WO2019063648A1 - Improved detection and characterization of anomalies in a continuum of water - Google Patents

Abstract

The present invention relates to a device for detecting and characterizing anomalies in a continuum of water. The device is configured to receive measurements of physical quantities from sensors situated in the continuum of water. The measurements are thereafter transformed, and anomalies are detected by an anomaly detector which is trained with transformed values obtained from the above sensors. In parallel with the detection, predefined rules make it possible to characterize a possible anomaly. Thus, an anomaly detection can be optimized for one continuum of water in particular, whilst the characterization with predefined rules allows the device to be operational without each anomaly having had to be detected in this particular continuum of water.

Description

 IMPROVED DETECTION AND CHARACTERIZATION OF ANOMALIES IN A WATER CONTINUUM

FIELD OF THE INVENTION

The present invention relates to the field of water management in water distribution networks or natural areas. More particularly, the present invention relates to the detection of anomalies affecting the quality of water.

STATE OF THE ART PREVIOUS

The water distribution networks can be affected by numerous anomalies affecting the quality of the water. For example, the water can be contaminated by various pollutants introduced voluntarily or involuntarily by return of water or connection error, by particles of matter regularly deposited in the pipes and suddenly resuspended, by the reaction with the constituent materials networks, or by bacteria during bacterial regrowth phenomena. To prevent the deleterious effects that may be caused by contamination, or more generally by reducing the quality of water that may cause discomfort to users (change of color, taste, smell ...) , any event affecting the quality of the water should be detected and characterized as quickly and accurately as possible. Problems that alter water quality may also occur in natural systems, stretches or streams such as lakes, ponds or rivers, or together with stretches and natural watercourses. These systems can be affected for example by accidental pollution, or the abnormal growth of algae, significantly degrading their state and preventing their use (potabilisation, bathing water ...).

The detection of such anomalies may in certain cases be essential, in order to initiate corrective actions to correct the anomaly. The different anomalies occurring in an aquatic system can be detected by monitoring a set of physical parameters in the water. For example, an abnormally low value of the parameters such as chlorine concentration or abnormally high for turbidity can detect an anomaly.

 However, this task can sometimes be difficult, for several reasons. On the one hand, the detection of an anomaly requires the deployment of sensors, to have a fine information on the physicochemical parameters of the water. On the other hand, it can sometimes be difficult, from sensor measurements, to distinguish normal variations in physical parameters of the water, due to the different hydraulic operations necessary for the proper functioning of the network, abnormal values reflecting an anomaly. In particular, the specific configuration of each system can give it specific values. For example, the concentration of chlorine may differ significantly within a drinking water distribution network, depending on whether it is close to a plant or a chlorination point or in remote areas with a long time stay. The quality of treated water can also vary more or less regularly depending on the different types of resources available throughout the year. It is then a question of being able to detect if a new mixture is met.

 In order to provide a finer detection of anomalies affecting a water distribution network, the applicant has filed a patent, published under the number WO 201 6/012972. The patent WO 201 6/01 2972 discloses a method of detecting anomalies in a distribution network particularly adapted to a drinking water distribution network, consisting in acquiring measurements of one or more sensors on the distribution network, forming time series of sensor measurements, extracting operational characteristics from sensor measurements, forming a vector of operational characteristics, and detecting an anomaly if the vector is significantly dissimilar from previously constructed vectors.

This method allows a fine and effective detection of anomalies on a water distribution network, since the detection of anomalies is performed by comparing a vector with transformed vectors from the same sensors on the same distribution network of water. This makes it possible to obtain a detection completely adapted to the characteristics of the water distribution network studied. For example, an aquatic system made up of different resources can usually be alternated consumption of water with a different composition (conductivity, temperature, pH ...). In such a system, an anomaly may be detected if a significantly different combination of the usual combinations of these parameters appears, reflecting a new water mixture. In addition, the sensitivity of the method can be defined for the different types of events and measurements.

 [0007] However, this method may require a long training time before presenting a fully operational anomaly characterization. Indeed, a detection and identification of a particular type of anomaly on the basis of this method would require that several vectors representative of this anomaly have been constructed and labeled, and therefore the anomaly has already occurred, in this system. aquatic, several times. Such a method would therefore be fully effective for detecting anomalies in a given aquatic system only after a sufficiently long period (up to several years) so that all the anomalies sought have already occurred several times in the past. the aquatic system studied.

There is therefore a need for anomaly detection in a water distribution network, making it possible to detect and characterize in a fine manner the events affecting the quality of the water, which can rapidly detect and characterize a set. anomalies in a new aquatic system with little historical data.

SUMMARY OF THE INVENTION

For this purpose, the invention describes a device capable of detecting and characterizing anomalies in a water continuum, comprising: at least one communication link, respectively to at least one sensor of at least one magnitude physical in the water continuum; a processor configured to: receive measurements of the at least one sensor by the at least one communication link; generating a plurality of time windows of the measurements; for each time window of said plurality: obtaining a set of values transformed by at least one transformation of the measurements on the time window; apply an anomaly detector to the set of transformed values to detect a state normal or abnormal of the time window, said anomaly detector being parameterized from sets of transformed values resulting from the application of the at least one transformation to time windows of previous measurements of the at least one sensor; detect an anomaly in the water continuum, according to a number of time windows having an abnormal state; performing on the measurements a set of predefined rules for detecting a variation of at least one of the physical quantities; if an anomaly in the water continuum is detected, assigning to the anomaly a type of anomalies associated with a variation of a subset of the physical quantities, if the variation of at least one of the magnitudes of the sub-set of set is detected.

 This allows effective detection of anomalies in the water continuum. Indeed, the detector is parameterized from learning on measurements from the same sensors in the same water continuum. Thus, the anomaly detector is parameterized from measurements taking into account the characteristics of the water continuum, which allows optimized detection of the occurrence of abnormal situations.

 An anomaly is typed, once detected, from predefined rules. This makes it possible to identify, from general rules, what would be the type of anomaly once the anomaly has been detected. This therefore makes it possible to identify the type of the anomaly, even if this anomaly has not yet been detected in the water continuum with the measurements of the at least one sensor.

The device of the invention thus makes it possible to obtain at the same time a fine detection, based on a parameterization specific to the water continuum, of the occurrence of an abnormal situation, and a general typing of the anomaly. based on a global knowledge of the effects of the different anomalies, even if the anomaly has not been detected in the water continuum.

[0013] Advantageously, one of the transformations of the measurements over a time window is a transformation of the measurements into values representative of the variability of the measurements within the time window.

This makes it possible to detect anomalies resulting in a rapid variation of a physical quantity in the water continuum. For example, a rapid change in turbidity may be representative of particle motions. A transformation of measures into values representative of the variability of the turbidity measurements within the time window can thus make it possible to obtain transformed values allowing detection of this type of anomalies.

 Advantageously, the transformation of the measurements into values representative of the variability of the measurements within the time window is performed by subtracting from each measurement on the time window a median of said measurements over a sliding time window comprising at least said time window.

 This makes it possible to overcome slow variations, for example at the scale of a day, measurements over a time window. This transformation has the advantage, in addition to reliably providing transformed values representative of the variability of the measurements within the time window, to be simple to implement and to require a reduced computing power to execute .

Advantageously, the transformation retains only measurements corresponding to a predefined direction of variation.

 This makes it possible to more effectively detect anomalies resulting in a rapid variation of the measurements in a given direction. For example, particle motions can result in a rapid increase in turbidity. Thus, keeping only the measurements corresponding to an increase in turbidity can make it possible to detect this anomaly more safely.

 [0019] Advantageously, one of the transformations of the measurements over a time window consists of a slope test on the time window.

