DEVICES

FIELD OF THE INVENTION

The present invention is in the field of computer networks. In particular, the present invention relates to a network comprising a plurality of nodes and a server or distribution of servers for providing web services.

RELATED PRIOR ART

Recently, there has been a discussion of utilizing network infrastructure to improve the way cities operate with regard to consumption of energy, traffic control, public transportation, air pollution and the like. These ideas have sometimes been referred to "wired cities" or "smart cities" in the art. The rational behind this concept is to make better use of limited resources available in a city, in particular space (e.g. for parking or transportation) and energy to name but two.

However, in "smart" or "wired" cities two general problems arise. The first problem is how to obtain the data that is meaningful for characterizing the "behavior" or "status" of the city as a highly complex system. The second problem is how to handle the vast amount of data that will usually be produced such as to obtain meaningful information therefrom with feasible computation at effort and computation time.

SUMMARY OF THE INVENTION

The present invention has been devised in consideration of the above two problems. The present invention provides a network comprising a plurality of nodes and a server or distribution of servers for providing web services, wherein at least a subset of the nodes are associated with outdoor lighting devices, in particular street lights, and comprise:

- one or more sensors, and

- communication means allowing for communication with other nodes and with said server or distribution of servers. Further, the server or distribution of servers is configured to

- build a model based on received data,

- compare received data with the current model, and

- update the model according to received data.

The inventors have recognized that outdoor lighting devices, such as street lights, are in many respect ideal objects for providing network nodes, for several reasons. First of all, street lights are approximately evenly distributed throughout a city, so that the network nodes can be evenly distributed throughout the city as well. A further advantage of streetlight is that it is per se connected to a power supply so that it can easily provide the electrical power needed to operate the network node and the one or more sensors associated with it. Also, since streetlights usually stand clear from other objects, they are an ideal place for placing different types of sensors, e.g. optical sensors for monitoring the nearby street. Finally, one aspect of a "smart city" would be to provide for an intelligent power management for the street lighting itself, so for this season too it is advantageous to include the street lights as nodes in the network themselves.

Further, the server or distribution of servers is configured to build a model based on the received data, including the measurement data obtained by the sensors, to compare the received data with the current model and to update the model according to the received data. As will be explained in more detail below, this construction and update of the model can be carried out by various machine learning methods, preferably including Support Vector Machines, and can be used to generate the computation learning model of the entire network. That is to say, the learned model may represent the current status of the complete network of nodes, or part of the network of nodes, including a single node only. An important advantage of a machine learned model for the network behavior is that it allows for a fast, typically realtime classification of any new data. Further, any information about the network can be obtained by querying the model, instead of analyzing a huge amount of data gathered from the network, therefore allowing for a real time-analysis of the network.

In a preferred embodiment, the server or distribution of servers further provides an interface for assessing and querying the model. Herein, "querying the model" means any way of interacting with or analyzing the model such as to retrieve any information of interest. Again, querying the model is much more efficient than for example trying to analyze the vast amount of data provided by the multitude of sensors.

Preferably, at least a subset of the nodes is provided with sensors for sensing one or more of the following conditions:

- energy consumption of street lamps, lighting conditions on the street, environment temperature, humidity, air pressure, C02 concentration, NHx concentration, Ozone concentration.

Further, preferably at least a subset of the nodes comprises one or more of the following devices:

- a color detector, an optical detector, a GPS device, a sound detector, an accelerometer, a gyroscope, a compass, a gas sensor, a Hall effect sensor, a vibration sensor, a motion sensor, an electronic range finder, and a capacitive or inductive touch sensor.

With these types of detectors and sensors, various kinds of information can be gathered that is indicative of complex processes of interest and can be introduced to the model by way of machine learning. Note that some of these detectors and sensors are directly related with the operation of street lamps, such as a sensor for detecting the energy consumption of a street lamp or a detector for detecting the light conditions on the street. Other sensors may provide information that is only indirectly related with the operation of street lamps such as temperature, humidity and air pressure, which in combination are related to the visibility (e.g. fog vs. clear view), that can be accounted for in an appropriate model as well. Finally, the nodes may comprise sensor's or detector's information which at first sight are not related to the street lighting at all, such as optical detectors or motion sensors which will e.g. allow for example estimating the traffic. However, one of the advantages of model building by machine learning is that correlations between parameters that at first sight may appear unrelated will be reflected in the model, such as an increased traffic under certain conditions and an increased demand for street lighting.

