i.e. the neighbouring cell indicated by the field PCI_ncx). The field Long_ncx represents the longitude of a neighbour cell location determined for the xth neighbouring cell associated with the training data record (i.e. the neighbouring cell indicated by the field PCI ncx). The fields Lat_ncx and Long_ncx may be included for every neighbouring cell included in a training data record (i.e. for each value of the index x from 1 and X). The additional fields Lat_ncx and Lon_ncx shown in Figure 10 may be considered to comprise additional input fields of the training data records.

[00133] The neighbour cell locations may be determined in a corresponding way to the determination of serving cell locations described above with reference to step 702 of Figure 7 and Figure 8. For example, a neighbour cell location for a given neighbouring cell may be determined as a geographic location of a base station 101 which operates the neighbouring cell. Additionally or alternatively, a neighbour cell location for a given neighbouring cell may be determined in dependence on training locations for which that neighbouring cell acts as the serving cell.

[00134] For example, with reference again to Figure 8 the first cell 102a may be a neighbouring cell for at least some of the second training locations 105b. Training data records associated with the second training locations 105b may therefore include a neighbour cell location associated with the first cell 102a. The neighbour cell location associated with the first cell 102a may be determined in dependence on the first training locations 105a at which the first cell 102a acts as a serving cell. For example, the neighbour cell location associated with the first cell 102a may be determined as an average (e.g. an average latitude and an average longitude) of the first training locations 105a. The neighbour cell location associated with the first cell 102a may therefore correspond to the serving cell location which is determined for the first cell 102a and included in the training data records associated with the first training locations 105a, as was described above with reference to step 702 of Figure 7.

[00135] Similarly, the second cell 102b may be a neighbouring cell for at least some of the first training locations 105a. Training data records associated with the first training locations 105a may therefore include a neighbour cell location associated with the second cell 102b. The neighbour cell location associated with the second cell 102b may be determined in dependence on the second training locations 105b at which the second cell 102b acts as a serving cell. For example, the neighbour cell location associated with the second cell 102b may be determined as an average (e.g. an average latitude and an average longitude) of the second training locations 105b. The neighbour cell location associated with the second cell 102b may therefore correspond to the serving cell location which is determined for the second cell 102b and included in the training data records associated with the second training locations 105b, as was described above with reference to step 702 of Figure 7.

[00136] Methods have been described above with reference to Figures 7-10 in which one or more locations (e.g. a serving cell location and/or a neighbour cell location) associated with cells of a mobile telecommunications network are determined and included in training data records. According to at least some examples contemplated herein, a plurality of serving cell locations may be determined for at least one serving cell. For example, training locations associated with a given serving cell may be grouped into two or more sub-groups and a serving cell location may be determined for each sub-group of training locations and included in the corresponding training data records.

[00137] Figure 11 is a schematic illustration of a section of an environment in which a mobile telecommunications network may operate. The section of the environment shown in Figure 11 includes a first cell 102a operated by a first base station 101 a. Also shown in Figure 11 are a plurality of training locations 105 indicated by black circles. Each of the training locations 105 shown in Figure 11 represent training locations at which the first cell 102a acts as the serving cell at that location.

[00138] According to at least some examples contemplated herein, training locations 105 associated with the same serving cell 102a may be grouped into a plurality of sub-groups. For example, the training locations 105 shown in Figure 11 may be grouped into a first sub-group 107a and a second sub-group 107b as indicated by the dashed ellipses shown in Figure 11 . For each sub-group 107a, 107b of training locations 105 a serving cell sub-group location 106a, 106b may be determined. For example, a first serving cell sub-group location 106a may be determined for the first sub-group 107a and a second serving cell sub-group location 106b may be determined for the second sub-group 107b. The serving cell sub-group locations 106a may be determined in dependence on the training locations 105 belonging to that sub-group. For example, the first serving cell sub-group location 106a may be determined in dependence on the training locations 105 belonging to the first sub-group 107a. The second serving cell sub-group location 106b may be determined in dependence on the training locations 105 belonging to the second sub-group 107b.

[00139] A serving cell sub-group location 106a, 106b for a given sub-group 107a, 107b may be taken as an average location of the training locations 105 in that sub-group 107a, 107b. For example, the first serving cell sub-group location 106a for the first cell sub-group may be taken as an average of the training locations 105 belonging to the first sub-group 107a. Similarly, the second serving cell sub-group location 106b for the second sub-group 107b may be taken as an average of the training locations 105 belonging to the second sub-group 107b. Determining an average of a plurality of training locations may, for example, comprise determining an average longitude of the training locations and an average latitude of the training locations. Any suitable average may be determined such as a mean, a median and/or a mode.

[00140] Grouping training locations into a plurality of sub-groups 107a, 107b and determining serving cell sub-group locations 106a, 106b for each sub-group 107a, 107b may be considered to be an example of step 702 of the method of Figure 7. In such examples, determining a serving cell location associated with a serving cell (as in step 702 of the method of Figure 7) may, for at least some serving cells comprise grouping the training locations 105 associated with that serving cell 102a (i.e. the training locations 105 for which the cell 102a acts as a serving cell) into a plurality of sub-groups 107a, 107b and determining a serving cell sub-group location 106a, 106b for each sub-group 107a, 107b. In such examples, the serving cell location included in the training data record (for example, a training data record formed according to step 703 of Figure 7) may, for at least some training locations, comprise a serving cell sub-group location 106a, 106b determined for the sub-group 107a, 107b which the training location belongs to.

[00141] Figure 12 is table showing example properties which may be included in two training data records formed according to step 703 of the method of Figure 7. Most of the data fields included in the data training records indicated in Figure 12 are the same as the data fields included in the data training records indicated in Figure 9 and the same conventions are used to label the data fields and data training records in Figures 9 and 12. No further detailed description of the same data fields will therefore be provided with reference to Figure 12.