This makes it possible to detect anomalies affecting slow tendencies of a physical quantity. For example, in a drinking water system, a slow increase in the number of bacteria coupled or not to a slow decrease in the concentration of chlorine may mean the risk of occurrence of a bacterial growth anomaly. It is advantageous to alert operational staff to this type of current phenomenon so that they can be followed and possibly corrected. In a natural aquatic system, an increase in chlorophyll-a over several days, leading first to an increase and then a progressive decrease of the dissolved oxygen can mean the appearance of phenomena of growth of algae. They may affect aquatic life and the use of the system, and should therefore be anticipated. A slope test over a time window advantageously makes it possible to detect such slow variations, and therefore any underlying anomalies.

 [0021] Advantageously, the slope test is a Mann-Kendall trend test.

 The Mann-Kendall test allows effective detection of the significant slope.

 Advantageously, at least two of the sensors are sensors of the same physical magnitude at two points of the water continuum, and one of the transformations of the measurements over a time window comprises a time difference between the measurements of the two sensors.

This makes it possible to detect anomalies resulting in abnormal changes in physical quantities between different points of the continuum.

[0025] Advantageously, the set of transformed values is a vector, and the anomaly detector is a carrier vector machine with a class.

 This allows a particularly effective detection of abnormal states of time windows. In fact, a Class 1 Support Vector Machine makes it possible to classify vectors in a particularly efficient manner from a set of normal or abnormal vectors. In addition, a Class Support Vector Machine can be parameterized by the proportion of vectors expected to be abnormal. Thus, the anomaly detector can be easily parameterized, using this proportion, to detect more or less abnormal time windows and help operational prioritize complementary analysis and interventions according to their available resources.

 Advantageously, the set of transformed values is obtained by at least two transformations of the measurements on the time window, and the anomaly detector is configured to determine a normal or abnormal state of the time window from the set. transformed values.

This allows the anomaly detector to detect to what extent a set of transformed values is different from the previously collected sets, from the set of values, and therefore from detect possible links between transformed values derived from physical quantities and / or different sensors. This solution is particularly effective in a water distribution network. Indeed, this makes it possible to detect abnormal simultaneous variations of several physical quantities and / or at several points of a water distribution network.

 Advantageously, the set of transformed values is obtained by at least two transformations of the measurements over the time window, and the anomaly detector is configured to: detect normal or abnormal states of at least two subsets of values respectively transformed by said at least two transformations of the measurements over the time window; detecting the normal or abnormal state of the time window from a combination of said normal or abnormal states of the subsets.

 This makes it possible to detect in a first stage the normal or abnormal state of the values resulting from each of the transformations, for example of each physical quantity and / or sensor. The normal or abnormal state of the set of transformed values is defined according to the normal or abnormal states of each subset. This makes it possible to take into account more finely the measurements of each physical quantity and / or each sensor. This solution is particularly effective in the natural environment. Indeed, a sensor in a natural environment is more likely to produce false positives: the variability of the values of the physical quantities is greater, and the sensors are likely to undergo slow drifts due for example to fouling, because the waters are more loaded with suspended matter. Sensors in natural environments are therefore more likely to fail and / or produce noisy measurements. Perform an anomaly detection on the measurements resulting from each transformation (and therefore of each sensor), then detect abnormal windows only if at least two subassemblies have abnormal values makes it possible to detect anomalies only the measurements resulting from at least two sensors have anomalies.

Advantageously, the processor is configured to calculate a variation intensity of the physical quantities, and the set of predefined characterization rules comprises a predefined rule for detecting a variation of a physical quantity, if the intensity of variation of said physical quantity on the measurements is greater than a normal variation threshold.

 This effectively detects a variation of a physical quantity.

Advantageously, the processor is configured to assign a criticality indicator to the anomaly, as a function of the intensities of the variations of the physical quantities of the subset.

 This makes it possible to automatically determine the gravity of the anomaly, and thus to engage the most appropriate corrective actions.

Advantageously, the detection of the variation of the physical quantity detects a variation only if the variation of the physical quantity is consistent with a direction of variation.

 This allows, in cases where an anomaly is related to a given direction of variation (increase or decrease), to retain a variation of the physical quantity only if the physical quantity varies in the defined direction.

Advantageously, the processor is configured, if an anomaly is detected, to affect the anomaly at least one of the following types: a type "bacterial growth", in the case of variation of physical quantities of a sub-component. set comprising: a decrease in chlorine concentration, an increase in temperature, an increase in the total organic carbon content, an increase in ultraviolet light absorbance of wavelength 254 nm, an increase in the number of bacteria; a "water mixture" type, in the case of variations in physical quantities of a subset comprising a conductivity, a pH; a temperature; a "colored water" type, in the case of variations the physical quantities of a subset comprising a chlorine concentration, a pH, an increase in color, an increase in turbidity, an increase in the absorbance of ultraviolet light of length 254 nm wave; overspeed, in case of abnormal increase of physical magnitudes among a subset including turbidity and particles.

These examples make it possible to reliably detect the aforementioned types of anomalies.

Advantageously, the processor is configured, if the output of predefined rules for detecting a variation of at least one of the magnitudes physical does not allow the assignment of a type to an anomaly, to assign to this anomaly an unknown type.

 This allows to generate in all cases an anomaly associated with a type.

Advantageously, the device comprises an interface (250) for displaying the anomaly and its type to an operator.

 This allows an operator to view the anomalies, and if necessary provide corrective operations.

 Advantageously, the interface is configured to receive from the operator a label concerning the anomaly; the time window of values, and the label concerning the anomaly are added to the training data.

This makes it possible to correct a possible incorrect detection of anomalies, and to improve the training of machine learning machines for the detection of anomalies.

The invention also describes a method for detecting and characterizing anomalies in a water continuum, comprising: receiving (810) measurements of a plurality of physical quantities from a plurality of sensors of the plurality of physical quantities in the water continuum; generating (820) a plurality of time windows of measurements; for each window of said plurality: obtaining a set of values transformed by at least one transformation (830) of the measurements over the time window; the detection (840) of a normal or abnormal state of the time window, said detection being parameterized from sets of transformed values resulting from the application of the at least one transformation to time windows of previous measurements of the at least one sensor; detecting (850) an anomaly in the water continuum, based on a number of time windows having an abnormal state; performing (860) a set of predefined rules for detecting a variation of at least one of the physical quantities; if an anomaly in the water continuum is detected, the assignment (870) to the anomaly a type of anomalies associated with a variation of a subset of the physical quantities, if the variation of at least one quantities of the subset is detected.

The invention also describes a computer program product comprising program code instructions recorded on a computer program. computer readable medium comprising a processor for detecting anomalies in a water continuum, said computer program comprising computer readable programming means for receiving measurements of a plurality of physical quantities from a a plurality of sensors of the plurality of physical quantities in the water continuum; generating a plurality of time windows of measurements; for each window of said plurality: obtaining a set of values transformed by at least one transformation of the measurements over the time window; applying an anomaly detector to the set of transformed values in order to detect a normal or abnormal state of the time window, said anomaly detector being parameterized from transformed sets of values resulting from the application of the least a transformation to time windows of previous measurements of the at least one sensor; detect an anomaly in the water continuum, according to a number of time windows having an abnormal state; execute a set of predefined rules for detecting a variation of at least one of the physical quantities; if an anomaly in the water continuum is detected, assigning to the anomaly a type of anomalies associated with a variation of a subset of the physical quantities, if the variation of at least one of the magnitudes of the sub-set of set is detected.

LIST OF FIGURES

Other characteristics will become apparent on reading the detailed description given by way of example and not limiting thereafter with reference to the accompanying drawings in which:

 FIG. 1 represents an example of a probe intended to detect anomalies in a water distribution network, according to a set of modes of implementation of the invention;

 FIG. 2 represents an exemplary device for detecting anomalies in a water continuum according to a set of embodiments of the invention;

FIG. 3 represents an example of transformation of measurements of a physical quantity in an implementation mode of the invention; FIG. 4 represents an example of detection of an abnormal time window in a drinking water distribution network;

FIG. 5 represents an example of detection of an abnormal time window in water in a natural environment;

 FIGS. 6a and 6b show two examples of combinations of events on a water distribution network, in a set of modes of implementation of the invention;

 FIG. 7 represents an interface presenting events occurring in a water distribution network to an operator in a set of modes of implementation of the invention;

 FIG. 8 represents an example of a method for detecting and characterizing anomalies in a water continuum according to a set of embodiments of the invention.

