FR3084046A1 - METHOD FOR MONITORING THE CONDITION OF AN OVERHEAD CONTACT LINE AND ASSOCIATED PROGRAM DEVICE - Google Patents

Abstract

This monitoring method (200) consists in: recovering a set of signals characteristic of the state of wear of an overhead contact line, and of a state of wear associated with each characteristic signal; generating a reference support vector model from the set of recovered characteristic signals and associated wear states; then acquire (210 to 220), via a measuring device, a raw signal located during the movement of a surveillance vehicle along the overhead contact line to be monitored, the gross localized signal comprising data characterizing the wear of the overhead contact line to be monitored; and determining (230 to 270) the state of wear of the overhead contact line by applying the reference SVM model to data obtained from the localized raw signal.

Description

Method for monitoring the status of an overhead contact line and associated programmed device

The invention relates to that of overhead contact lines and, more particularly, that of methods for monitoring the state of an overhead contact line.

By overhead contact line is meant, in the present invention, not only overhead contact lines in the strict sense, but also neighboring overhead power supply systems, for example catenary.

Malfunctions affecting an overhead contact line are rare, but when a malfunction occurs, such as for example the rupture of the line, the consequences are catastrophic on the operation of the corresponding railway line.

It is therefore necessary to regularly assess the condition of a given contact airline and the likelihood of a malfunction.

Document JP 2017/035973 A describes a process for monitoring the condition of an overhead contact line allowing the development of a real-time and remote diagnosis of the line brought to high potential.

In this prior art document, a railway surveillance vehicle is traveling on the railway whose overhead contact line is to be monitored. The pantograph of the vehicle is equipped with a plurality of sensors making it possible to measure the vibrations of the pantograph in contact with the overhead line during the movement of the vehicle.

From this raw data, a segmentation operation is implemented, which makes it possible to obtain segmented data for each segment of the overhead contact line, a segment being defined between two successive support elements of the overhead contact line. (such as pendulum clocks).

These segmented data are subjected to different statistical processing so as to determine the values of a plurality of characteristic quantities.

Then, a self-organizing map is used. Such a card makes it possible to obtain, from the segmented data, the current state of the corresponding segment: normal state or abnormal state.

It should be noted that this self-adapting card is obtained following a "learning" phase on batches of segmented data corresponding either to contact lines which are known to be in the normal state , or to contact lines which we know to be in an abnormal state.

However, the use of a self-adapting card is not entirely satisfactory. In particular, with this method many types of dysfunctions cannot be distinguished, since they are embedded in the background noise associated with the normal state.

The object of the present invention is therefore to provide an improved monitoring method compared with this state of the art.

For this, the invention relates to a method for monitoring a state of wear of an overhead contact line, the method comprising the following steps:

- recovery of a set of signals characteristic of the state of wear of an overhead contact line, and of a state of wear associated with each characteristic signal;

- generation of a model with a reference support vector, known as the SVM reference model, from the set of recovered characteristic signals and associated states of wear;

- Acquisition, via a measuring device, of a gross signal located during the movement of a surveillance vehicle along the overhead contact line to be monitored, the gross signal localized comprising data characterizing the state of wear of the overhead contact line to be monitored; and

- determination of the wear and tear of the overhead contact line by applying the SVM reference model to data from the localized raw signal.

According to particular embodiments, the method comprises one or more of the following characteristics taken in isolation or according to all the technically possible combinations:

the recovery step comprises a sub-step consisting in supplying a first batch of characteristic signals each associated with a first state of wear known as normal wear state and a second batch of characteristic signals each associated with a second state of wear said abnormal state of wear; and in which the generation step comprises sub-belts consisting of:

- segmenting each characteristic signal to obtain segmented signals, each segmented signal corresponding to an overhead contact line segment between two successive support elements;

- process each segmented signal to obtain, by segment, the values of a set of descriptors;

- select a subgroup of descriptors from the set of descriptors; and,

- build the SVM reference model with the selected subgroup of descriptors.

- the supply of the first and second lots results either from a digital simulation of the interaction between an overhead contact line and a pantograph of a surveillance vehicle, or from tests carried out on a test overhead contact line with a vehicle monitoring ;

- The acquisition step comprises the segmentation of the localized raw signal to obtain a plurality of segmented signals, each segmented signal corresponding to a segment of the overhead contact line to be monitored between two successive support elements; and in which the determining step comprises the following sub-steps consisting in:

- process each segmented signal to extract, by segment, the values of the descriptors of the subgroup of descriptors used to build the reference SVM model; and,

- apply the SVM model of reference to the descriptors extracted to classify each segment of the overhead contact line in the normal state of wear or in the abnormal state of wear.