[00142] The training data records shown in Figure 12 include a first training data record shown in the first row of the table of Figure 12 and a second training data record shown in the second row of the table of Figure 12. The first training data record and the second training data record correspond to training locations for which the serving cell at those locations are the same, as indicated by the field PCI_sc1 which is the same in each of the training data records. The training location with which the first training data record is associated is grouped into a first sub-group, whereas the training location with which the second training data record is associated is grouped into a second sub-group. For example, the first training data record may be associated with a training location belonging to the first sub-group 107a indicated in Figure 11. The second training data record may be associated with a training location belonging to the second sub-group 107b indicated in Figure 11. The serving cell location included in the first training data record comprises a first serving cell sub-group location (e.g. the first serving cell sub-group location 106a indicated in Figure 11) determined for the first sub-group. This is indicated by the fields Lat_scsg1 and Long_scsg1 in Figure 12, which correspond to the latitude and longitude of the first serving cell sub-group location respectively. The serving cell location included in the second training data record comprises a second serving cell sub-group location (e.g. the second serving cell sub-group location 106b indicated in Figure 11) determined for the second sub-group. This is indicated by the fields Lat_scsg2 and Long_scsg2 in Figure 12, which correspond to the latitude and longitude of the second serving cell sub-group location respectively.

[00143] As explained above, in at least some examples, a serving cell location included in different training data records may be different from each other even when the different training data records are associated with the same serving cell. For example, the first and second training data records shown in Figure 12 are associated with the same serving cell but are grouped into different sub-groups. The serving cell location included in the first and second training data records are therefore different (corresponding to the first and second serving cell sub-group locations respectively) because the training locations with which the first and second training data records are associated are grouped into different sub-groups.

[00144] Also included in each training data record indicated in Figure 12 is a sub-group identifier for identifying the sub-group to which the training location has been grouped. The sub-group identifiers may take any suitable form to identify the sub-group relative to other subgroups. In the table of Figure 12 the sub-group identifiers are labelled as SGID 1 for the first sub-group and SGID 2 for the second sub-group.

[00145] Training locations may be grouped into sub-groups according to any suitable method. For example, a clustering technique may be used to group training locations into sub-groups. In at least some examples, an unsupervised learning technique may be used to group training locations into sub-groups. Training locations may be grouped into sub-groups according to one or more properties associated with the training locations. For example, training locations may be grouped into sub-groups according to their location such that training locations positioned relatively close to each other are grouped into the same sub-group. This is the case for the sub-groups 107a, 107b shown in Figure 11 , where the sub-groups 107a, 107b represent training locations which are clustered together geographically.

[00146] In some examples, one or more properties in addition to or as an alternative to geographic location may be used to group training locations into sub-groups. For example, properties such as a measure of propagation time (e.g. a timing advance) at each training location may be used to group training locations into sub-groups. Propagation time (e.g. timing advance) for signals propagating between a given training location and a base station is at least a function of distance between the training location and the base station. Training locations sharing the same serving cell may therefore have different propagation times depending at least in part on their distance from a base station operating the serving cell. In at least some examples, training locations may be grouped into sub-groups based at least in part on a measure of propagation time (e.g. a timing advance) associated with the training locations. For example, training locations for which the propagation time (e.g. timing advance) is relatively long (e.g. being over a threshold propagation time) may be grouped together in a first sub-group. Training locations for which the propagation time (e.g. timing advance) is relatively short (e.g. being below a threshold propagation time) may be grouped together in a second sub-group. In such an example, the first sub-group may represent training locations which are situated relatively far away from a base station and the second sub-group may represent training locations situated relatively close to the base station.

[00147] Grouping training locations into sub-groups may serve to link training locations having one or more similar properties. For example, training locations grouped into the same subgroup may be located at relatively similar geographic locations and/or may be situated at relatively similar distances from a base station. Linking training locations having one or more similar properties by grouping them into the same sub-group serves to add additional information to the training data records which has been found to improve the accuracy of location determination carried out using a prediction model trained using the training data records.

[00148] Figure 13 is a table showing an example of the properties which may be included in the data obtained for providing inputs to a trained prediction model for determining the geographic location of a device. For example, the properties indicated in Figure 13 may correspond to data obtained at step 501 of the method of Figure 5. The data indicated in Figure 13 closely corresponds to the data indicated in the table of Figure 6. The same conventions are used in Figure 13 as those described above with reference to Figure 6 and no further detailed description will be provided with reference to Figure 13. The data fields which are indicated in the table of Figure 13 correspond to the input data fields included in each training data record indicated in the table of Figure 12.

[00149] In addition to the data fields included in the table of Figure 6, the table of Figure 13 also includes a sub-group identifier field labelled SGID in Figure 13. It will be appreciated that for an electronic device whose location is to be determined (e.g. using a method according to Figure 5) the identity of a sub-group may not be a property which can be directly measured by the device, since the sub-groups were determined as part of the preparation of training data records. This contrasts, for example, with identifying a serving cell (e.g. a PCI) which is directly determined by an electronic device during its normal course of operation in the network. However, linking a device to a sub-group may improve the accuracy of a location determination where the prediction model was trained using training data records including a serving cell sub-group location since location information linked to sub-groups is included in the trained model.

[00150] A sub-group identifier may be obtained for a device whose location is to be determined, for example, by performing an initial location determination and using the initial location to link the device to a sub-group. The initial location determination may be performed, for example, by using a prediction model which is trained without grouping training locations into sub-groups and using inputs to such a prediction model which does not include a subgroup identifier. An initial determined location may be used to match the initial location to a sub-group. For example, the initial location may be matched to a sub-group having a serving cell sub-group location which is closest to the initial location. Such a determination and matching may be performed, for example, as part of the data obtaining step at step 501 of Figure 5. A sub-group to which the initial location is matched may be indicated by a sub-group identifier included in the input to a trained prediction model (e.g. the inputs provided at step 502 of the method of Figure 5).