DETAILED DESCRIPTION

In the following description the method according to the invention is mainly illustrated by examples relating to the detection of anomalies in a drinking water distribution network. However, the invention is not restricted to these examples, and can be applied to any event detection related to water quality, from measurements of at least two sensors. FIG. 1 represents an example of a probe intended to detect anomalies in a water distribution network, according to a set of embodiments of the invention.

 The probe 100 for detecting anomalies in a water distribution network. The probe can be placed at any location in a water distribution network, for example when leaving a drinking water production plant, at the outlet of the tank, at a point of consumption, where in all other point.

The probe 100 is connected to pipes 130 of the water distribution network, for example by a set of sensor cables 120 or directly connected to the network (insertion), and is supplied with electricity 1 10 or battery.

 The probe 100 may comprise one or more physical size sensors of the water distribution network. For example, the probe 100 may comprise one or more sensors selected from chlorine concentration, temperature, TOC (Total Organic Carbon, Total Organic Carbon), UV 254 (absorbance of water for a light) sensors. ultraviolet wavelength 254 nm), conductivity, pH, color, turbidity, number of particles, number of bacteria, dissolved oxygen, chlorophyll a or any sensor of a physical quantity that can characterize water.

 The probe 100 thus makes it possible, in a set of embodiments of the invention, to measure a set of parameters representative of the quality of the water at a point in the water distribution network. .

 In a set of embodiments of the invention, the probe 100 comprises communications means for transmitting the measurements of the onboard sensors. For example, the probe may include a wired or wireless connection to a server to send the measurements to a server configured to detect anomalies in the water distribution network. The probe 100 can thus be coupled to an intelligent water consumption sensor sending consumption data by remote reading, by sending, in combination, water consumption data and measurements of the sensors.

In a set of embodiments of the invention, the probe comprises a processor configured to detect and characterize, from sensor measurements, anomalies in the water distribution network.

Examples of detection and characterization of anomalies by a processor will be given below, the techniques for detecting and characterizing anomalies described with reference to the following figures being applicable to a processor embedded in the probe 100.

Although the probe 100 represents an example of a probe in a water distribution network, such probes can also be deployed in water in a natural environment, for example in a lake, pond, river or any other aquatic system, possibly at different depths.

FIG. 2 represents an exemplary device for detecting anomalies in a water continuum.

 The network 210 is a water distribution network equipped with at least one sensor 21 1, 212 of at least one physical quantity. For example, the sensors 21 1, 212 may comprise one or more sensors chosen from chlorine concentration, temperature, TOC, UV 254, conductivity, pH, color, turbidity, number of particles, bacteria, dissolved oxygen, chlorophyll a or any sensor of a physical quantity that can characterize water. The sensors 21 1, 212 may be both insulated sensors, as sensors located in multi-sensor probes, such as for example the probe 100 shown in FIG.

Although FIG. 2 represents sensors in a water distribution network, the invention is also applicable to water sensors in a natural environment, for example in a lake, a pond or a river, or more generally in an aquatic system. The sensors may be located at the same location, or spread across multiple locations in the water distribution system or the natural environment. However, the sensors must be located in the same water continuum, ie, if the sensors are not located in the same place, they must be located in the same aquatic system or communicating aquatic systems between they, so that, when several sensors are present, their measurements can be cross-checked to detect and characterize the anomalies in the same continuum of water. This condition is fulfilled if the sensors are located in the same water distribution network, at two points of a lake or river, in a lake and a tributary or effluent river of that lake, or more generally in two points of the same continuum of water. The term "water continuum" can therefore as well designate a water distribution network as an aquatic system in a natural environment.

The device 200 makes it possible to detect and characterize anomalies in a continuum of water, for example the water distribution network 210. It comprises for this purpose a processor 240 and at least one link 221 and 222 with the sensors 21 1, 212. In a set of embodiments of the invention, the communication with the sensors 21 1, 212 can be performed by a measurement acquisition system, for example a SCADA platform (Supervisory Control And Data Acquisition). In one set of embodiments of the invention, the communication links 221, 222 are a radio link with a receiver 220. Many other implementations of the communication links are possible. For example, the sensors 220, 221 may send values to a hub, which is connected by a radio link or a cellular telephone link to a modem in the device 200. Some parts of the communication link may be wired links. The person skilled in the art can define, without effort, the relevant communication links for a given use case, for example by selecting one of the well-known communication links in the field of the management of water distribution networks. . According to various embodiments of the invention, the device 200 may belong to different types of computing devices. For example, the device 200 may be a personal computer, a workstation, a server, a digital tablet, or any other suitable device.

The system 200 further comprises a storage medium 230. In the example shown in FIG. 2, the storage medium is located in the device 200. In other embodiments of the invention, the supports storage can be located outside the computing device. For example, it may be a shared hard disk, or a remote database, accessible through queries by the computing device.

In the example shown in FIG. 2, the storage medium comprises parameters 231 of an anomaly detector in the water continuum, a measurement database 232. Parameters 231 make it possible to parameterize a anomaly detector in the water continuum. As will be described in more detail below, the parameters are obtained from transformations of previous measurements. For example, the parameters 231 may be transformed previous measurement vectors, or a configuration of an automatic learning engine obtained using previously transformed measurement vectors. The measurement database 232 is used to store measurements from the sensors 21 1, 212. The measurements include values of physical quantities measured by the sensor or sensors 21 1, 212, and a time stamp with the date / time of the measurements. The storage medium 230 may for example be a hard disk, a solid state drive, a flash memory or any other type of known storage. In other embodiments, the parameters 231 and / or the measurement database 232 may be stored on different storage media, inside or outside of the computing device 200.

 The device 200 may also include a set of inputs / outputs 250 such as a screen, a keyboard or a router.

The device 200 comprises a processor 240. Different types of processor can be used in the context of the invention: the processor may be for example a microprocessor, a microcontroller or a digital signal processor (also known in English Digital Signal Processor or DSP). The processor is not limited to any type or processor architecture, and may be configured to perform operations by loading executable code elements. The processor can be as well located in a probe comprising the at least one sensor as in a personal computer or remote server.

 The processor is configured to receive measurements of the at least one sensor 21 1, 212. In the example of Figure 2, the processor is configured to read the measurements in the database of measurements 232. However, the processor 240 can also receive the measurements in another way, for example by directly receiving the measurements of the sensors and storing them in a local working memory.

The processor 240 is further configured to generate a plurality of time windows of measurements. According to different embodiments of the invention, the time windows can be both successive time windows, as successive windows overlapping.

Many times of windows are possible. For example, the time windows can have durations of 5min, 15min, 1h, 6h, 12h or 24h.

The measurements can be sampled at regular intervals within a window. For example, all the sensors can be sampled at the same frequency, and thus produce measurements simultaneously. In other cases, the sampling frequencies can be variable. For example, some sensors can be sampled every 5 minutes while others are sampled every 15 minutes. It is then possible to standardize the time steps using interpolation of the data. The measurements are thus synchronized by taking into account regular time steps at the smallest available acquisition time. This makes it possible to have the same time steps for all the sensors, sampled at the highest acquisition frequency.

The processor is configured, for each time window, to obtain a set of values transformed by at least one transformation 242 of the measurements over the time window.

 The at least one transformation 242 makes it possible to obtain transformed values, some of whose properties may be characteristic of anomalies.