- a localized raw signal results from the localization, along the overhead contact line, of the pantograph of a surveillance vehicle at each instant of acquisition of a raw signal;

- the raw signal is a measurement at all times of the contact force between the overhead contact line and the pantograph and / or a measurement at all times of the vertical acceleration of the pantograph;

- the step of processing the or each segmented signal consists in applying statistical processing in the time domain, statistical processing in the frequency domain, and / or a wavelet transformation, so as to obtain descriptor values;

- the reference SVM model is multi-class so as to classify a segment in the normal state of wear or the abnormal state of wear, and in the event of classification in the state of abnormal wear, to label the segment with one type of fault among a plurality of possible faults, each possible fault corresponding to a class of the SVM reference model;

The subject of the invention is also a device for monitoring an overhead contact line programmed so as to implement the above monitoring method and a vehicle for monitoring the state of wear of an overhead contact line, including the previous device.

The invention and its advantages will be better understood on reading the detailed description which follows of a particular embodiment, given solely by way of nonlimiting example, this description being made with reference to the appended drawings in which:

- Figure 1 shows, in the form of blocks, the training phase of the method according to the invention; and,

- Figure 2 shows, in the form of blocks, the monitoring phase of an overhead contact line of the method according to the invention.

The method 10 for monitoring an overhead contact line according to the invention is subdivided into a first phase 100, represented in FIG. 1, of determining a reference model according to the technique of support vector machines (“Support Vector Machine "in English or SVM) from batches of training data, and a second phase 200, shown in FIG. 2, of determining the current state of an overhead contact line using the SVM model of reference developed at the end of the first phase 100.

The first phase 100 aims to determine a relevant SVM model allowing to discriminate between an overhead contact line in the normal state and a defective line in the abnormal state. As a variant, and preferably, in this latter state, the SMV model, in the event of an abnormal condition, makes it possible to identify the dysfunction affecting the line from among a set of possible malfunctions.

Different techniques can be envisaged for obtaining batches of training data. For example, it is possible to perform a digital modeling of the overhead contact line and the pantograph interaction of a railway surveillance vehicle with this line, then to use a Monte-Carlo type simulation in order to obtain raw signals. The model can describe a line in the normal state and the signals obtained will constitute a batch of training data in the normal state. The model can describe a line in the abnormal state and the signals obtained will constitute a batch of abnormal state training data. Eventually, a particular malfunction will be introduced into the model to obtain a set of abnormal state learning data specific to this type of malfunction.

Another technique is preferably used, which consists in obtaining batches of training data from tests carried out with an actual overhead contact line and a pantograph of a real test rail vehicle.

Thus in step 105, an overhead test contact line is set up. It has either a normal state or an abnormal state. As a variant, and for this latter state, particular malfunctions from among a set of possible malfunctions are recreated on the line.

The pantograph of a surveillance vehicle is equipped with various sensors. Preferably, a force sensor is used to continuously measure the contact force between the pantograph and the line, and / or an acceleration sensor to continuously measure the vertical acceleration of the pantograph.

In step 110, during the movement of the surveillance vehicle along the test track, the sensors allow the acquisition of raw signals, for example a first raw contact force signal and a second raw signal pantograph acceleration. Raw signals are time dependent.

Depending on how the overhead contact line was prepared (normal contact line, contact line with first fault, second fault, etc.) corresponding batches of raw signals are obtained.

At each instant in the movement of the rail vehicle, its position along the test track is precisely measured using, for example, a GPS positioning device of the satellite type coupled with odometric means. The position of the vehicle is thus obtained along the overhead contact line at each instant of data acquisition.

In step 115, a raw signal is located by means of the positions measured at each instant. Raw localized signals are obtained at the output of step 115.

The surveillance vehicle is advantageously equipped with a system capable of detecting, in a video stream delivered by a camera mounted near the pantograph, the support pendulums of the overhead contact line.

In step 122, the times of observation of the pendulums and therefore their position along the overhead contact line are saved.

As a variant, other processes can be envisaged to obtain the position of the line support elements, such as for example a network map indicating the locations of these elements.

