CLUSTER ANALYSIS

The present invention described herein generally relates to systems for monitoring fluids such as water, oil, gases and other fluid products in a fluid supply network

Pipeline networks are the most economic and safest mode of transportation for fluids like water, oil, gases and other fluid products. As a means of long-distance transport, pipelines have to fulfill high demands of safety, reliability and efficiency. If properly maintained, pipelines can last indefinitely without leaks. The leaks in the pipeline network can be caused by various reasons : damage from nearby excavation equipment, corrosion of pipes, accidents, earth movement, cyclic loads, related to changes in the network operation mode, etc.

Such a system of interconnected pipes that carries a pressurized fluid such as water, oil, gases and other fluid products is called hydraulic supply network. When monitoring hydraulic supply networks, one often faces the task of monitoring states of fluid supply networks including leakage detection and leakage localization. If there is a problem in the network, it is very important to troubleshoot the defect in short time. Timely localization of the problem allows reducing cost of repairing and possible liquid losses on the network. Consequences of the leakage can be very destructive.

Any fluid supply network, including a water supply system, typically includes: fluid sources or fluid storage facilities such as reservoirs, tanks, etc.;

pressurizing components such as pumping stations, pumps, etc . ;

a network of pipes for transportation and distribution of fluid to the consumers.

Further the fluid supply network is considered on an example of a water supply network. A water supply system or water supply network also belongs to the hydraulic supply networks which provide water supply to different types of consumers.

The water in the supply network is maintained at positive pressure to ensure that water reaches all parts of the network, that a sufficient flow is available at every take-off point, i.e. at every consumer, and to ensure that untreated water in the ground cannot enter the network. The water is typically pressurized by the pressurizing components.

Different types of pipes are used in the pipe network of the water supply network. In general, the pipes can be classified in two categories depending on purpose : main pipes or transportation pipes, that are mainly long pipes located underground with large diameters of, for example, 300 - 700 mm, but can be of giant diameters of more than 3m, moving pressurized water from the water storage facilities into the town or district of the town. distribution pipes, that are pipes with small diameters of, for example, 80 - 300 mm, used to take the water from the main pipes to the consumers, which may be private houses as whole or each apartment individually, or industrial, commercial or institution establishments, and other usage points such as fire hydrants. By now water has become one of the most important goods in the 21st century. However, sometimes considerable water losses occur in water supply networks .

The term "water loss" is generally adopted to indicate the difference between the overall amount of water supplied into the network and the sum of the water volumes corresponding to the consumers' consumption recorded by flow meters installed on consumers ' nodes .

These water losses can be divided into two groups: the apparent losses, e.g. unrecorded water volumes used for public functions, such as cleaning of roads and urban areas, irrigation of green spaces, operation of public fountains, fire-extinguishing service, which consist of water volumes actually consumed but not accounted for, and the real losses, that are caused by damages that may have occurred to the network pipes or by the deterioration of the pipe junctions or the hydraulic devices. Real losses are the physical losses of water from the water supply system, also referred as "water leakages".

These losses put a strain on water supply and inflate the management cost for the water utilities since they represent water that is extracted and treated but never reaches the consumers . In many cases, minor water leakages deriving from the inefficient hydraulic seal of junctions or from small cracks on pipes may lie hidden for a long time, sometimes for months or even years. Major leakages can be easily observed when significant damages to the pipes occur, as they usually result in large amounts of water erupting from ground or flowing in the consumer properties.

The proven method around the world to reduce leakage from the water supply system is to proactively find the leaks before they appear at the surface. This can be achieved by monitoring the network and has the benefits of reducing the time the leaks are running, and wasting water.

The supply network monitoring can be performed using the local leak detection methods by using various equipment, such as acoustic equipment, thermograph, ground penetration radar, etc. In most cases such method requires a lot of labor since experts with the respective equipment should trace the whole supply network to find the leakage.