 For example, a transformation 242 may consist of a suppression of the long tendencies (typically daily or weekly) to keep only the rapid variations of the measurements of a physical quantity, that is to say that the transformation 242 transforms measurements in values representative of the variability of the measurements within the time window: the higher the transformed values, the faster the time window is affected, and the more the measurements are dispersed. For example, one of the transformations of the measurements on a window can consist of a filtering of the rapid variations, for example of the order of a few hours, compared to the daily trend, equivalent to a high-pass filter, measurements of a Physical size. Such a transformation makes it possible to keep only the rapid variations of the measurements of a physical quantity, and thus to detect anomalies related to an abnormally rapid variation of this quantity. For example, sediment movements can be detected by abnormally rapid changes in turbidity.

The transformation of the measurements into values representative of the variability of the measurements within the time window can be achieved by subtracting from each measurement on the time window a median of said measurements over the time window or over a longer time window. This allows efficient high-pass filtering while retaining the information given by each measure. As an alternative, filtering high pass can also be achieved by other means. For example, it can be done by subtracting from each measurement on the time window an average of said measurements over the time window, or by performing a frequency transform.

The transformation may furthermore keep only measurements corresponding to a predefined direction of variation. For example, in the case of turbidity measurements, only sharp increases can be maintained, which can be caused by sediment movements. In the case of chlorine concentration measurements, only those measures representative of a decrease in the chlorine concentration, which may characterize an increase in the number of bacteria or in the level of organic matter, may be retained. Thus, keeping only the measurements corresponding to an increase in turbidity can make it possible to detect this anomaly more safely. This can be done, for example, in the case mentioned above, where the filtering of the rapid variations is carried out by subtracting from each measurement on the time window a median of said measurements over the time window, while retaining after subtraction:

 - that positive or zero values, if the desired direction of variation is an increase;

 - Negative or zero values, if the desired direction of variation is a decrease.

A transformation may also consist of the application of a slope test over a time window. This makes it possible to detect slow variations over a time window. Some anomalies can cause slow changes in some physical quantities. For example, in a drinking water system, a slow increase in the number of bacteria coupled or not to a slow decrease in the concentration of chlorine may mean the risk of occurrence of a bacterial growth anomaly. It is advantageous to alert operational staff to this type of current phenomenon so that they can be followed and possibly corrected. In a natural aquatic system, an increase in chlorophyll-a over several days leading first to an increase and then a progressive decrease in dissolved oxygen can mean the appearance of growth phenomena. seaweed. They may affect aquatic life and the use of the system, and should be anticipated.

 The size of the time window can be adapted to the estimated duration of the phenomena involved. For example, a slope test can be performed by comparing the medians of the measurements per day, over a time window of 5 successive days. The slope test can be, for example, a Mann-Kendall tendency test, which is particularly reliable. A slope coefficient can be estimated for example using a Theil-Sen estimator.

The transformations mentioned above are applied to time windows of measurements of a physical quantity from a single sensor. However, the invention is not restricted to these examples. For example, a person skilled in the art could define transformations making it possible to establish a link between measurements of the same physical quantity from two different sensors, in order to obtain transformed values representative of the evolution of a physical quantity. in the continuum of water. For example, two dissolved oxygen sensors may be located upstream and downstream of a river, and a joint transformation of the measurements from the two sensors makes it possible to obtain transformed values characterizing the evolution of the dissolved oxygen in the river. river.

For example, at least two of the sensors may be sensors of the same physical magnitude at two distinct points of the water continuum, and one of the transformations of the measurements over a time window may include a time difference between the measurements of the measurements. two sensors.

In many cases, the flow of water in the continuum defines temporal correlations between the physical quantities at different points of the continuum. For example, if the water continuum is a river, and two conductivity sensors are located, one downstream of the other, measurements from the downstream sensor will be correlated with upstream sensor measurements, a time difference corresponding to the average time taken by the water to travel the distance between the upstream sensor and the downstream sensor.

An optimal time offset between the measurements of two sensors of the same physical magnitude at two points of the continuum can be estimated from the historical data, by performing on the previous measurements cross correlations between the measurements of the two sensors, by performing a correlation significance test, and selecting the most important peak.

 When a transformation of the measurements consists in making a time difference according to the optimal time offset between the measurements of two sensors of the same physical quantity at two distinct points of the continuum, the transformed values will be small if the flow is normal. , and if no event impacting the physical quantity tested has occurred between the two sensors.

 On the contrary, high values could be representative:

 or a modification of the flow of water in the continuum, implying that the time taken by the water to travel the space between the two sensors becomes different from the optimal time offset previously calculated, the measurements of the magnitude physical being no longer correlated according to this shift;

 or an event affecting the physical quantity occurring between the two sensors.

 Such a transformation therefore makes it possible not only to detect an anomaly from measurements distributed in the network, but also to provide location information of a possible anomaly.

The one or more transformations thus defined makes it possible to obtain characteristic transformed values, also called "features", making it possible to detect anomalies. According to various embodiments of the invention, different transformations can be applied to the same time window to measurements of different physical magnitudes. It is also possible to apply several transformations to the same physical quantity. Those skilled in the art can thus, knowing the characteristics associated with potential anomalies in the water continuum, define the transformations most suited to the types of measurements available, and the anomalies sought.

The set of transformed values can take many forms, from the moment when the structure of the values transformed on the different time windows remains identical, in order to allow identification of the abnormal sets. It is necessary that the transformations be applied in a similar way to the data of the current vector and past vectors, serving as drive data for anomaly detection.

 For example, the set of transformed values can be obtained in the form of a vector. If only one transformation is applied, the vector can simply include the transformed values on the different parameters on the time window. It can also contain the successive measurements transformed on the time window. If multiple transformations are applied, the transformed values at the end of each transformation can be concatenated to form a single vector.

 In a set of embodiments of the invention, the processor is configured to normalize the set of transformed values, in order to detect anomalies on homogeneous values. For example, the processor may be configured to calculate the standard deviation of the transformed values over a time window, and divide all values of the window by this standard deviation. The processor can also be configured to remove singular extreme values, in order to calculate the features as well as the statistical indicators (standard deviation, mean, median) used for normalization. Several embodiments are possible for this purpose. For example, the processor can be configured to calculate the mean and standard deviation of the values on the window, and to suppress the higher values by a predefined number of times the standard deviation from the mean, and / or the values lower than this predefined number of times to the mean, for example above average values plus three times the standard deviation, or below average minus three times the standard deviation. Another option is to delete a given percentage (for example 5%) of the highest and / or lowest transformed values on the time window.

When the set of transformed values is obtained by concatenating transformed values resulting from several transformations, the normalization of the values can be carried out separately on the values resulting from each transformation. This makes it possible to obtain a set of homogeneous transformed values, even if the applied transformations and the amplitudes of the initial measurements are very different. It does not exist no intrinsically dominant parameter in the learning space.

 In a set of embodiments of the invention, the set of transformed values may also include short-term data that may have an impact on the measured physical quantities. For example, these data may include meteorological data such as measurements of humidity, air temperature, sunshine, cumulative rainfall or a number of successive days without rain at the time the measurements are made. These parameters affect in particular the water quality of natural aquatic systems. The rain accumulation data is for example used, via a threshold, to remove the time steps corresponding to a rain time from the detection of anomalies. FIG. 3 represents an example of transformation of measurements of a physical quantity in one embodiment of the invention.

The curve 310 represents a time window of chlorine concentration measurements. The horizontal axis represents the elapsed time, in number of time steps of 5 minutes. The vertical axis represents the measurement intensities, in ppm (parts per million).

 A first transformation step consists in calculating a median of the measurements over the time window, and subtracting this median from the measurements, in order to obtain a first level of transformations 320. The transformed values are then representative of the rapid variations of the measurements. .

 A second transformation step consists in keeping only the negative values (thus representative of a decrease in chlorine concentration). The transformed values 330 therefore make it possible to evaluate a level of rapid decrease in chlorine concentration. Since an excessive decrease in chlorine concentration can be caused by a bacterial growth phenomenon, these transformed values can be used to detect this anomaly.