In step 120, each localized raw signal is subdivided into a plurality of segmented signals using the positions of the pendulums obtained in step 122. Each segmented signal corresponds to the portion of the localized raw signal associated with a segment of the overhead line of contact between two successive pendulums.

Segmented signals 125 are thus obtained.

In step 130, each segmented signal is processed by the implementation of different statistical techniques making it possible to extract from each segmented signal a plurality of values of quantities of interest or descriptors.

A first processing can thus consist in applying statistical processing while remaining in the time domain so as to obtain for each segment of the overhead contact line an average of the signal amplitude, a standard deviation, a variance, a minimum value. , a maximum value, a distance between two successive peaks, a coefficient of asymmetry ("skewness") a coefficient of flattening (kurtosis), ...

A second statistical processing is carried out which consists in applying a Fourier transform to the segmented signal to pass into the frequency domain, then in calculating descriptors, such as for example the coefficients of the three main frequencies of the Fourier transform or the spectrum in power in spectral bands corresponding to the wavelength between two successive pendulums.

A third processing is carried out which consists in applying a wavelet transform, then in applying statistical processing so as to obtain values of mean, standard deviation, variance, etc. on the average wavelet power weighted by a scale factor.

If the first treatment is preferably applied to the contact force signal and to that of the pantograph acceleration, the second and third treatments are rather reserved for the contact force signal.

As a variant, only one or the other of these three treatments is carried out. However, it is preferable to carry out several of them simultaneously so as to obtain a large number of descriptors.

The calculated quantities are stored in a table in step 135. This table therefore includes an entry for each segment of the line between two support pendulums and, for each entry, a value for each of the descriptors from the set of descriptors selected. .

Step 140 aims to select the descriptor subgroup among the set of descriptors which allows the best possible separation between a segment in the normal state and a segment in the abnormal state.

Then, a linear discriminant analysis, allowing the calculation of the Fisher score for each descriptor, is applied. This analysis makes it possible to order the descriptors of all the descriptors according to their impact on the separation between a segment in the normal state and a segment in the abnormal state.

In addition, a two-class SVM model is implemented. The two-class case makes it possible to distinguish between a normal state and an abnormal state. The method then appears as a method of detecting a malfunction of the overhead contact line.

As a variant, the SVM model implemented is multi-class so as to make it possible to distinguish not only between the normal state and the abnormal state, but also among various possible malfunctions each having a particular signature. The process then appears as a process for detecting a malfunction and identifying the nature of the malfunction affecting the overhead contact line.

Step 140 of selecting the most relevant quantities is therefore based on a linear discriminant analysis and an SVM model. Reducing the number of descriptors to be considered, on the one hand, and selecting the most relevant for training the SVM model, on the other hand, makes it possible to reduce the calculation time and the resources allocated to this calculation. In this way, the real-time analysis aspect is improved.

More specifically, step 140 is performed iteratively. The descriptors are first ordered according to the Fisher score obtained for each of them. Then several subgroups of descriptors are formed

The different subgroups are formed by selecting the first and the second descriptor, then the first, the second and the third descriptor, then the first, the second, the third and the fourth descriptor, and so on until the last descriptor .

For each selected subgroup, an SVM model is trained.

Each SVM model trained from a subgroup of descriptors is evaluated using the score F1 advantageously defined by:

Precision Recall

Fi = 2-Precision + Recall

TP

Precision = ————

TP + F P

TN

Recall = ————

T N + F N where TP is the number of true positives, TN is the number of true negatives, FP is the number of false positives and FN is the number of false negatives. During the iterations, the subgroup which allowed to train the SVM with the best value of the classification F1 score is selected.

At the end of step 140, the best subgroup of characteristics is therefore determined and makes it possible, in step 150, to build the SVM model of reference 155.

Once the reference SVM model has been determined, it is used in real time on the data acquired on a contact line to be diagnosed.

Thus, the second phase 200 of method 10 is carried out during the movement of the railway surveillance vehicle along a railway whose state of the overhead contact line must be determined.

The second phase 200 repeats some of the steps of the first phase 100. The corresponding steps are identified by a reference number increased by one hundred.

Thus, the second phase 200 comprises a step 210 of acquiring the contact force between the pantograph and the contact line; acquisition of the pantograph acceleration: as well as the determination of the position of the test vehicle at all times so as to be able to locate in step 215 the raw signals acquired.