Also there are global methods of the supply network monitoring that are based on gathering a lot of data about the network condition such as nodal pressure and pipe flows and analyzing them. According to international and national standards, the best practice method for monitoring a water supply network is to sectorize it into district metered areas (DMA) . A DMA is an area with strict boundaries within the water supply network with measured inflow into this discrete area. This technique was first introduced at the beginning of the 90s. However the DMA' s approach has its disadvantages . Firstly, dividing the water supply network into smaller areas comes with a cost - the cost of an area survey, installation design, flow meter and chamber installation etc., is substantial, particularly if small areas (say less than 1000 consumers) are chosen. Since the main pipe diameters are big (typically more than 300 mm) , flow meters and valves to be installed on this main pipes are big and therefore expensive as well. Also, creating DMAs requires the 'permanent' closure of many boundary valves - and, because of area supply arrangements and network characteristics (such as topography and low system pressures) some networks are hydraulically difficult to divide into single-feed DMAs without disadvantaging consumers.

Most of the utilities are introducing the systems for monitoring of the network states based on the installation of the sensor networks, providing the functionality, e.g. for the control system to detect deviations of the parameters from the desired values. However setting the proper thresholds, as the most popular approach to monitor measured parameters, is in the most of the cases a complicated task requiring experience and deep domain understanding. In some cases the modeling can support in this by providing the calculated values, but compared to the model the variation of the parameters in field is usually fuzzier.

In general, the existing ways of the fluid supply network monitoring are complex, time consuming, costly, and require very sophisticated and expensive equipment . In the light of the foregoing discussion, it is evident that there is a strong need of an easy and convenient detection of a fluid supply network state. The object is solved by a method for detection of a fluid supply network state based on cluster analysis as defined in claim 1, and a system for detection of a fluid supply network state based on cluster analysis as defined in claim 7. Consequently, the present invention provides a method for detection of a fluid supply network state based on cluster analysis, where the fluid supply network comprises a network of pipes for delivering a fluid to consumers, wherein the network of pipes comprises a plurality of pipes. The fluid supply network also comprises a plurality of metering points wherein each metering point is located on a pipe and equipped with at least one sensor. For each metering point, the at least one sensor is adapted to measure at least one parameter of a fluid condition in the pipe of the fluid supply network where the at least one sensor is installed, and to provide data, which comprises information about the measured parameter.

According to the present invention in the beginning metering points are selected from the plurality of metering points for further steps. Also the method comprises at least a training mode and at least a test mode .

Within the training mode, a plurality of historical data sets is collected. The plurality of historical data sets corresponds to the provided data measured with the sensors of the selected metering points. Each one of the historical data sets corresponds to the provided data measured with the sensors at a particular point in time and different historical data sets of the plurality of historical data sets correspond to different particular points in time. After that this plurality of historical data sets are clustered to form at least one S-cluster wherein each S- cluster corresponds to a specific state of the fluid supply network. Further a one-class classifier is trained for each S-cluster. In addition to that a state of the fluid supply network is assigned for each S-cluster. Within the test mode an actual data set that corresponds to the provided data measured with the sensors of the selected metering point is collected. The collected actual data set is provided to each trained one-class classifier that corresponds to the respective S-cluster. As a result of this step a confidence value is received from the each trained one-class classifiers. Each received confidence value means with what probability the collected actual data set provided to the respective trained one-class classifier corresponds to the respective S-cluster.

Finally judgment about the actual state of the fluid supply network is made based on the received plurality of the confidence values and the assigned state of the fluid supply network for the respective S-cluster. Furthermore the present invention provides a system for detection of a fluid supply network state based on cluster analysis. According to the present invention the system comprises a plurality of metering points wherein each metering point is located on a pipe and equipped with at least one sensor. For each metering point the at least one sensor is adapted to measure at least one parameter of a fluid condition in the pipe of the fluid supply network where the at least one sensor is installed, and to provide data, which comprises information about the measured parameter. In addition to that the system comprises a control unit configured to perform cluster analysis to detect a fluid supply network state in accordance with method of any one of the claims 1 - 4.