In a preferred embodiment, at least a subset of the nodes comprises actuator means, in particular actuators for adjusting the color of street lighting or the position of the illumination spot provided by the streetlight. By providing this type of actuators, the individual nodes can make practical, physical use of the knowledge about the network state represented by the model by corresponding operation of the actuators.

Preferably, at least a subset of the nodes comprises computation capabilities and preferably employs a local model derived from the model developed by the server or distribution of servers. This way, the network nodes can make direct use of the local information, i.e. information from the node's own sensors or information from sensors of nearby nodes. By means of such local models, the individual nodes acquire cognitive abilities of their own.

In the preferred embodiment, at least a subset of the nodes associated with the outdoor lighting devices has a modular sensor design, the modular sensor design being characterized by comprising plural receiving sections for selectively receiving one out of a plurality of sensor modules and/or communication modules. The rationale behind this modular sensor design is that a large variety of sensors may in general be desirable for characterizing the status of the network, but that only a limited choice of sensors is actually required at any particular node. Based on a modular design, it is possible to provide exactly the selection of sensors at any given node that is needed, by simply providing the appropriate selection of sensors at the node's receiving sections.

In a preferred embodiment, the communication means of the node comprise one or more of the following: ZigBee, 6L0WPAN, IEEE 802.15.4, Blue Tooth, WLAN, GSM/GPRS, Ethernet or proprietary communication means.

While for the reasons given above street lights are regarded as ideal locations for nodes of the network of the invention, further nodes at other locations may also be part of the network. In particular, the network may include further nodes provided in vehicles. In fact, vehicles such as cars are particular useful network nodes, because drivers are typical recipients of information provided by the network, for example navigation instructions, traffic information, parking space directions and the like, and because cars are at the same time particularly suitable for monitoring the traffic flow and providing corresponding information to the network.

A further example of network nodes not associated with streetlights are nodes which are provided in operative connection with municipal supply systems, in particular one or more of a water supply system, a waste water system, a high power cable, a low power cable, a transformer, a data cable, a data switch, a gas supply system and a district heating system. Herein, the communication means of a node associated with a streetlight may serve as a gateway to the network for a node provided in a vehicle or a node provided in operable connection within municipal supply systems. That is to say, the streetlight-nodes may provide the "backbone" of the network infrastructure to which other nodes can be coupled.

Preferably, the server or distribution of servers comprises at least one machine learning algorithm, where the machine algorithm includes one or more of the following: support vector machines (SVM), Bayesian network, reinforcement learning, principal component analysis, clustering, association rule learning, decision tree learning. Using this type of machine learning algorithms, the model can be built and updated efficiently and automatically.

Preferably, the server or distribution of servers is adapted to clean data received from the nodes such as to exclude errors in sensor reading and/or communication errors. This cleaning may for example be performed using semi-supervised learning methods with background common-sense knowledge. This way, errors in the model due to erroneous sensor values can be efficiently avoided.

Preferably, the server or distribution of servers is adapted to extract information from the data received. In particular, the server or distribution of servers may be adapted to extract information about the energy efficiency from received data related to energy consumption, lighting and external factors. Based on this extraction, a more efficient use of energy can be achieved.

In a preferred embodiment, the model is presented as a multi-dimensional matrix. In particular the model may be developed in the form of a hyperplane in a multi-dimensional space. Herein, the model allows for classification of received data with regard to the location of corresponding data points with regard to said hyperplane in multi-dimensional space. An example of such hyperplanes are hyperplanes constructed by Support Vector Machines in a high or infinite dimensional space which can be used for classification or regression in a way per se known in the art.

In a preferred embodiment, the server or distribution of servers is adapted to develop a plurality of models and storing the same. Herein, the server or distribution of servers is preferably adapted to store data that triggered deviances in the model. While it is generally advantageous to deal with the model rather than the vast amount of data underlying the model, for analysis purposes it may be advantageous to take notice of the data that triggered significant changes to the model.