This transformation is given by way of example only of a transformation according to the invention. Those skilled in the art can easily define the most appropriate transformations according to the types of anomalies sought, for example by keeping, on a time window only the rapid or slow variations, and the increases or decreases of a given physical quantity. Returning to FIG. 2, the processor is further configured to apply an anomaly detector 243 to the set of transformed values in order to detect a normal or abnormal state of the time window, said anomaly detector being parameterized. from sets of transformed values resulting from the application of the at least one transformation to time windows of previous measurements of the at least one sensor.

 The anomaly detector can take different forms. For example, the fault detector may be a set of processor instructions, a software module, or an automatic learning module configured using previous anomaly detections. The anomaly detector is configured to detect, for a time window, whether the set of values is representative of an abnormal state of the time window. For this purpose, the anomaly detector is parameterized with sets of previous transformed values, originating from the at least one transformation 242 to previous measurement time windows of the at least one sensor 21 1, 212. Thus, the An anomaly detector is driven with transformed values representative of the water continuum, which allows an anomaly detection well adapted to the studied water continuum. Although the anomaly detector improves as the amount of training data increases, a detector indicating a normal or abnormal state, on data from the same sensors, can be set. place quickly. In particular, such a detector can detect anomalies as soon as transformed values are significantly different from the transformed values of the learning base. On the other hand, it does not require to have met all the anomalies previously, nor that anomalies have been labeled. Thus, the anomaly detector can be functional very quickly, as soon as a few transformed measurements are available.

The detection of an abnormal state of the window can for example be carried out by the anomaly detector by affecting a score, for example between 0 and 1, to the set of transformed values of a window indicating how much this set is different from the sets of previous transformed values, and determining the window state as abnormal if the score is greater than a threshold predefined. It is also possible to detect an abnormal state of a time window if the transformed values are part of a given percentage of the most different transformed values of the previously collected values.

 [0098] Many embodiments of the anomaly detector are possible. In particular, the anomaly detector may be a machine learning algorithm, automatically learning, from the transformed values previously obtained, the transformed values representative of a normal or abnormal state of a time window.

In a set of embodiments of the invention, the set of transformed values is a vector, and the anomaly detector is a Class 1 (or one-class SVM) Vector Support Vector Machine. English one-class Support Vector Machine, literally a Vector Support Class Machine). This type of anomaly detector allows a particularly effective detection of the abnormal states of the time windows. Indeed, a Class Supported Vectors Machine is a particularly efficient way of representing the boundary between normal and abnormal vectors (too far from the normals), in order to determine a level of differences between a vector and the previously studied vectors. In addition, a Single Class Support Vector Machine Machine is set using the proportion of vectors expected to be abnormal. An anomaly detector of the Support Class Vector Class has the advantage of providing an efficient classification of the data in normal or abnormal states. In addition, it can be set up very quickly, with a limited amount of training data.

The carrier vector machine with a class can thus be parameterized with a variable abnormal vector ratio, for example between 1/100 and 1/10000 (that is to say with a set ratio of abnormal transformed values between 1/100 and 1/10000, the Class Support Vector Machine automatically adapts its definition of the abnormal set to achieve this value ratio abnormal). The ratio can be adjusted at any time, to detect a larger or smaller number of data sets as abnormal.

In a set of embodiments of the invention, the set of transformed values is obtained by at least two transformations of the measurements over the time window. This is for example the case when the set of transformed values is obtained by transforming measurements of several physical quantities and / or measurements from several sensors. For example, such a configuration can be obtained if the set of transformed values comprises values resulting from a transformation of conductivity measurements, a transformation of pH measurements and a transformation of temperature measurements.

In these cases, the question may be posed in the most appropriate manner of detecting anomalies on this set of transformed values.

In a set of embodiments of the invention, the anomaly detector is configured to determine a normal or abnormal state of the time window from the set of transformed values.

 This allows the anomaly detector to detect, from the set of transformed values from different physical quantities and / or sensors, to what extent this set is different from the previously collected sets, and therefore of the detector. possible links between transformed values derived from physical quantities and / or different sensors. This solution is particularly effective in a water distribution network. This solution makes it possible to detect at the same time important variations on a single parameter or smaller variations but applying to several parameters simultaneously. The detection can thus highlight in a particularly fine manner phenomena affecting several parameters. In the case of drinking water networks, the variability of the signal is generally better controlled because of regular business operations. In addition, the sensors are in an environment, drinking water, cleaner. They generally experience less drift.

In one set of embodiments of the invention, the set of transformed values is obtained by at least two transformations of the measurements on the time window, and the anomaly detector is configured to:

 detecting normal or abnormal states of at least two subsets of values respectively transformed by said at least two transformations of the measurements over the time window;

detecting the normal or abnormal state of the time window from a combination of said normal or abnormal states of the subsets.

 The different subsets typically correspond to subsets of measurements from each sensor and / or each physical quantity on the time window. For example, if a time window includes conductivity measurements from a single sensor, and turbidity from two different sensors, the set of measurements can be separated into three subsets: a subset of conductivity measurements , and two subsets of turbidity measurements, corresponding to the measurements from each of the two sensors.

It is also possible to define subsets according to different rules. For example, a subset could correspond to measurements of the same multi-sensor probe, or measurements of the same physical quantity from several sensors at several different points.

This makes it possible to detect in a first stage the normal or abnormal state of the values resulting from each of the transformations, for example of each physical quantity and / or sensor. The normal or abnormal state of the set of transformed values is defined according to the normal or abnormal states of each subset. This makes it possible to take into account in a differentiated manner the measurements of each physical quantity and / or each sensor. This solution is particularly effective in the natural environment. Indeed measurements in the natural environment are subject to a greater variability of each parameter, as well as to failures or drifts of the sensors more frequent because of the hostility of the environment. It is therefore relevant to try to ensure that the parameters present abnormal values simultaneously or slightly offset in time, to reduce false alarms. In embodiments of the invention in which the detection of anomalies is carried out using machine learning algorithms, for example a Class Support Vectors Machine, a separate learning may be realized for each of the subsets: the different transformations are applied to each of the previous measurement time windows used for the learning, in order to obtain, for each of these windows, the subsets of corresponding transformed values. The training is then carried out separately, for each subset, on the training data thus obtained. This allows a more accurate detection of abnormal situations on a sensor and / or a physical quantity, eliminating some of the false alarms related to sensor faults or the higher natural variability of the measured parameters.

 The combination of the normal or abnormal states of the subassemblies can be done in different ways. For example, an abnormal state of the time window may be detected, if at least one of the subsets has an abnormal state. This makes it possible to detect an anomaly if the transformed measurements of at least one physical sensor or quantity are abnormal. It is also possible to detect an abnormal state of the time window, only if all the subsets have an abnormal state, or if at least a given number of subsets (for example, at least two subsets out of three) present an abnormal state. In the natural environment, there is no anomaly generation when there is detection on a single sensor only, even after temporal smoothing (taking into account the previous time steps to merge the potential anomalies that occurred shortly before).

[00111] Figure 4 shows an example of detection of an abnormal time window in a drinking water distribution network.

The curves 410, 420, 430, 440, 450, 460, 470 and 480 respectively represent the evolutions of measurements of 8 chlorine concentration sensors 41 1, level of a water tank 421, flow 431, conductivity 441, temperature 451, TOC 461, turbidity 471 and UV254 481.

These measurements can be obtained from different sensors, for example the sensors 21 1, 212. [00114] The measurements are taken over a week, between January 26 and February ^1. The measurements are transformed by 24-hour sliding windows, in order to detect abnormal windows. In this example, the measurements follow a regular distribution to the ^1st of February. This regular distribution represents, day after day, the mixing of two water sources for the filling of a tank. Then a problem in the filling of the tank leads to a quality anomaly visible on many parameters over the same time window: the chlorine concentration suddenly decreases 412, the water level in the tank rises then decreases 422, the flow rate water remains zero for 24 hours 432, conductivity of water collapses 442, temperature increases significantly 452, TOC (Total Organic Carbon concentration) decreases 462, turbidity increases 472, absorbance of UV254 decreases 482.