Therefore there is a fluid supply network with the system for detection of a fluid supply network state based on cluster analysis .

The present invention is based on combined approach involving one-class classification technique and domain expert knowledge, applied for training of the one-class classifiers. In other words as soon as a plurality of historical data sets are collected they can be clustered, i.e. similar historical data sets can be grouped in S-clusters. Domain experts can assign a state to each S-cluster, i.e. to define / describe what state of the fluid supply network was when these historical data sets were collected.

And additionally to that one-class classifiers can be trained for each such S-cluster. Therefore after submitting an actual data set to such trained one-class classifiers, each trained one-class classifier provides a confidence value as a response to the actual data set provided to the trained one- class classifier. Confidence value means with what probability the actual data set provided to the respective trained one-class classifier corresponds to the respective S- cluster. After that it is possible to make judgment about the actual state of the fluid supply network that corresponds to the actual data set. This particular method gives additional advantages in the cases when the location of the metering point can be hardly changed, as well as overall number of the metering points.

Thus, the present invention is proposed to provide a new method and a system for detection of a fluid supply network state based on cluster analysis.

Further embodiments of the present invention are subject of the further sub-claims and of the following description, referring to the drawings .

In a possible embodiment selection of metering points from the plurality of metering points is based on topological criteria. In other words the plurality of metering points can be clustered based on different topological criteria: geographic distance between centroids (geometrical center) , topological distance, minimal or maximal number of the metering points within the cluster, etc. In other words metering points selected for further performance of the method should be topological proximal ones .

This feature allows providing more precise S-clustering on the further steps, therefore the overall precise detection of an actual fluid supply network state. In other possible embodiment, the actual state of the fluid supply network is judged to be a state of the fluid supply network assigned to the S-cluster for which the respective one-class classifier provided the highest confidence value as a response to the collected actual data provided to the respective trained one-class classifier.

This feature allows properly interpreting the results of the step of making judgment about the actual state of the fluid supply network. In case all confidence values are below a predefined threshold the actual state of the fluid supply network is to be judged as an unassigned state of the fluid supply network that has not been classified yet.

Such case should be additionally investigated by domain experts. Since one of the reasons of such low confidence values can be incorrect selection of metering points for further method performance, and / or incorrect S-clustering at the step of clustering the plurality of historical data sets to form at least one S-cluster. This feature allows properly interpreting the results of the step of making judgment about the actual state of the fluid supply network.

In another embodiment the at least one metering point is installed on a distribution pipe. As it was mentioned above the network of pipes for delivering a fluid to consumers comprises at least one main pipe to transport fluid from a source into the fluid supply network, and at least one distribution pipe, wherein the at least one distribution pipe is adapted to transport fluid from the main pipe to a consumer fluidly connected to the distribution pipe.

The distribution pipes are pipes with small diameter (e.g. less than 300 mm) to carry the pressurized fluid from the main pipe to the consumers. The sensors that are installed on the distribution pipes are relatively cheap in comparison with the ones that installed on the main pipes. In addition to that the installation process of sensors on a distribution pipe is not connected with the interruption of fluid supply to the consumers of the fluid supply network.

Therefore using metering points equipped with sensors installed on the distribution pipes makes this method more cost effective in comparison with the case when metering points are installed on the main pipe.

In other possible embodiment the fluid supply network is a water supply network.

In enhanced embodiment of the system the at least one metering point is installed on a distribution pipe. Wherein the network of pipes for delivering a fluid to consumers comprises at least one main pipe to transport fluid from a source into the fluid supply network, and at least one distribution pipe, wherein the at least one distribution pipe is adapted to transport fluid from the main pipe to a consumer fluidly connected to the distribution pipe.

In other possible embodiment of the system the fluid supply network is a water supply network.