Further, the server or distribution of servers is preferably adapted to analyze data received from the network nodes with reference to the available model or models in real time.

In the preferred embodiment, the server or distribution of servers is configured to receive, in addition to data from the network, information from external information sources, in particular weather conditions and/or traffic information. Further, the interface for assessing and querying the model is preferably adapted to allow one or more of the following: status monitoring, keyword-based searching and browsing, complex events browsing, alerting and searching, visualization of short-term trends, visualization of long-term trends, visualization of predictions.

In a preferred embodiment, the server or distribution of servers is configured to optimize energy consumption of the outdoor lighting based on a model reflecting information from the past and current information received from network nodes. It has been confirmed by the inventors that this way the energy consumption of the outdoor lighting can be dramatically decreased without compromising the lighting quality at any particular point in time.

In a preferred embodiment, the server or distribution of servers is configured to compare a recent model with at least one previous model and to detect an anomaly if the difference between the two models exceeds a previously defined or an expected threshold. Herein, the server or distribution of servers may be configured to report an error, a warning or an information message in response of detecting said anomaly.

However, the network is not only suitable for detecting anomalies, but also suitable for predicting the behavior of the network or subset of nodes of the network. In particular, the prediction may be based on historic models from the past and current data received from the nodes and/or from external information sources.

In particular, the prediction may be based on one or more of the following:

- parametric modeling, in particular autoregressive and moving-average-modeling,

- non-parametric modeling, in particular wavelet-based modeling, and

- regression models. In particular, the server or distribution of servers may be configured to predict the consumption of energy of individual nodes or the entire network of nodes for a period of time ahead, based on past models.

In a further preferred embodiment, the server or distribution of servers is configured to predict traffic jams or to predict air pollution. In yet a further preferred embodiment, the server or distribution of servers is configured to predict failure of single devices associated with a node, failure of a single node or a group of nodes. Herein, the server or distribution of servers may be configured to report a prediction of failures as a rank list of failures that are most likely to occur in a given period of time. The failure prediction is preferably based on an analysis of historic models of the past, technical data of the related devices and detected trends.

In a preferred embodiment, the server or distribution of servers is configured to monitor the evolution of the updates of the model to determine trends in the behavior of the network or parts of the network. Herein, the server may be configured to determine micro-trends related to the behavior of individual nodes or a subset of nodes and/or macro-trends related to the behavior of the entire network.

In particular, the server or distribution of servers may be configured to monitor the time evolution of trends and to compare this time evolution with an expected trend, wherein the server or distribution of servers may be configured to issue a warning if a divergence between the monitored trend and the expected trend is detected, that exceeds a predetermined threshold. This again is a very sensitive way of detecting anomalies or imminent failures in the intervals at an early stage where it is sill possible to take countermeasures.

In a further embodiment, the server or distribution of servers is configured to extract patterns of information that frequently emerge. In addition or alternatively, the server or distribution of servers may be configured to model the behavior of individual nodes.

The present invention further relates to a method of processing data according to claim 44 and a network node according to claim 46.

BRIEF DESCRIPTION OF THE FIGURES

Fig. 1 is a schematic illustration of elements of the network of the invention. Fig. 2 is a schematic diagram illustrating the mechanics for an SVM machine learning algorithm that is employed in one embodiment of the present invention.

Fig. 3 is a schematic drawing illustrating a divergence between an expected trend and observed measurements.

Fig. 4 is a block diagram illustrating components of the network according to an embodiment of the present invention.

Fig. 5 is a diagram similar to Fig. 5, wherein the structure of an individual node is further illustrated.

Fig. 6 is a schematic figure showing the integration of nodes that are in operative connection with municipal supply systems with nodes associated with street lights.

DESCRIPTION OF THE PREFERRED EMBODIMENT

In a preferred embodiment, a node in a network is composed of a set of sensors, communication means and computational abilities. In a subset of nodes that are associated with the outdoor lighting devices, sensors provide to every node information about energy consumption, lighting and environment conditions (see Fig. 1). The data being measured in a basic sensor set is energy consumption, lighting on the street, environment temperature, humidity, air pressure, C0 ₂ rate and NHx rate. Other sensors that can be included are Colour light, Optical Detector, GPS, Sound, Accelerometer, Gyroscope, Compass, Gas sensor, Hall effect, Vibration, Motion, Ultrasonic range finder, Capacitive/inductive touch sensors. The sensor information is transmitted to the server or distribution of servers and between neighbouring network nodes by using various communication means like ZigBee, 6LoWPAN,IEEE 802.15.4, Bluetooth, WLAN, GSM/GPRS, ETHERNET, or Proprietary communication means.