[00151] Several methods have been described above with reference to Figures 1 -13 for training a prediction model and implementing a trained prediction model for determining a geographic location of a device. According to at least some examples, prediction model may be trained using training data records which are associated with training locations corresponding to a plurality of different serving cells. For example, referring again to Figure 3 a plurality of training locations 105 are shown. The training locations 105 are situated in a plurality of different cells 102. For example, a serving cell at some of the training locations 105 is the first cell 102a, whereas a serving cell at some others of the training locations 105 is the second cell 102b, the third cell 102c or the fourth cell 102d. According to at least some examples, a prediction model may be trained using training data records associated with a plurality of different serving cells. For example, training data records associated with all of the training locations 105 including locations having different serving cells shown in Figure 5 (e.g. all of the locations 105 shown in Figure 5) may be used to train a prediction model. The same prediction model may then be used to determine a geographic location of devices having a plurality of different serving cells. For example, the same prediction model may be used to determine the location of both the first terminal 104a (for which the third cell 102c is the serving cell) and the second terminal 104b (for which the second cell 101 b is the serving cell) shown in Figure 3.

[00152] Whilst in some examples, the same prediction model may be used across a plurality of serving cells, in other examples, separate prediction models may be trained and implemented for different serving cells. For example, taking the example training locations shown in Figure 8, the first training locations 105a for which the serving cell is the first cell 102a may be used to form training data records to train a first prediction model. The second training locations 105b for which the serving cell is the second cell 102b may be used to form training data records to train a second prediction model. The formation of training data records and the training of separate prediction models may include any of the features and method steps described throughout this specification for the formation of training data records and training of prediction models.

[00153] Figure 14 is a flow chart of an example method for training a prediction model for determining the geographic location of a terminal based on data representative of coverage of a mobile telecommunications network. The method may be implemented on any suitable computing device. In some examples, each method step may be implemented on the same computing device. In other examples, different parts of the method may be implemented on different computing devices which may be in communication with each other.

[00154] The method of Figure 14 is similar to the methods described above with reference to Figures 2 and 7. Any of the steps or features described above with reference to the method of Figures 2 or 7 may also apply to the method of Figure 14. A detailed description of the same or corresponding method steps with reference to Figure 14 may be omitted and only the differences between the methods of Figure 2, 7 and 14 will be described in detail.

[00155] At step 1401 of Figure 14, data representative of coverage of a mobile telecommunications network at a plurality of training locations is received. Step 1401 corresponds to steps 201 and 701 of the methods of Figures 2 and 7 and the data received at step 1401 may include any of the features described above with reference to Figures 2 and 7. In particular, the data received at step 1401 may include an indication (such as a PCI) of a serving cell of the network for each training location. The received data may include data associated with training locations which are served by a plurality of different serving cells. As was explained extensively above, the data may also include coverage properties (e.g. received signal power, received signal quality and/or a measure of propagation time) for each training location.

[00156] At step 1402, training data comprising a plurality of training data records is formed, where each training data record is associated with a training location of the plurality of training locations. Step 1402 corresponds to steps 202 and 703 of the methods of Figures 2 and 7 and the training data records and their formation may include any of the features described above with reference to Figures 2 and 7. Each training data record may include at least the training location with which the training data record is associated (e.g. the latitude, longitude and/or altitude of the training location) and the data representative of network coverage at that training location. Each training data record may be associated with a serving cell at the training location for that training data record. The serving cell with which a given training data record is associated may comprise a cell which acts as a serving cell at the training location with which the training data record is associated. The training data records created at step 1402 may include training data records associated with training locations which are served by a plurality of different serving cells. That is, at least some of the training data records may be associated with different serving cells.

[00157] At step 1404, a subset of the training data records are selected, where each of the selected subset of training data records are associated with the same serving cell. That is, a subset of the training data records are selected which are all associated with the same serving cell. For example, taking the example training locations shown in Figure 8, training data records associated with a subset of all of the training locations shown in Figure 8 may be selected. For example, the selected subset may comprise training data records associated with either the first training locations 105a or the second training locations 105b.

[00158] At step 1404, a prediction model for determining the geographic location of an electronic device is trained using the subset of training data records selected in step 1403. Step 1404 corresponds to steps 203 and 704 of the methods of Figures 2 and 7 and the prediction model and the training of the prediction model may include any of the features described above with reference to Figures 2 and 7. By using the selected subset of training data records to train a prediction model at step 1404, the trained prediction model is specific to the serving cell for which the subset of training data records was selected and is therefore associated with that serving cell.

[00159] Training a prediction model based on a selected subset of training data records such that the prediction model is associated with a single serving cell serves to capture the unique network conditions and other factors for that serving cell in the trained prediction model. It has been found that this can significantly improve the accuracy of location determinations made using such a trained prediction model for devices having the serving cell with which the prediction model is associated (when compared, for example, to using a prediction model which is trained using training data records associated with a plurality of different serving cells).

[00160] In examples in which different prediction models are trained and implemented for different serving cells (e.g. training a prediction model using the method of Figure 14), a plurality of different prediction models may be trained and implemented in order to determine the geographic location of devices having different serving cells. This may be achieved, for example, by performing steps 1403 and 1404 of the method of Figure 14 a plurality of times to train a plurality of prediction models associated with different serving cells. For example, referring to the training locations shown in Figure 8, steps 1403 and 1404 may be performed a first time in which a first subset of training data records are selected, where the first subset of training data records are associated with the first training locations 105a (which are served by the first cell 102a). A first prediction model may be trained using the first subset of training data records and may be associated with the first cell 102a. Steps 1403 and 1404 may be performed a second time in which a second subset of training data records are selected, where the second subset of training data records are associated with the second training locations 105b (which are served by the second cell 102b). A second prediction model may be trained using the second subset of training data records and may be associated with the second cell 102b. In general, as many prediction models may be trained based on as many subsets of training data as needed to provide geographical coverage across all areas and/or cells for which training data is available.

[00161] In some examples, prediction models may be trained based on a sub-group of training locations for a particular cell. For example, in an analogous manner to the grouping of training locations into sub-groups as was described above with reference to Figure 11 , training data records associated with a given serving cell may be grouped into sub-groups to train a prediction model associated with a given sub-group of training data records.