For a more complete understanding of the present invention and advantages thereof, reference is now made to the following description taken in accompanying drawings. The invention is explained in more detail below using exemplary embodiments which are specified in the schematic figures of the drawings, in which:

Fig. 1 shows a block diagram of a fluid supply network; Fig. 2 shows a flow diagram of an embodiment of a method of a state of a fluid supply network based on cluster analysis in accordance with the present invention; Fig. 3 shows a block diagram of the system a system 16 for detection of a state of a fluid supply network based on cluster analysis according to the present invention;

Various embodiments are described with reference to the drawings, wherein like reference numerals are used to refer to like elements throughout. In the following description, for purpose of explanation, numerous specific details are set forth in order to provide a thorough understanding of one or more embodiments. It may be evident that such embodiments may be practice without these specific details.

The invention relates to a system 16 for detection of a state of a fluid supply network based on cluster analysis.

FIG 1 illustrates a water supply network 1 with a water source (not shown in FIG 1) , a supply network pressurizing component 2, main pipes 3 and distribution pipes 4. The fluid pressurized by the supply network pressurizing component 2 is being transported through main pipes 3. Distribution pipes 4 are used to connect the consumers 5 with the main pipes 3.

The main pipes 3 and distribution pipes 4 might be equipped with metering points 6. Such metering points 6 can be installed directly in the main pipes 3, or in the distribution pipes 4. Each metering point 6 is equipped with the at least one sensor 7. Such sensor 7 can measure different parameters of a fluid condition in the pipe of the fluid supply network 1 where such sensors 7 are located. Each sensor 7 can measure one parameter or plurality of them. Also a plurality of sensors 7 measuring at least one parameter can be installed in the same metering point 6.

Such parameters measured by the at least one sensor 7 can be pressure of the fluid in the pipes 3, 4, pressure of the fluid provided by the supply network pressurizing component 2, temperature, consumptions, flows, fluid quality parameters such as fluid chemical composition, turbidity, etc. In addition to that such sensors 7 can fix particular time when the parameters of a fluid condition were measured. The sensor 7 can be a pressure sensor, a flow meter and etc. Different metering points 6 can have different sensors 7 installed on.

However the observability of fluid supply network state should be achieved by measuring such parameters . In other words in case the state of the fluid supply network 1 is changing the value of such parameters are also changing. Therefore it is possible to determine a state of the fluid supply network 1 by measuring such parameters .

For example, measuring pressure in the fluid supply network 1 makes possible to determine emergency of the fluid leakage. Since emergency of the fluid leakage somewhere in the fluid supply network 1 will bring to the pressure drop in the pipe 3, 4 where the leakage is appeared and can be detected by the sensors 7 installed on the pipes 3, 4.

As it is mentioned above said distribution pipes 4 are pipes with small diameter (e.g. less than 300 mm) to carry the pressurized fluid from the main pipe 3 to the consumers 5. The metering points 6 equipped with the at least one sensors 7 that are installed on the distribution pipes 4 are relatively cheap in comparison with the ones that are installed on the main pipes 3. In addition to that the installation process of said metering points 6 on the distribution pipe 4 is connected with the interruption of fluid supply to the limited number of consumers 5 in comparison with the situation when such metering points 6 are being installed on the main pipe 3.

The more metering points 6 with more sensors 7 are installed in the fluid supply network 1 the more precise the state of the fluid supply network 1 can be detected.

The topology of the fluid supply network 1 is well known by the utilities companies who service the fluid supply network 1. Therefore location of the metering points 6 equipped with the at least one sensor 7 with respect to the topology of the fluid supply network 1 are well known. FIG 2 shows a flow diagram of an embodiment of a method for detection of a state of a fluid supply network based on cluster analysis.

At step 8 metering points 6 from the plurality of metering points 6 should be selected for further implementation of the further steps of the method. The metering points 6 can be selected based on different criteria and different approaches. It can be all available metering points 6 or just some metering points 6 out of the plurality of metering points 6 available.