On the server side, the information received from plural nodes is processed with a series of machine learning algorithms including Support vector machines, Bayesian networks, Reinforcement learning, Principal components analysis, Clustering, Association rule learning, and/or Decision tree learning. While reference is made to "a" server in the following, it is understood that a distribution of servers, such as a server cloud may be employed. First, the data is automatically cleaned by using semi-supervised learning methods supported with background common-sense knowledge to exclude obvious errors in sensor reading and/or communication errors. The data gathered can be in a form of different modalities, such as: structured numerical data, unstructured textual data, picture data, streamed video data and other data streams. In the preferred embodiment, several methods for information extraction are employed including name entity extraction, co-reference, relation extraction, SIFT features and others. After the data from all modalities is presented in a common description, the cleaning is carried out. For example, if two temperature outdoor sensors that are only e.g. 25m apart have measured two values that are significantly different, it is concluded that one of the measurements was wrong. Because of the network density, the wrong measurement can be identified and excluded in the course of the cleaning .

After the data has been cleaned, the data is fused into meaningful information. The measured data about energy consumption, lighting and external factors generates coherent information for the energy efficiency in respect to the effect (i.e. lighting value) and environmental information.

Such information is used for further processing and model construction. Various machine learning methods, preferably including support vector machines can be used to generate the computation learned model of the network (see Fig. 2). A model that represents the whole network of nodes, specifics of a subset of nodes or specifics of a particular node is presented as a multidimensional matrix with many multidimensional vectors. The learned model shows the current status of the complete network e.g. outdoor city lighting network, the subset of nodes, e.g. one particular street in a city or a model of a particular node, respectively. The first learned model in a setup process is the default model representing the normal situation. The next instance of the learned model in time represents the deviation of the previous model, the next the deviation of the previous and so on. Based on the deviances that are based on comparing a previous model with the current information from the network, the previous model is updated with a new one.

Fig. 2 schematically illustrates the mechanics of Support Vector Machines (SVM) machine learning algorithms. The algorithm develops a model in the form of a hyperplane based on the learning data. A developed model allows for fast classification of new data. Fig. 2 illustrates a simplified case of classification of two classes based on the two values (categories). Preferably, all models are stored to be used for further analytics, in particular time dependent analytics like trend detection and prediction.

The server stores two types of information, namely built models that are used for the real time processing and analytics, and the data that triggered the deviances in the models for long term off-line analytics. Built models cannot be represented in formalised and structured way as "if- then" rules, but are rather represented as a multidimensional matrix. In respect to the top- down codified rules, the machine learning methods applied herein allow for modelling the vast complex space of multisensory nodes network.

Since the introduction of machine learning methods allows handling mass amount of data by interacting in the analysis only with models and not with the data gathered from the network, this approach allows to process a mass amount of information in real-time.

The server stores models that are than being used for further analysis and querying. The server also provides interfaces for status monitoring, simple and keyword based searching and browsing, complex events browsing, alerting and searching, short and long term trends visualisation and prediction visualisation. Different interactive visualisation methods are implemented to allow various views on the developed models, the methods also include geographically and temporarily information and information received from external and public sources.

In the preferred embodiment, network nodes have cognitive abilities of their own. Accordingly, every network node preferably has abilities to locally behave as an independent entity that is able to respond accurately on the information received from the environment. Data gathered from several sensors, data received from the neighbouring nodes and from the server is processed by the node's own processing capabilities to form a local control loop. The network node may pre-process the information that is gathered locally and adopt a local developed model that has been sent to the node by the server. The local model can be "localised" from the global model by the server. Using integrated causal and temporal reasoning methods, a node can decide how to react to information gathered from local sensors that significantly modify the current local model. For example, by receiving information from the neighbouring light about its local problem of increased heating, the light can decide to increase its lighting power in order to give the neighbouring light enough time to communicate its problem to the server and to be repaired. On a basic level of the possible embodiments of the present invention, developed models at the server side, current data from the network and external information sources like weather conditions, traffic information etc. form input information for the analytics, optimisation, prediction and simulation features.