[00162] In some examples, selecting a subset of training data records at step 1403 of the method of Figure 14 may comprise grouping the training data records associated with the same serving cell into a plurality of sub-groups of training data records. For example, taking the training locations 105 depicted in Figure 11 , which all have the same serving cell 102a, the training data records associated with the training locations 105 may be grouped into a plurality of sub-groups of training records. For example, the training location 105 may be grouped into a first sub-group 107a and a second sub-group 107b. A first sub-group of training data records may comprise training data records associated with the first sub-group 107a of training locations. A second sub-group of training data records may comprise training data records associated with the second sub-group 107b of training locations.

[00163] The grouping into sub-groups of training records may be similar to the grouping of training locations which was described above with reference to Figure 11 . For example, the training records associated with the training locations 105 may be grouped into sub-groups based on one or more properties such as location and/or a measure of propagation time for each training location. Any suitable grouping methods may be used such as clustering techniques and/or unsupervised learning. Any of the features described above with reference to Figure 11 may also apply to the grouping of training data records into sub-groups for training a prediction model associated with a sub-group. [00164] Referring again to Figure 14, selecting a subset of training data records at step 1403 may further comprise selecting a first sub-group of the sub-groups as the selected subset of the training data records associated with the same serving cell. For example, the first subgroup of training data records associated with the first sub-group 107a of training locations may be selected as the subset of training data records each associated with the same serving cell at step 1403. Training a prediction model based on the selected subset of training data records at step 1404 of Figure 14 may comprise training a first prediction model using the first sub-group of the sub-groups of training data. For example, a first prediction model may be trained using the first sub-group of training data records associated with the first sub-group 107a of training locations. The first prediction model is associated with the first-sub group.

[00165] At least parts of steps 1403 and 1404 may be performed a plurality of times for a given serving cell in order to train a plurality of prediction models each associated with different sub-groups of training data records. For example, in addition to the steps described above for training a first prediction model, step 1043 may further comprise selecting a second-sub group of training data as the selected subset of the training data records associated with the same serving cell. For example, the second sub-group of training data records associated with the second sub-group 107b of training locations may be selected as the subset of training data records each associated with the same serving cell at step 1403. Training a prediction model based on the selected subset of training data records at step 1404 of Figure 14 may further comprise training a second prediction model using the second sub-group of the sub-groups of training data. For example, a second prediction model may be trained using the second subgroup of training data records associated with the second sub-group 107a of training locations. The second prediction model is associated with the second-sub group.

[00166] As was described above, for a given serving cell a plurality of prediction models may be trained, where each prediction model is associated with a different sub-group of training locations available for that cell. Similarly to the training of different prediction models for different serving cells, training different prediction models based on different sub-groups of training data records all associated with a single serving cell may serve to capture the unique network conditions and other factors for that sub-group in the trained prediction model. It has been found that in at least some situations this can significantly improve the accuracy of location determinations made using such a trained prediction model for devices having one or more similar properties for the relevant sub-group (when compared, for example, to using a prediction model which is trained using training data records associated with a plurality of different serving cells, or trained using all of the training data records associated with a serving cell). [00167] For example, training locations for which coverage properties are available for a given serving cell may include first training locations at which network coverage is relatively stable, received signal power is relatively strong, received signal quality is relatively good and/or propagation time is relatively low. The training locations may also include second training locations at which network coverage is relatively unstable, received signal power is relatively low, received signal quality is relatively poor and/or propagation time is relatively high. The difference in properties of the first training locations and the second training locations may lead to the training data records associated with the first and second training locations being grouped into different sub-groups. First and second prediction models may be trained using the training data records associated with the first and second training locations respectively. The first and second prediction models may therefore separately capture the different network conditions which apply to the first and second locations, which may improve the accuracy of subsequent location determinations when a suitable prediction model is chosen to carry out the location determination.

[00168] As was explained above with reference to Figure 14, in some examples a plurality of different prediction models for determining a location of a device may be trained. Each prediction model may be associated with a serving cell. Furthermore, in at least some examples, a plurality of different prediction models may be trained for a given serving cell (e.g. for different sub-groups all associated with the same serving cell). In such examples, when a plurality of different prediction models are available, determining the location of a device may comprise selecting a suitable prediction model to use for carrying out the location determination.

[00169] Figure 15 is a flow chart of an example method for determining a geographic location of an electronic device (e.g. a terminal) using a trained prediction model selected from a plurality of trained prediction models. The plurality of trained prediction models may, for example, be trained according to a method described above with reference to Figure 14. The method may be implemented on any suitable computing device. In some examples, each method step may be implemented on the same computing device. In other examples, different parts of the method may be implemented on different computing devices which may be in communication with each other.

[00170] The method of Figure 15 is similar to the methods described above with reference to Figure 5. Any of the steps or features described above with reference to the method of Figure 5 may also apply to the method of Figure 15. A detailed description of the same or corresponding method steps with reference to Figure 15 may be omitted and only the differences between the methods of Figures 5 and 15 will be described in detail. [00171] The method of Figure 15 may, for example, be implemented for an electronic device situated at an unknown location. The electronic device, which may be referred to as a terminal, is configured to communicate over a mobile telecommunications network and is operable to make measurements indicative of the coverage of the mobile telecommunications network at the location of the terminal.

[00172] At step 1501 of Figure 15 data representative of coverage of the mobile telecommunications network at the location of the terminal is obtained. Step 1501 corresponds to steps 501 of the method of Figure 5 and the data obtained at step 1501 may include any of the features described above with reference to Figure 5. The data may correspond to, or at least be based, on one or more measurements made by the device at its current location and may include coverage properties such as received signal power, received signal quality and/or a measure of propagation time for one or more cells. The data may further include an indication (e.g. PCI) of a serving cell of the network at the location of the device. The serving cell may comprise the serving cell currently being used by the device to connect to the network which may be a property which is routinely established as part of the normal operation of the device.

[00173] At step 1502 a prediction model from a plurality of prediction models is selected. The plurality of prediction models may each be associated with a serving cell of the network. For example, as was described above with reference to Figure 14, each of the plurality of prediction models may have been trained using training data records associated with a particular serving cell. In some examples, some or all of the prediction models may be associated with a sub-group of training data records associated with a serving cell.