The various transformations of these measurements over the time window make it possible to highlight that the measurements on this window do not follow their usual pattern, and thus to detect an abnormal time window, composed of several abnormal time steps.

These curves are given by way of example only, and the person skilled in the art may, depending on the types of anomalies sought and / or the available sensors, detect abnormal time windows on another set of physical quantities.

FIG. 5 represents an example of detection of an abnormal time window in water in a natural environment.

 FIG. 5 represents measurements taken from a set of sensors of the same probe, comprising:

 conductivity measurements 510;

 measurements of an oxygen dissolution rate 520;

 pH measurements 530.

[00119] Cyclical data (in this case a number of consecutive days without rain 540, recovered from meteorological data) are also used. The measurements are transformed in order to obtain transformed values 51 1, 521, 531 taking the median over periods of 6 hours of the conductivity, the dissolved oxygen level and the pH.

The detection of anomalies is made by time window of 6 hours: a time window is analyzed every 6 hours. An anomaly detector is applied to each time window. In this example, a variation is firstly detected on each physical quantity, and an abnormal state of the window is detected, if at least two physical quantities have a variation within 4 time slots of difference. This makes it possible to avoid detecting an abnormal window, if a sensor is defective.

 [00122] The measurements are here made from November 27 to January 5. The measurements are relatively regular in November and December, in particular by removing the time steps corresponding to rain (inverted blue curves); at the beginning of January, there is a variation of the parameters that can reflect a pollution by product rejection in the medium, and resulting in an increase 512 in the conductivity, and a decrease 522 in the dissolved oxygen level. These anomalies result in abnormally high variations, which make it possible to detect an anomaly of each of these physical quantities over a time window.

The curve 550 represents an evaluation of the intensity of the anomalies on each time window of 6 hours. In this example, the intensity of the anomalies is calculated, only in the case where an anomaly is previously detected, subtracting the cumulative median of the measurements on the available data history of each physical quantity from the previously calculated transformed values, multiplied by 100 and divided by this same cumulative median to normalize these differences. The maximum on the different physical quantities is then calculated in order to produce the indicator represented on the graph. A value of 0 corresponds to the absence of anomalies. However, this is only given by way of example only, and other ways of evaluating a criticality of an anomaly resulting in significant differences between the values within the same time window, or relative to to a data history, could be used. It can be observed that the anomaly detector does not detects no abnormal windows in November and December, then detects a significant proportion of abnormal windows in early January.

These curves are given by way of example only, and the person skilled in the art may, depending on the types of anomalies sought and / or the available sensors, detect abnormal time windows on another set of physical quantities.

[00125] Returning to Figure 2, the processor 240 is further configured to detect 244 an anomaly in the water continuum, according to a number of time windows having an abnormal state.

 In one set of embodiments of the invention, an anomaly is detected in the continuum of water, as soon as a time window has an abnormal state.

 In a set of embodiments of the invention, in order to limit the number of false alarms if a single time window is identified as abnormal, an anomaly in the water continuum is detected only if several time windows are detected. present an abnormal state. For example, an anomaly can be detected if a predefined number of successive windows (for example 2, 3, 4, 10 ...) has an abnormal state. Another option is to detect an anomaly, if, on a number of successive time windows, a percentage of successive windows greater than a given threshold has an abnormal state. For example, an anomaly can be detected in the water continuum if, among 10 successive time windows, at least 75% have an abnormal state. This makes it possible to generate an alert only if an abnormal state is proved. In order to avoid an alert being raised too late, the time windows used may be slippery windows. For example, windows of 6 hours of measurements can be generated every hour (thus, a time window includes 5 hours of measurements in common with the previous window). This therefore makes it possible at the same time to detect an anomaly, only if it is proven because it is highlighted over several time windows, to use long time windows, and to avoid a too important delay in the detection of anomalies.

The processor 240 is also configured to execute on the measurements 245 a set of predefined rules for detecting a variation. of at least one of the physical quantities, and if an anomaly is detected, assigning to the anomaly 246 a type of anomalies associated with a variation of a subset of the physical quantities, if the variation of at least one quantities of the subset is detected.

These rules make it possible to characterize possible anomalies in the water continuum. It is possible to use predetermined rules corresponding to known types of anomalies. For example, it is known to those skilled in the art that a change in water source results in a change in the conductivity, pH and temperature of the water. It is therefore possible, in parallel with the anomaly detection, to execute predefined anomaly characterization rules: thus, if an anomaly has been detected, and that in parallel a modification of the conductivity, pH and the temperature of the water has been observed, it can be deduced that this anomaly is a change of water source.

[00130] The combination of the detection is the characterization of anomalies is particularly effective. Indeed :

 the detection is carried out by comparing transformed values of measurements with transformed values resulting from previous measurements from the same sensors on the same water continuum, which makes it possible to obtain abnormal situation detection particularly suited to the continuum of water, on the basis of statistical occurrences. In addition, as indicated above, such detection can be implemented quickly with limited training data;

 - the characterization can be based on general rules, since the same type of anomaly is reflected, on all water continuums, by a variation of the same physical quantities. Thus, the characterization or the typing of anomalies can be carried out starting from predefined rules applicable to all the aquatic systems, without needing to have been parameterized for a continuum of water, or sensors in particular.

Thus, the device according to the invention makes it possible to detect anomalies in a water continuum reliably and adapted to the characteristics of the environment studied, and to determine their type, without having to have already observed each of the possible anomalies in the continuum of water. The device according to the invention can therefore be operational very quickly, with few prior observations from the sensors 21 1, 212. When a larger set of training data is available, and a sufficient number of anomalies have been labeled, the anomaly detector can be replaced, if the detection is improved, by an anomaly detector directly detecting the type of anomaly.

 According to various embodiments of the invention, the rules for detecting variations in physical quantities can be performed in parallel with the detection of anomalies, or only if an anomaly has been detected.

The detection of variations of physical quantities can for example be performed by calculating an intensity of variation of the physical quantities, and then comparing this intensity to a threshold, the variation being detected if the intensity of variation is greater than the threshold. This intensity of variation can be calculated in different ways. For example, the intensities of variation can be calculated using the transformed values previously calculated for a physical quantity, for example, a slow variation intensity may have already been calculated from a slope test and from the calculation of the slope. , and a fast variation intensity can be calculated by applying a moving 95% quantile to values representative of rapid variation, such as deviations of each measurement to a median over a time window. Thus, the average of the absolute values of the representative values of a fast variation can be calculated. If this average is greater than a threshold, corresponding here to the limit between the lowest 95%, and the highest 5% of the average of the absolute values of the values representative of a rapid variation over a set of previous time windows from of historical data, a variation of the physical quantity is detected. The average can be replaced here by other characteristic values, such as, for example, the sum of the absolute values of the representative values of a rapid variation. The threshold can be calculated from historical data by adding to the median 3 standard deviations. This then represents a high value of the data distribution. However, these rules are given by way of example only, and those skilled in the art can implement any type adapted rule for detecting a particularly large variation of a physical quantity, using or not the transformed values calculated by the at least one transformation 242.

 According to different embodiments of the invention, the intensity of variations can be calculated for all or part of the physical quantities.

For certain physical quantities, a variation can be considered as significant or problematic, only for a given direction of variation. For example, a decrease in chlorine concentration could be considered problematic, while an increase would not be. In a set of embodiments of the invention, a variation of certain physical quantities is retained only if the direction of variation of the physical quantity on the measurements is consistent with a direction of variation.