The server may use the information from the past in the form of models together with the current information to develop an optimisation model for energy consumption in the outdoor lighting. Based on the measured lighting environment and the past occurrences, the server may propose an optimized energy consumption scenario. For example, consider that the model at the server indicates that during the period from November to February, traffic in a particular zone is usually increased due to the skiing season. The model may further indicate that based on the previous similar weather condition as the current one, the light has to be switched on for additional 30 min as compared to normal to decrease the risk for traffic accidents. Furthermore, local light's cognitive abilities enables to react on the information at the respective location. For example, if there is no traffic, the lights may be dimmed by 50%.

The difference between the current model and the previous model is spotted as the difference in the network behaviour. If the difference exceeds an expected or pre-defined difference, this is regarded as an anomaly that server has to report. For example, the failure of a lamp, decreased traffic due to the changed traffic regime in the city, increased value of C0 ₂ or NHx is defined as anomaly. Based on the detected anomalies that exceed certain threshold, the server reports an error, warning or information.

Another example is finding anomalies in the road segments that are then used for optimal route planning. For each segment one can determine the tracks that are anomalous. Removing the anomalous tracks and computing a set of features for each path can enable a global path based analysis. The analysis can, however, be applied only to a set of short road segments that are well represented with a large number of tracks. It has been found that low numbers of tracks that match a particular segment can lead to unreliable analysis.

Given the information of the global network behaviour, the behaviour of local network subsets and individual nodes from the past, and given the technical data related to specific devices, the server is able to predict the forthcoming failure of the nodes, devices and the network. Predictions are reported as a ranked list of failures that might occur in the next period (hours, days, weeks). The severity measure is calculated on the basis of the historical models, technical data and detected trends. On a more advance level, using stored models and the current information from the network can be used by the server to perform simple and deep analytics by combining machine learning methods with semantics and reasoning methods.

By monitoring model evolution (updates) during the time, the server may calculate micro and macro trends in the complete network, in the subset of nodes (sub-network) and on a particular node or device. Fig. 3 schematically illustrates the method for detecting and calculating unexpected trends. Micro trends are trends that cannot be spotted by using only traditional analytic techniques. Micro trends are usually local and are not detected in the complete network. Examples of micro trends include energy consumption increase in a subnetwork, decrease of the lighting intensity in a particular light over the time, increase of the air pollution (e.g. a combined value of different variables like C0 ₂ , NHx, Ozone, etc.). Macro trends are the ones that are usually detected on the complete network or on large subnetworks. Examples include detecting changed environmental data due to pollution, drivers behaviour in a street, changed energy consumption due to economic or environmental influences, change in the topology of a network due to a different lighting policies, etc.

Detected trends are being further monitored by the server. When the trend reaches the critical point in a predefined warning range, the server reports a critical warning to the monitoring service. The server ranges all detected trends in respect to their severity and lists them accordingly.

Using the historical models and the current data, the server may calculate the predicted behaviour of a network, subset of nodes (sub-network) or a particular device. In respect to the monitored and calculated trends, the server may in particular also employ an interaction between different models. Similarly to trends, the server may calculate macro predictions and micro predictions. Exemplary methods used in prediction are: parametric modelling (e.g. autoregressive models, moving average models), nonparametric modelling (e.g. wavelet based), and regression approaches. In the preferred embodiment, the server mostly uses a data mining oriented regression approach.

An example of a micro prediction is a prediction of the energy consumption for the next period (hour) that is calculated from the past models. Another example is the prediction of light (or other device) failure that is calculated based on the stored models from the past, detected trends on the light, current data and light technical specifications. An example of the macro prediction is the prediction of the energy consumption of the complete network over larger time scales (several weeks, months). A further example is the prediction of network failures that are caused by failure of dependent devices. A model developed by the server may also include information about device dependencies that cause cascade failure effects in the network. Even more complex scenarios can be handled, like predicting traffic jams, community infrastructure network failures, environment pollution prediction etc.