[00174] Selecting a prediction model from the plurality of prediction models may comprise selecting a prediction model which is associated with the indicated serving cell at the location of the device. That is, the serving cell which is indicated in the data obtained at step 1501 may be used to select a prediction model which is associated with the same serving cell. In some examples, the plurality of prediction models may include a single prediction model associated with each serving cell. In such examples, selecting a prediction model may comprise selected the single prediction model which is associated with the same serving cell which is indicated in the data obtained at step 1501.

[00175] In other examples, the plurality of prediction models may include a plurality of prediction models associated with at least some of the serving cells with which the prediction models are associated. For example, for the serving cell which is indicated in the data obtained at step 1501 , the plurality of prediction models may include a plurality of prediction models associated with the indicated serving cell. In such a scenario, selecting a prediction model from the plurality of prediction models may comprise selecting a prediction model from a plurality of prediction models associated with the indicated serving cell.

[00176] As was explained above, different prediction models associated with the same serving cell may comprise prediction models which are trained using different sub-groups of training data records. Selecting a prediction model from a plurality of prediction models associated with the indicated serving cell may comprise selecting a sub-group from a plurality of sub-groups associated with the indicated serving cell. A number of different methods may be used to select a sub-group from a plurality of sub-groups associated with the indicated serving cell. For example, a sub-group may be chosen which has one or more similar properties to the properties indicated in the data obtained at step 1501 . For example, different sub-groups may be associated with training data records having different ranges of a propagation time (e.g. timing advance). In such a scenario a sub-group may be chosen which has a range of propagation times which is most similar to a propagation time included in the data obtained at step 1501 .

[00177] In some examples, selecting a sub-group may comprise performing an initial location determination and using the initial location to select a sub-group. The initial location determination may be performed, for example, by using a prediction model which is not associated with a particular sub-group of training locations or training data records. For example, a prediction model trained using all of the training data records associated with the indicated serving cell or a prediction model trained using training data records associated with a plurality of different serving cells may be used to determine an initial determined location. An initial determined location may be used to match the initial location to a sub-group of training locations. For example, the initial location may be matched to a sub-group of training locations which is closest to the initial location. Such a determination and matching may be performed, for example, as part of the selection step at step 1502 of Figure 15.

[00178] A sub-group may be selected from a plurality of sub-groups associated with the indicated serving cell according to any suitable method. Selecting a prediction model from the plurality of prediction models may then comprise selecting a prediction model which is associated with the selected sub-group.

[00179] At step 1503 of Figure 15, the data obtained at step 1501 is provided to the selected trained prediction model selected at step 1502. Step 1503 of Figure 15 corresponds to step 502 of the method of Figure 5 and any of the features described above with reference to step 502 of Figure 5 may also apply to step 1503 of Figure 15.

[00180] At step 1504 of Figure 15, the prediction model selected at step 1502 of Figure 15 is implemented to generate an output representative of the geographic location of the electronic device. The output of the selected prediction model is dependent on the inputs provided at step 1503. Step 1504 of Figure 15 corresponds to step 503 of the method of Figure 5 and any of the features described above with reference to step 503 of Figure 5 may also apply to step 1504 of Figure 15. The output of the prediction model may, for example, include one or more of a latitude, longitude and/or altitude of the electronic device.

[00181] Several examples have been described above in which one or more prediction models are trained using training data records corresponding to training data collected by one or more devices operating in the network. In the description of such training data it has been generally assumed that all of the available training data is used in the training of relevant prediction models. However, according to some examples contemplated herein training data may be filtered, supplemented or otherwise modified before using the data to train a prediction model.

[00182] In some situations it may be beneficial to filter available data so as to exclude some training data from training data records used to train a prediction model. For example, some of the available data representative of network coverage (e.g. coverage properties measured or otherwise determined for a plurality of training locations) may correspond to training locations at which the network coverage is relatively poor or unstable. Consequently the coverage properties determined for such training locations may be relatively unreliable. Such training locations may, for example, correspond to locations located at relatively large distances from a serving base station and/or locations at which signals exchanged with a base station suffer from obstruction and/or interference. If training data records are formed using training locations at which the network coverage is relatively poor and unstable then this has the potential to influence the accuracy and/or reliability of a prediction model trained using such training data records.

[00183] According to at least some examples, a subset of available training locations may be selected to form training data records for training one or more prediction models. The subset of available training locations may be selected to exclude training locations at which the network coverage is relatively poor and/or unstable. One or more filters may be used to select the subset of available training locations. For example, one measure which may be indicative of relatively poor and/or unstable network conditions may be a measure of propagation time of signals exchanged between a device and a serving base station. One such measure may be a timing advance, as has been described above. Training locations for which a timing advance is relatively high for a serving base station may represent locations at which a device is situated relatively far from the serving base station. [00184] Figure 16 is a flow chart of an example method for training a prediction model for determining the geographic location of a terminal based on data representative of coverage of a mobile telecommunications network. The method may be implemented on any suitable computing device. In some examples, each method step may be implemented on the same computing device. In other examples, different parts of the method may be implemented on different computing devices which may be in communication with each other.

[00185] The method of Figure 16 is similar to the methods described above with reference to Figures 2, 7 and 14. Any of the steps or features described above with reference to the methods of Figures 2, 7 or 14 may also apply to the method of Figure 16. A detailed description of the same or corresponding method steps with reference to Figure 16 may be omitted and only the differences between the methods of Figure 2, 7, 14 and 16 will be described in detail.

[00186] At step 1601 of Figure 16 data representative of coverage of a mobile telecommunications network at a plurality of training locations is received. Step 1601 corresponds to steps 201 , 701 and 1401 of the methods of Figures 2, 7 and 14 respectively and the data received at step 1601 may include any of the features described above with reference to Figures 2, 7 or 14. In particular, the data received at step 1601 may include a measure of propagation time of signals exchanged between each training location and at least one base station. Such a measure may, for example, comprise a timing advance for each training location. As was explained extensively above, the data may also include other coverage properties (e.g. received signal power and/or received signal quality) for each training location.