 In a set of embodiments of the invention, the intensities of variations of the physical quantities also make it possible to determine a criticality of the anomaly. A criticality score can thus be assigned to the anomaly as a function of the intensities of the variations of the physical quantities of the subset associated with the anomaly. This score is calculated according to the intensities of the various previously calculated parameters. These intensities are compared with the intensities calculated on the previous time windows in the data history, only for the critical parameters. They are then normalized with respect to the robust maximum observed on each parameter, then the maximum is taken, in order to determine the criticality as a number between 0 and 1.

The person skilled in the art can provide numerous assignment rules of type to an anomaly, depending on the available sensors, and physical quantities associated with the types of anomalies sought.

For example, the following table gives some examples of types of anomalies that can be affected, depending on the abnormal variations of physical quantities:

Quantities with variation Types of anomalies

 Chlorine concentration, Bacterial growth

temperature, TOC, UV254, number

of bacteria Conductivity, pH, temperature Mixture of water

 Concentration in chlorine, pH, colored water

color, turbidity, UV254

 Turbidity, particles Movements of sediments

 All. Global anomaly

 These possibilities are given by way of example only, and the person skilled in the art can choose, for each anomaly, the physical magnitudes whose variation makes it possible to better characterize it. The invention thus makes it possible to characterize a large number of different anomalies.

According to various embodiments of the invention, a type of anomaly can be detected, if all or part of the associated physical quantities vary. In the example of the table above, an anomaly could be characterized with the type "water mixture", that is if the three quantities conductivity, pH, temperature have an abnormal variation, or if two of the three at least vary.

 Once an anomaly has been detected, it may, in a set of embodiments of the invention, be displayed to an operator via the interface 250. The interface 250 may notably display the type of anomaly, the place where this anomaly has been detected and, where appropriate, the criticality of the anomaly.

 The anomalies detected can be crossed with other sources of information, such as customer complaints, to confirm or deny the anomaly.

 [00143] The operator can take all the measures required by the anomaly. In a set of embodiments of the invention, the operator can also validate or not the anomaly, that is to say, indicate after verification, whether the detection of the anomaly was justified or not. This makes it possible, when the anomaly detector is an automatic learning machine, to have more reliable training data, and thus to improve the performance of the anomaly detection.

In a set of embodiments of the invention, a detected anomaly can generate an event. Different events can be generated, depending on measurements from different sections of a water distribution network, or different sensors. Events can also be combined with each other, a combination local events in a water distribution system that can be interpreted as a global event. If similar events are detected on multiple sensors, the area covered can be displayed to the user for appropriate action. If phenomena are detected both at the factory and on the rest of the network, it is possible to deduce that the anomaly concerns the water treatment plant, and an appropriate event can be generated. . It is therefore necessary to investigate at this level and not that of the network. [00145] FIGS. 6a and 6b show two examples of combinations of events on a water distribution network, in a set of modes of implementation of the invention.

 In these two examples, a drinking water distribution network 620 is located downstream of a water treatment plant 610. Several event generators are based on anomaly detections as represented in FIGS. examples relating to FIG. 2. A first event generator 630 detects possible anomalies from measurements made at the outlet of the drinking water production plant.

In the example of FIG. 6a, three event generators 631a, 632, 633 are executed in parallel, from measurements of physical quantities at three different points of the network. The events thus combined can be combined into events of the network as a whole. For example, a mixture of water should be visible on several points of the same zone.

In the example of FIG. 6b, the event detector 631 has been modified into an event detector 631b, taking into account the occurrence of events 630 at the output of the water production plant. drinking. Thus, when an anomaly is detected by the measurements used by the event detector 631b, the characterization of the type of anomaly or event depends on event detection 630. For example, a change in the production should can not be generated on the network sensors because already detected at the factory.

These examples demonstrate the ability of the invention to detect anomalies or events in a water distribution network, while taking into account the interactions between the different points of the network, in order to detect more global anomalies.

 However, the invention is not limited to these examples, and a combination of events as described with reference to Figures 6a and 6b could for example be applied to the detection of anomalies in the natural environment.

[00151] Figure 7 shows an interface presenting events occurring in a water distribution network to an operator.

The interface 700 is configured to present the anomalies detected by a device according to the invention in a water distribution network to an operator. In this example, the anomalies are detected on a sector 710, and are located on a map. The types of anomalies are represented by different pictograms. In this example, the pictograms respectively represent the positions of the measurement points: chlorine concentration 720, 721, 722 and 723, pH 730 and 731, temperature 740, absorbance of water at UV 750, TOC 760, multi-parameter probes 760 and point sampling points 770.

This example thus demonstrates the ability of the invention to show clearly intuitive to an operator the various anomalies occurring in a water distribution network. However, the invention is not restricted to this example, and other types of representation could be used. In the same way, this representation could be used for water in a natural environment.

FIG. 8 represents an exemplary method 800 for detecting and characterizing anomalies in a water continuum according to a set of embodiments of the invention.

 Method 800 comprises receiving 810 measurements of a plurality of physical quantities from a plurality of sensors of the plurality of physical quantities in the water continuum.

The method 800 further comprises the generation 820 of a plurality of time windows of measurements.

The method 800 further comprises, for each window of said plurality: obtaining a set of values transformed by at least one transformation 830 of the measurements over the time window;

 the detection 840 of a normal or abnormal state of the time window, said detection being parameterized from sets of transformed values resulting from the application of the at least one transformation to time windows of previous measurements of the at least one sensor.

In one set of embodiments of the invention, at least one of the transformations 830 is a transformation of the measurements into values representative of the variability of the measurements within the time window.

 In a set of embodiments of the invention, the transformation of the measurements into values representative of the variability of the measurements within the time window is carried out by subtracting from each measurement on the time window a median of said measurements over a period of time. sliding time window.

 In a set of embodiments of the invention, the transformation retains only measurements corresponding to a predefined direction of variation.

In one set of embodiments of the invention, at least one of the transformations 830 consists of a slope test on the time window.

 In one set of embodiments of the invention, the slope test is a Mann-Kendall trend test.

In one set of embodiments of the invention, at least two of the sensors are sensors of the same physical magnitude at two points of the water continuum, and one of the transformations of the measurements over a time window comprises a time difference between the measurements of the two sensors.

In a set of embodiments of the invention, the set of transformed values is a vector, and the anomaly detector is a Class 1 Support Vector Machine.

In one set of embodiments of the invention, the set of transformed values is obtained by at least two transformations of the measurements over the time window, and the detection of anomalies determines a normal or abnormal state of the time window from the set of transformed values.

 In a set of embodiments of the invention, the set of transformed values is obtained by at least two transformations of the measurements over the time window, and the detection of anomalies:

 detects normal or abnormal states of at least two subsets of values respectively transformed by said at least two transformations of the measurements over the time window;

 detects the normal or abnormal state of the time window from a combination of said normal or abnormal states of the subsets.

 The method 800 further comprises the detection 850 of an anomaly in the water continuum, as a function of a number of time windows having an abnormal state.

 The method 800 further comprises the execution 860 of a set of predefined rules for detecting a variation of at least one of the physical quantities.

In one set of embodiments of the invention, the method 800 comprises a calculation of a variation intensity of the physical quantities, and the set of predefined characterization rules comprises a predefined rule for detecting a variation. of a physical quantity, if the intensity of variation of said physical quantity on the measurements is greater than a threshold of normal variation.

 In a set of embodiments of the invention, the detection of the variation of the physical quantity detects a variation only if the variation of the physical quantity is consistent with a direction of variation.

The method 800 finally comprises, if an anomaly is detected, the assignment 870 to the anomaly of a type of anomalies associated with a variation of a subset of the physical quantities, if the variation of the at least one of the magnitudes of the subset is detected.

In one set of embodiments of the invention, the method 800 includes the assignment of a criticality indicator to the anomaly, according to the intensities of the variations of the physical quantities of the subset.