Knowing the model, the current measurement data, the model's dependency on external factors like data measured from natural environment, social environment data, global factors data etc, the server is able to calculate the optimal outcomes by using optimisation methods. Such optimisation methods may be chosen from simple mathematical optimisation methods to complex constraint logic programming methods and forward chain reasoning. The server may calculate different optimisation plans, evaluate them and propose the ranking list of potential scenarios. For example, based on the predictive energy consumption, the server may calculate the optimal scenario given the expected external conditions. This optimisation may include energy optimisation, road traffic optimisation, lighting optimisation, network devices optimisation etc. Several criteria that are included in the developed models are taken into consideration, primary and secondary natural environment factors, primary social environment factors, economic factors, wellbeing factors etc.

By using predefined common-sense knowledge, the server can describe a decision that has been taken in the modelling phase. "A common sense knowledge base" is available as a network of concepts that are connected through many different types of relations. The server may use predefined knowledge to reason about the developed model, gathered information and decisions taken. Nodes that are able to perform some of the cognitive abilities make own plans in a limited context that is provided by the server. The server thus models the network and node's behaviour. For example, the server knows from the developed model how a particular node in the network behaves under particular sensed conditions. For example, when there is a need for higher luminance on the street, particular lights coordinate their behaviour. The server also knows how the sub-network or network as a whole behaves in particular conditions.

The server may use knowledge extraction methods to extract meaningful patterns from the data gathered. A pattern is a set of similar information that emerges very often. The server may transform the patterns to the formalised descriptions that are inserted in the common- sense knowledge base. Formalised descriptions are in the form of "if-thenrules" or semantic networks presented in RDFs. The server will typically use data mining and text mining algorithms for knowledge extraction and association rules for knowledge formalisation.

In Fig. 4, the architecture of the network and method of the invention is illustrated. Reference sign 10 denotes a geographic area, where a plurality of nodes are provided at street lights, and further nodes may be provided, such as on-board devices of vehicles or municipal supply systems. The sensor data provided by the nodes is communicated to the server or distribution of servers which is adapted to carry out the steps described above, i.e. the cleaning and fusion of data, model building, anomaly detection, trend detection, prediction, knowledge discovery, reasoning and context generation.

Further, an interface 14 is provided for assessing and querying the model developed by the server or distribution of servers 12. Via this interface 14, numerous GUI applications can query the model in any desired way, such as for decision making support, visualisation and the like.

Fig. 5 is a diagram that is partly identical with the diagram of Fig. 4. However, instead of the geographic region, in Fig. 5 a node 16 is symbolically represented. As explained above, according to the present invention at last some of the nodes 16 are associated with street lights, symbolically represented at reference sign 18, and comprise a modular sensor 20. Further, the local node 16 has local computation capability. Together with a local model, that has been localized from the global model by the server, the node 16 acquires cognitive abilities of its own. To illustrate this, in Fig. 5 the difference between the "global intelligence" held by the server 12 and the "local intelligence" acquired by the individual node 16 is highlighted.

Nodes in the network can also perform actions and can provide wireless access points for further devices. Several additional scenarios are to be implemented in the network.

By installing cameras on the network nodes scenarios for security, traffic control and different monitoring systems can be developed.

By installing nodes on other infrastructures like waste water management, water pipes, gas distribution network, electricity network, telecommunication network and the like, existing network allows for integrated city infrastructure management. This is illustrated in Fig. 6. By installing wireless access points of either 3G or 4G, the network can also be used to automatically connect mobile devices. Mobile devices can be either personal or public. Similarly, if nodes are installed in cars and other vehicles, they can be automatically connected to the network.

By installing active projectors on lights, it becomes possible to project information directly into the environment e.g. by employing the light source of the street light that is available anyhow. Such setups can be used for marketing purposes, helping driving by directing drivers and support for automated driving.

In addition, actuators can be installed and managed by the network. For example, Pulse Width Modulation can be employed that manages the colour and frequency of the RGB LED lights or the actuation of a motor. Possible scenarios include light shows, or the modification of the colour of the light in extreme weather conditions on the roads, such as to enhance visibility. Servo motors may be employed to move the lighting device. Further scenarios include controlling the light dispersion, moving lighting to a needed spot, wave lighting function, etc.