[00187] At step 1602, a subset of the plurality of training locations is determined. The subset of the plurality of training locations may be determined as all training locations for which the measure of propagation time (included in the data received at step 1601 ) is less than a threshold propagation time. The threshold propagation time may be chosen such that propagation times greater than the threshold propagation time represent relatively poor and/or unstable network conditions. The selected subset of training locations may therefore represent locations at which the network coverage is relative good and/or stable.

[00188] At step 1603, training data comprising a plurality of training data records is formed, where each training data record is associated with a training location of the determined subset of the plurality of training locations (the subset determined at step 1602). Step 1603 corresponds to steps 202, 703 and 1402 of the methods of Figures 2, 7 and 14 respectively and the training data records and their formation may include any of the features described above with reference to Figures 2, 7 or 14. Each training data record may include at least the training location with which the training data record is associated (e.g. the latitude, longitude and/or altitude of the training location) and the data representative of network coverage at that training location. Each training data record may be associated with a serving cell at the training location for that training data record.

[00189] At step 1604, a prediction model for determining the geographic location of an electronic device is trained using the training data records formed at step 1603. Step 1604 corresponds to steps 203, 704 and 1404 of the methods of Figures 2, 7 and 14 respectively and the prediction model and the training of the prediction model may include any of the features described above with reference to Figures 2, 7 or 14. As was explained above, the training data records used to train the prediction model may represent training locations for which the network coverage is relatively good and/or stable. It has been found that, for at least some situations, the accuracy and/or reliability of location determinations carried out using such a trained prediction model may be improved (relative to, for example, using a prediction model trained using training data records for training locations at which the network conditions are relatively poor and/or unstable).

[00190] In some situations training locations for which network coverage data is available may be non-uniformly distributed in space. Figure 17 is a schematic depiction of an example distribution of training locations 105 in a cell 102. The training locations 105 represent example locations for which coverage properties are available and for which training data records may be created as has been described in detail above. In the example of Figure 17 there is a relatively high density of training locations 105 in a first region generally indicated by the arrow numbered 701 in Figure 17. However, in a second region generally indicated by the arrow numbered 702 in Figure 17 there is a lower density of training locations 105.

[00191] It has been realised that when a prediction model is trained with training data records associated with unevenly distributed training locations 105, the location determinations made by the trained prediction model may be biased towards regions of high density training locations 105. Taking the example of Figure 17, if data corresponding to the depicted training locations 105 are used to form training data records and a prediction model is trained using the training data records then the resulting prediction model may be overfitted to the first region 701 and underfitted to other regions such as the second region 702. This is a consequence of the higher spatial density of training locations in the first region 701 than other regions, such as the second region 702.

[00192] Such overfitting of a prediction model to regions 701 with a higher density of training locations may result in errors in location determinations carried out using the trained prediction model. For example, a prediction model trained using training data corresponding the training locations 105 shown in Figure 17 may be used to determine the geographical location of a terminal 104 whose location is shown in Figure 17. An example location which might be determined by such a prediction model is shown by a black square labelled 120 in Figure 17. It can be seen that the determined location 120 is closer to the first region 701 (having a relatively high density of training locations 105) than the true location of the terminal 104 due to the overfitting of the prediction model to regions 701 with a higher density of training locations.

[00193] It has been realised that the accuracy of a prediction model may be improved by over- sampling and/or under-sampling available coverage data in order to form training data records having a more uniform geographic distribution than the available data. Figure 18 is a flow chart of an example method for training a prediction model for determining the geographic location of a terminal based on data representative of coverage of a mobile telecommunications network. The method may be implemented on any suitable computing device. In some examples, each method step may be implemented on the same computing device. In other examples, different parts of the method may be implemented on different computing devices which may be in communication with each other.

[00194] The method of Figure 18 is similar to the methods described above with reference to Figures 2, 7, 14 and 16. Any of the steps or features described above with reference to the methods of Figures 2, 7, 14 or 16 may also apply to the method of Figure 18. A detailed description of the same or corresponding method steps with reference to Figure 18 may be omitted and only the differences between the methods of Figure 2, 7, 14, 16 and 18 will be described in detail.

[00195] At step 1801 of Figure 18 first data representative of coverage of a mobile telecommunications network at a plurality of first measurement locations is received. Step 1801 corresponds to steps 201 , 701 , 1401 and 1601 of the methods of Figures 2, 7, 14 and 16 respectively and the data received at step 1801 may include any of the features described above with reference to Figures 2, 7, 14 or 16. The first data may include coverage properties (e.g. received signal power, received signal quality and/or a measure of propagation time) for each measurement location.

[00196] The first plurality of measurement locations in the method of Figure 18 may correspond with the plurality of training locations in the methods of Figures 2, 7, 14 or 16. The measurement locations represent locations for which coverage properties have been measured or otherwise determined. The measurement locations therefore generally represent locations at which at least one device has been situated in order to measure or otherwise determine coverage properties at that location. An example of measurement locations (which may not be uniformly geographically distributed) are the locations 105 depicted in Figure 17.

[00197] At step 1802, second data representative of coverage of the mobile telecommunications network at a second plurality of training locations is generated. The second data is based on the first data. The second plurality of training locations are more evenly geographically distributed than the first plurality of measurement locations. At least some of the second plurality of training locations may be the same as at least some of the first plurality of measurement locations. For locations which belong both to the first plurality of measurement locations and the second plurality of training locations, the generated second data for those second training locations may simply comprise the first data received for the corresponding measurement location. That is, at least some of the generated second data may correspond directly to at least some of the first data for corresponding measurement and training locations.

[00198] At least one of the second training locations may not be included in the first measurement locations. That is, the second data may include data for at least one second training location for which there is no directly corresponding data in the first data received at step 1801 . For such second training locations second data may be generated for the second locations based on first data at nearby first measurement locations. For example, in regions (such as the second region 702 depicted in Figure 17) where there is a relatively low distribution of first measurement locations, coverage properties (included in the first data) at the first measurement locations in that region may be interpolated to generate coverage properties at second training locations for which there is no corresponding first measurement location. Such second training locations may, for example, be situated in between nearby first measurement locations and the coverage properties at the nearby first measurement locations may be used to interpolate corresponding coverage properties at the second training locations. Such a technique amounts to adding additional data at additional second training locations relative to the first data available for the first training locations and may therefore be referred to as a form of oversampling.