The preferable way of selecting metering points 6 from the plurality of metering points 6 is the selection based on topological criteria. The topological criteria can be, for example, geographic distance between centroids (geometrical center) , topological distance, minimal or maximal number of the metering points 6 within the cluster, etc.

Therefore the state of the fluid supply network 1 on such topologically proximal metering points 6 can be observed by taking measurements of the at least one parameter of a fluid condition in the pipe 3, 4 where the at least one sensor 7 is installed on these metering points 6.

In fact it makes sense to cluster the metering points 6 based on the topological criteria and to form sets of the selected metering points 6 wherein each metering point 6 is included into the only one set of the selected metering points 6. Further steps of the method should be performed for each set of the selected metering points 6.

Further the method comprises at least a training mode I and at least a test mode II. The training mode I can be processed also in off-line regime.

Within the training mode I at step 9 a plurality of historical data sets is collected. Wherein the plurality of historical data sets corresponds to the provided data measured with the sensors 7 of the selected metering points 6, wherein each one of the historical data sets corresponds to the provided data measured with the sensors 7 at a particular point in time, and

different historical data sets of the plurality of historical data sets correspond to different particular points in time.

The measurements of the parameters with the sensors 7 installed on the pipes 3, 4 are done with a given time step within a given period of time. The given period of time can be, for example, 1 day, 1 month, 3 months, 6 months, etc. The given time step can be, for example, 1 minute, 10 minutes, 15 minutes, etc. A value of the given time step can be limited by the equipment, for example, by ability of a sensor 7 to make such measurements. The given period of time and the given time step are defined with expert knowledge and also depend on the ability of equipment to provide such data.

Since the historical data sets correspond to the to the provided data measured with the sensors 7 at a particular point in time, and due to the observability of the fluid supply network 1 with the measured parameters, such each historical data set describe the particular state of the fluid supply network 1 at this particular point in time.

The more historical data sets are available, the more precise an actual state of the fluid supply network 1 can be detected by means of the method.

At step 10 the collected plurality of historical data sets are clustered to form at least one S-cluster wherein each S- cluster corresponds to a specific state of the fluid supply network 1. Each historical data set can be included only to one S-cluster.

Cluster analysis or clustering is a well known task of grouping a set of objects in such a way that objects in the same group (called a cluster) are more similar (in some sense or another) to each other than to those in other groups (clusters) . Cluster analysis itself is not one specific algorithm, but the general task to be solved. It can be achieved by various algorithms that differ significantly in their notion of what constitutes a cluster and how to efficiently find them. The clustering method is chosen with the domain expert knowledge . As a result of this step 10 at least one S-cluster should be formed. Preferably plurality of S-clusters should be formed.

In case the only one S-cluster is formed as a result of the step 10, it likely means that whether there is a lack of collected historical data sets, and / or the step of clustering the collected plurality of historical data sets 10 is done incorrectly.

At step 11 a one-class classifier is trained for each S- cluster. Such the trained one-class classifier is adapted to provide a confidence value as a response to a data set provided to the one-class classifier, wherein the data set comprises the provided data measured with the sensors 7 of the selected metering point 6. Each provided confidence value means with what probability the data set provided to the respective one-class classifier corresponds to the respective S-cluster.

In machine learning, one-class classification, also known as unary classification, tries to identify objects of a specific class amongst all objects, by learning from a training set containing only the objects of that class. The confidence value can be measured in percents (for example, 20%, 40%, 85%, etc.), or can be dimensionless value (for example, 0.2, 0.4. , 0.85).

Further at step 12 a state of the fluid supply network (1) is assigned for each S-cluster. Such assignment should be done with domain expert knowledge. The assigned state of the fluid supply network 1 can comprise, for example, information about the existence of leakages with certain parameters (including location), or absence ones, etc.

The more complete that description of the assigned state is made, the more detailed information about the actual state of the fluid supply network will be received later within the test mode II.