 In a set of embodiments of the invention, the method 800 comprises, if an anomaly is detected, the assignment to the anomaly of at least one of the following types:

 - a "bacterial growth" type, in the case of variation of physical quantities of a subset comprising: a decrease in the concentration of chlorine, an increase in temperature, an increase in the total organic carbon content, an increase in light absorbance

Ultraviolet wavelength 254 nm, an increase in the number of bacteria;

 a "water mixture" type, in the case of variations in physical quantities of a subset comprising a conductivity, a pH; a temperature;

 a "colored water" type, in the case of variations in the physical quantities of a subset comprising a chlorine concentration, a pH, an increase in color, an increase in turbidity, an increase in the light absorbance of wavelength 254 nm;

- an overspeed, in case of abnormal increase of physical quantities among a subset including turbidity and particles.

 In a set of embodiments of the invention, the method 800 comprises, if the output of predefined rules for detecting a variation of at least one of the physical quantities does not allow the assignment of a type to an anomaly, the assignment to this anomaly of an unknown type.

 In one set of embodiments of the invention, the method 800 comprises a step of using an interface to display the anomaly and its type to an operator.

 In one set of embodiments of the invention, method 800 comprises:

the use of the interface to receive from the operator a label concerning the anomaly; - the addition of the time window of values, and the label concerning the anomaly to the training data.

The method 800 thus makes it possible to detect different types of anomalies in a continuum of water, and to assign them types. All the embodiments described with reference to FIGS. 2 to 7 are applicable to method 800.

The above examples demonstrate the ability of the invention to detect events or anomalies related to water quality. They are however given only by way of example and in no way limit the scope of the invention, defined in the claims below.

Claims

Device (100, 200) for detecting and characterizing anomalies in a water continuum, comprising:  at least one communication link (221, 222) respectively to at least one sensor (21 1, 212) of at least one physical quantity in the water continuum;  a processor (240) configured to:  - receiving measurements of the at least one sensor by the at least one communication link;  generating a plurality of time windows of the measurements;  for each time window of said plurality:  obtaining a set of values transformed by at least one transformation (242) of the measurements over the time window;  - applying an anomaly detector (243) to the set of transformed values in order to detect a normal or abnormal state of the time window, said anomaly detector being parameterized from sets of transformed values from the application at least one transformation to time windows of previous measurements of the at least one sensor;  detecting (244) an anomaly in the water continuum, as a function of a number of time windows having an abnormal state; executing (245) on the measurements a set of predefined rules for detecting a variation of at least one of the physical quantities;  if an anomaly in the water continuum is detected, assigning (246) to the anomaly a type of anomalies associated with a variation of a subset of the physical quantities, if the variation of at least one of the subset quantities is detected. 2. Device according to claim 1, wherein one of the transformations of the measurements over a time window is a transformation of the measurements into values representative of the variability of the measurements within the time window. Device according to claim 2, wherein the transformation of the measurements into values representative of the variability of the measurements within the time window is performed by subtracting from each measurement on the time window a median of said measurements over a sliding time window comprising at least said time window. Device according to one of claims 1 to 3, wherein the transformation retains only measures corresponding to a predefined direction of variation. Apparatus according to claim 1, wherein one of the transformations of the measurements over a time window consists of a slope test on the time window. Device according to Claim 1, in which at least two of the sensors are sensors of the same physical quantity at two points of the water continuum, and one of the transformations of the measurements over a time window comprises a time difference between the measurements of the two. sensors. Apparatus according to one of claims 1 to 6, wherein the set of transformed values is a vector, and the anomaly detector is a one-class Support Vector Machine. Device according to one of claims 1 to 7, wherein the set of transformed values is obtained by at least two transformations of the measurements on the time window, and the anomaly detector is configured to determine a normal or abnormal state of the time window from the set of transformed values. Device according to one of claims 1 to 7, wherein the set of transformed values is obtained by at least two transformations of the measurements on the time window, and the anomaly detector is configured to: detecting normal or abnormal states of at least two subsets of values respectively transformed by said at least two transformations of the measurements over the time window;  detecting the normal or abnormal state of the time window from a combination of said normal or abnormal states of the subsets. Apparatus according to one of claims 1 to 9, wherein the processor is configured to calculate an intensity of variation of the physical quantities, and the set of predefined characterization rules comprises a predefined rule for detecting a variation of d. a physical quantity, if the intensity of variation of said physical quantity on the measurements is greater than a threshold of normal variation. 1 1. The apparatus of claim 10, wherein the processor is configured to assign a criticality indicator to the abnormality based on the intensities of variations in the physical quantities of the subset. 12. Device according to one of claims 1 to 1 1, wherein the detection of the variation of the physical quantity detects a variation only if the variation of the physical quantity is consistent with a direction of variation. Apparatus according to claim 1 to 12, wherein the processor is configured, if an abnormality is detected, to assign to the abnormality at least one of the following types: - a "bacterial growth" type, in the case of variation of physical quantities of a subset comprising: a decrease in the concentration of chlorine, an increase in temperature, an increase in the total organic carbon content, an increase in Ultraviolet light absorbance of wavelength 254 nm, an increase in the number of bacteria; a "water mixture" type, in the case of variations in physical quantities of a subset comprising a conductivity, a pH; a temperature;  a "colored water" type, in the case of variations in the physical quantities of a subset comprising a chlorine concentration, a pH, an increase in color, an increase in turbidity, an increase in the absorbance of ultraviolet light, wavelength 254 nm; - an overspeed, in case of abnormal increase of physical quantities among a subset including turbidity and particles. 14. Device according to one of claims 1 to 1 3, wherein the processor is configured, if the output of predefined rules for detecting a variation of at least one of the physical quantities does not allow the assignment of a type with an anomaly, to assign to this anomaly an unknown type. 15. Device according to one of claims 1 to 14, comprising an interface (250) for displaying the anomaly and its type to an operator. The device of claim 15, wherein:  the interface is configured to receive from the operator a label concerning the anomaly;  - the time window of values, and the label concerning the anomaly are added to the training data. 17. Method (800) for detecting and characterizing anomalies in a water continuum, comprising:  receiving (810) measurements of a plurality of physical quantities from a plurality of sensors of the plurality of physical quantities in the water continuum;  generating (820) a plurality of time windows of measurements; for each window of said plurality: obtaining a set of values transformed by at least one transformation (830) of the measurements over the time window;  the detection (840) of a normal or abnormal state of the time window, said detection being parameterized from sets of transformed values resulting from the application of the at least one transformation to time windows of previous measurements of the at least one sensor;  the detection (850) of an anomaly in the water continuum, as a function of a number of time windows presenting an abnormal state; the execution (860) of a set of predefined rules for detecting a variation of at least one of the physical quantities;  if an anomaly in the water continuum is detected, the assignment (870) to the anomaly a type of anomalies associated with a variation of a subset of the physical quantities, if the variation of one at less of the subset quantities is detected. A computer program product comprising program code instructions recorded on a computer-readable medium including a processor for detecting anomalies in a water continuum, said computer program including program readable programming means. computer for:  receiving measurements of a plurality of physical quantities from a plurality of sensors of the plurality of physical quantities in the water continuum;  generating a plurality of time windows of measurements;  for each window of said plurality:  obtaining a set of values transformed by at least one transformation of the measurements over the time window; applying an anomaly detector to the set of transformed values in order to detect a normal or abnormal state of the time window, said anomaly detector being parameterized from sets of transformed values resulting from the application of the at least one transformation to time windows of previous measurements of the at least one sensor; - detect an anomaly in the water continuum, according to a number of time windows presenting an abnormal state;  execute a set of predefined rules for detecting a variation of at least one of the physical quantities; if an anomaly in the water continuum is detected, assigning to the anomaly a type of anomalies associated with a variation of a subset of the physical quantities, if the variation of at least one of the magnitudes of the sub -set is detected.