[00199] Additionally or alternatively, coverage properties included in the first data for at least one first training location may be repeated a plurality of times in the second data for the same second training location. For example, for a first measurement location situated in a region (such as the second region 702 depicted in Figure 17) where there is a relatively low distribution of first measurement locations, coverage properties (included in the first data) at the first measurement location may be added to the second data a plurality of times (e.g. twice or three times). Such added data may be associated in the second data with a second measurement location which directly corresponds with the first measurement location with which the data is associated in the first data. Such a technique also amounts to adding additional data to the second data relative to the first data available for the first training locations and may also therefore be referred to as a form of oversampling.

[00200] Additionally or alternatively, coverage properties included in the first data for a least one first training location may not be included in the second data. For example, for a region (such as the first region 701 depicted in Figure 17) where there is a relatively high distribution of first measurement locations, coverage properties (included in the first data) for at least one first measurement location may be omitted from the second data. Such a technique amounts to not including data available at one or more first measurement locations in the second data and may be referred to as a form of under-sampling.

[00201] By using over-sampling and/or under-sampling techniques (such as those described explicitly above and/or other techniques) second data is generated at second training locations which are more evenly geographically distributed than the first measurement locations. A more even geographic distribution may mean, for example, that a variance or standard deviation of density per unit area or volume for different regions may be less for the second training locations than for the first measurement locations. For example, the cell shown in Figure 17 may be divided into a plurality of regions and a density per unit area or volume of first measurement training locations may be determined in each region. The densities of first measurement training locations in the different regions have a first standard deviation or variance. A corresponding density per unit area or volume of second training locations may be determined in the same regions. The densities of second training locations in the different regions have a second standard deviation or variance. Generally if the second training locations are more evenly geographically distributed than the first measurement locations then the second standard deviation or variance is less than the first standard deviation or variance.

[00202] At step 1803, training data comprising a plurality of training data records is formed, where each training data record is associated with a training location of the second plurality of training locations (the second plurality of training locations for which the second data is generated at step 1802). Step 1803 corresponds to steps 202, 703, 1402 and 1603 of the methods of Figures 2, 7, 14 and 16 respectively and the training data records and their formation may include any of the features described above with reference to Figures 2, 7, 14 or 16. Each training data record may include at least the second training location with which the training data record is associated (e.g. the latitude, longitude and/or altitude of the training location) and the generated data representative of network coverage at that training location.

[00203] At step 1804, a prediction model for determining the geographic location of an electronic device is trained using the training data records formed at step 1803. Step 1804 corresponds to steps 203, 704, 1404 and 1604 of the methods of Figures 2, 7, 14 and 16 respectively and the prediction model and the training of the prediction model may include any of the features described above with reference to Figures 2, 7, 14 or 16.

[00204] It has been found that, for at least some situations, over-sampling and/or undersampling the data available for first measurement locations to generate second data at more evenly distributed second training locations can improve the accuracy and/or reliability of location determinations carried out using a prediction model trained with the generated second data. Such an improvement in accuracy and/or reliability may in particular be achieved when the measurement locations for which coverage data is available are relatively non-uniformly distributed.

[00205] Various methods have been described above with reference to Figures 1 -16 for training one or more prediction models for determining the geographic location of a device. In some examples, features from two or more of these methods may be combined to train a prediction model. For example, the inclusion of a serving cell location in the training data records as described above with reference to Figure 7 may be used to train a prediction model which is specific to a serving cell (e.g. using only training locations associated with the same serving cell) as described above with reference to Figure 14. Furthermore, as part of the same training process, the training locations used may be filtered to remove training locations at which a propagation time (e.g. timing advance) exceeds a threshold as was described above with reference to Figure 16. That is, features of the methods of Figures 7, 14 and 16 may be combined to train the same prediction model. In general any of the features described herein may be combined to train the same prediction model.

[00206] As was explained above a number of different features and/or combinations of features described herein may be used to train a prediction model for determining the geographic location of a device. In some examples, a plurality of different prediction models may be trained which are all capable of determining the location of the same device. For example, a first prediction model may be trained which is specific to a serving cell (e.g. using only training locations associated with the same serving cell as described above with reference to Figure 14), by including a serving cell location in the training data records (as described above with reference to Figure 7) and by filtering training location to remove training locations at which a propagation time (e.g. timing advance) exceeds a threshold (as was described above with reference to Figure 16). A second prediction model may be trained using training data associated with a plurality of different serving cells (i.e. which is not specific to a single serving cell) by also including a serving cell location in the training data records (as described with reference to Figure 7) and by also filtering training locations at which a propagation time exceeds a threshold (as described with reference to Figure 16). [00207] The first and second prediction models may both be implemented to determine the location of the same device. The different training data records used to train the first and second prediction models may mean that a first location determination made by the first prediction model may be different from a second location determination made by the second prediction model.

[00208] In some examples, an ensemble method may be used to combine two or more prediction models to provide a single location determination. For example, ensemble methods such as maximum voting, averaging and/or weighted averaging may be used to combine the output of two or more prediction models (e.g. the output of the first and second prediction models described above) to provide a single location determination. In some examples, more advanced ensemble methods such as stacking, blending, bagging and/or boosting may be used to combine two or more prediction models to provide a single location determination. For example, the first prediction model and the second prediction model may be combined using a bagging method to provide a single location determination.

[00209] It has been found that using ensemble techniques to combine a plurality of prediction models trained using different features and/or combinations of features can significantly improve the accuracy of location determination when compared to, for example, using a single prediction model to determine the location of a device.