For example, including in the assigned state, for example, information about the existence of leakages with different parameters (for example, of different size, in different location), or absence ones, allows to detect leakages and localize them in the fluid supply network further within the test mode II .

The steps 11 and 12 can be performed in any sequence or simultaneously.

Within the test mode at step 13 an actual data set is collected in a particular point in time for which an actual state of the fluid supply network 1 should be determined. The actual data set corresponds to the provided data measured with the sensors 7 of the selected metering point 6.

At step 14 the collected actual data set is provided to each trained one-class classifier that corresponds to the respective S-cluster. As a result of it a confidence value is received from each trained one-class classifier as a response to the collected actual data set provided to the respective trained one-class classifier. Each received confidence value means with what probability the collected actual data set provided to the respective trained one-class classifier corresponds to the respective S-cluster. Further at step 15 judgment about the actual state of the fluid supply network 1 is made based on the received confidence values and the assigned state of the fluid supply network 1 for the respective S-cluster.

The actual state of the fluid supply network 1 is judged to be a state of the fluid supply network assigned to the S- cluster for which the respective one-class classifier provided the highest confidence value as a response to the collected actual data ^" provided to the respective trained one- class classifier. In ordinary situation one confidence value of the provided plurality of the confidence values should be clearly higher than others. Therefore it means that with this high confidence value the actual state of the fluid supply network can be judged as the state that was assigned to the S-cluster for which the trained one-class classifier provided such high confidence value as a response to the provided actual data set .

In case all received confidence values are below a predefined threshold the actual state of the fluid supply network is judged to be a new unassigned state of the fluid supply network.

The predefined threshold should be defined by domain experts. The predefined threshold can be, for example, 0.2 (or 20%), 0.4 (or 40%), or 0.6 (or 60%).

Such case should be additionally investigated by domain experts. Since one of the reasons of such low confidence values can be incorrect selection of metering points for further method performance, and / or incorrect S-clustering at the step of clustering the plurality of historical data sets to form at least one S-cluster.

Another reason for such low confidence values, especially when all confidence values are of the same range and do not differ from each other significantly, is that this unassigned state is a completely new state of the fluid supply network. In this case the actual data set that corresponds to the new, unassigned state of the fluid supply network should be processed through the training mode of the method. In other words more historical data sets related to this new unassigned state should be collected, a one-class classifier should be trained with such historical data sets, and a particular state of the fluid supply network should be assigned by domain experts.

FIG 3 shows a system 16 for detection of a state of a fluid supply network based on cluster analysis according to the present invention. The system 16 as it was described above on the FIG 1 comprises a plurality of metering point 6, wherein each metering point 6 is located on a pipe 3, 4 and equipped with at least one sensor 7. For each metering point 6, the at least one sensor 7 is adapted

to measure at least one parameter of a fluid condition in the pipe 3, 4 of the fluid supply network 1 where the at least one sensor 7 is installed, and

to provide data, which comprises information about the measured parameter. The system 16 also comprises a control unit 17 configured to detect a state of the fluid supply network 1 in accordance with method of any one the claims 1 - 4.

The network of pipes can comprise

at least one main pipe 3 to transport fluid from a source into the fluid supply network 1, and

at least one distribution pipe 4, wherein the at least one distribution pipe 4 is adapted to transport fluid from the main pipe 3 to a consumer 5 fluidly connected to the distribution pipe 4. Therefore at least at least one metering point 6 of the plurality of the metering points 6 can be installed on a distribution pipe 4.

Also the fluid supply network (1) can be a water supply network . While the invention has been illustrated and described in detail with the help of preferred embodiment, the invention is not limited to the disclosed examples . Other variations can be deducted by those skilled in the art without leaving the scope of protection of the claimed invention. Reference numerals

1 - fluid supply network

2 - supply network pressurizing component

3 - main pipe

4 - distribution pipe

5 - consumer

6 - metering point

7 - sensor

8 - 15 - method steps

16 - system

17 - control unit