[00210] Various methods have been described above for training and implementing prediction models for determining a geographic location of a device. Such methods may find particular application and advantages in the determination of the location of airborne devices (such as network connected drones) situated at above ground locations. As was explained above, the availability of training data for above ground locations may be relatively limited. Furthermore, location determination at above ground locations involves an extra dimension of altitude, when compared to determining the location of ground based devices. Consequently, the accuracy of location determinations of airborne devices may be lower than corresponding location determinations of ground based devices. It has been found that the various methods and features described herein, whether used alone or in combination, can significantly improve the accuracy with which the location of an airborne device situated at an above ground location can be determined. However, the methods and features described herein may also find application in the determination of the location of one or more ground based devices.

[00211] Various computer implemented methods and devices have been described above. In general any method described herein may be implemented on one or more electronic devices. Figure 19 is a schematic illustration of an example electronic device which may be used to implement all or part of any method described herein. The general structure of the device depicted in Figure 19 may be applicable to any terminal, base station, network node and/or any other electronic device contemplated herein.

[00212] The device 1000 may include at least one processing unit 1001 , memory 1002 and an input/output (I/O) interface 1000. The processing unit 1001 may include any suitable processer and/or combination of processors. For example, the processing unit 1001 may include one or more of a Central Processing Unit (CPU) and a Graphical Processing Unit (GPU). The memory 1002 may include volatile memory and/or non-volatile/persistent memory. The memory 1002 may, for example, be used to store data such as an operating system, instructions to be executed by the processing unit (e.g. in the form of software to be executed by the processing unit), configuration information related to the device 1000, session information and/or configuration or registration information associated with any other device, node or module in the network. For example, the memory 1002 may be used to store data representative of coverage of a mobile telecommunications network at one or more training locations and/or to store parameters of a trained prediction model. In some examples, the memory 1002 may be used to store instructions for executing any of the methods disclosed herein.

[00213] At least the processing unit 1001 is connected to an input/output (I/O) interface 1003. The I/O interface 100 facilitates communication with one or more other devices, network nodes or modules in a network. For example, the I/O interface 1003 may be operable to transmit and/or receive communications to/from other devices in a network. In some examples, the I/O interface 1003 may be operable to transmit and/or receive communications over an air interface. For example, the I/O interface 1003 may include a transmitter and/or a receiver for transmitting and/or receiving wireless communication (e.g. radio frequency signals). In some examples, the I/O interface 1003 may include a transceiver configured to receive and transmit wireless communication (e.g. radio frequency signals). In some examples, the I/O interface

1003 may be operable to additionally or alternatively communicate over one or more wired connections.

[00214] Optionally, the device 1000 may further include a display 1004. For example, where the device 1000 is a UE, the UE may include a display 1004 for displaying information to a user of the UE. The display 1004 may comprise any suitable electronic display such as a touch sensitive display. The display 1004 may be connected to at least to the processing unit 1001 . The processing unit 1001 may generate display signals which are sent to the display

1004 in order to cause the display information.

[00215] In general the methods disclosed herein may be implemented on any suitable computing device and/or combination of computing devices. Generally methods for training one or more prediction models may be implemented on one or more fixed computing devices such as one or more servers. For example, one or more devices which form part of a core network may implement methods for training one or more prediction models. Additionally or alternatively, one or more devices such as a server which does not form part of the core network but is in communication with the network (e.g. capable of receiving data from the network) may be used to implement method for training one or more prediction models.

[00216] Methods for implementing one or more trained prediction models to determine the location of a device may be implemented on the device for which the location is being determined. For example, such methods may be at least partly executed on a terminal device operating in the network. Additionally or alternatively, methods for implementing one or more trained prediction models to determine the location of a device may be implemented by one or more devices which form part of a core network and/or are in communication with the network. For example, a terminal device operating in the network may report one or more coverage properties at its current location to the network (e.g. the core network). The network may use the reported coverage properties to implement one or more trained prediction models to determine the location of the device. The determined location of the device may be reported to the device over the network.

[00217] In the interest of conciseness not all possible alternatives which fall within the scope of the present disclosure have been explicitly discussed herein. As the skilled person will appreciate, in the present disclosure any aspect discussed from the perspective of an element being operable to do an action also discloses the same feature from the perspective of a method including a method step corresponding to the action. Similarly, any discussion presented from the perspective of a method step also discloses the same features from the perspective of any one or more suitable elements being operable or configured to carry out some or all of the method step. It is also considered within the present disclosure that for any method step(s), there can be a computer program configured to carry out, when executed, the method step(s).

[00218] Within the context of the present disclosure a device, such as a terminal, a base station, or a network module or node, or server is generally considered from a logical perspective, as the element carrying out the appropriate function. Any such device may be implemented using one or more physical elements as deemed appropriate. For example, it may be implemented in one (or more) of: a standalone physical device, in two or more separate physical devices, in a distributed system, in a virtual environment using any suitable hardware or hardware combination, etc. [00219] It will be appreciated that embodiments of the present invention can be realised in the form of hardware, software or a combination of hardware and software. Any such software may be stored in the form of volatile or non-volatile storage such as, for example, a storage device like a ROM, whether erasable or rewritable or not, or in the form of memory such as, for example, RAM, memory chips, device or integrated circuits or on an optically or magnetically readable medium such as, for example, a CD, DVD, magnetic disk or magnetic tape. It will be appreciated that the storage devices and storage media are embodiments of machine-readable storage that are suitable for storing a program or programs that, when executed, implement examples of the present disclosure. Accordingly, examples provide a program comprising code for implementing a system or method as claimed in any preceding claim and a machine readable storage storing such a program. Still further, examples of the present disclosure may be conveyed electronically via any medium such as a communication signal carried over a wired or wireless connection and examples suitably encompass the same.

[00220] Features, integers, characteristics, or groups described in conjunction with a particular aspect, embodiment or example of the invention or disclosure are to be understood to be applicable to any other aspect, embodiment or example described herein unless incompatible therewith. All of the features disclosed in this specification (including any accompanying claims, abstract and drawings), and/or all of the steps of any method or process so disclosed, may be combined in any combination, except combinations where at least some of such features and/or steps are mutually exclusive. The invention is not restricted to the details of any foregoing examples.