Each localized raw signal is then segmented in step 220 knowing the positions of each of the support pendulums of the overhead contact line determined during step 222. Segmented signals are obtained (step 225).

These segmented signals are processed in step 230 to extract only the values of the descriptors corresponding to the optimal subgroup which has allowed the development of the reference SVM model.

A table of characteristic sizes 235 is constructed. This table includes an entry for each segment of the contact line to be diagnosed and, for each entry, the value of the descriptors belonging to the optimal subgroup.

In a step 270, the reference model SVM of reference obtained at the end of the step 150 is then used to classify each entry of the table 235 either in the normal state or in the abnormal state for the case of a dual-class SVM model.

For the variant in which the reference SVM model is multiclass, in step 270, the reference SVM model makes it possible to classify each entry of the table 235 either in the normal state or in the abnormal state, and for the abnormal state of associating a type of dysfunction among the possible types, which each correspond to a class of the SVM model of reference.

In step 275, the state of each segment of the line to be diagnosed is labeled with a normal or abnormal state (and possibly with a type of dysfunction affecting this segment). This information is stored in the updated table 235.

Advantageously, the method according to the invention is implemented by a programmed device on board the surveillance vehicle and connected to the various sensors (for example contact force, acceleration, or other), by means of locating the train, and possibly detection of the overhead contact line support pendulums. This device is suitable for storing in a database the table 235 associating with each segment of the line a detected state and possibly a type of identified malfunction.

Thus, a person skilled in the art will recognize that the main contribution of the present invention consists in training a model based on an algorithm of the SVM type which will then make it possible to predict the state and possibly the type of fault affecting the overhead contact line. The method therefore makes it possible to determine the current state of the line and to detect and locate the occurrence of fault (s). This process is carried out in real time and remotely from the elements brought to high voltage.

Claims (10)

1. - Method (100, 200) for monitoring a state of wear of an overhead contact line, the method comprising the following steps: - recovery (110) of a set of signals characteristic of the state of wear of an overhead contact line, and of a state of wear associated with each characteristic signal; - generation (120 to 150) of a model with a reference support vector, known as the SVM reference model, from the set of recovered characteristic signals and associated states of wear; - Acquisition (210 to 220), via a measuring device, of a raw signal located during the movement of a surveillance vehicle along the overhead contact line to be monitored, the gross localized signal comprising data characterizing the state of wear of the overhead contact line to be monitored; - determination (230 to 270) of the state of wear of the overhead contact line by applying the SVM reference model to data from the localized raw signal. 2. - Method (100, 200) according to claim 1, wherein the recovery step (110) comprises a substep consisting in supplying a first batch of characteristic signals each associated with a first state of wear known as state d normal wear and a second batch of characteristic signals each associated with a second state of wear called abnormal state of wear; and in which the generation step (120 to 150) comprises sub-belts consisting of: - segmenting (120) each characteristic signal to obtain segmented signals, each segmented signal corresponding to an overhead contact line segment between two successive support elements; - process (130) each segmented signal to obtain, by segment, the values of a set of descriptors; - select (140) a subgroup of descriptors from the set of descriptors; and, - construct (150) the SVM reference model with the selected subgroup of descriptors. 3. - Method (100, 200) according to claim 1 or claim 2, wherein the supply of the first and second batches results either from a digital simulation of the interaction between an overhead contact line and a pantograph of a surveillance vehicle, i.e. tests carried out on an overhead test contact line with a surveillance vehicle. 4. - Method (100, 200) according to claim 2, in which the acquisition step (210 to 220) comprises the segmentation (220) of the localized raw signal to obtain a plurality of segmented signals, each segmented signal corresponding to a segment of the overhead contact line to be monitored between two successive support elements; and in which the determination step (230 to 270) comprises the following sub-steps consisting in: processing (230) each segmented signal to extract, by segment, the values of the descriptors of the descriptor subgroup used to construct the reference SVM model; and, applying (270) the SVM model of reference to the descriptors extracted to classify each segment of the overhead contact line in the normal state of wear or in the abnormal state of wear. 5. - Method (100, 200) according to any one of claims 1 to 4, in which a localized raw signal (115, 215) results from the localization, along the overhead contact line, of the pantograph of a surveillance vehicle at each instant of acquisition of a raw signal. 6. - Method (100, 200) according to claim 5, wherein the raw signal is a measurement at each instant of the contact force between the overhead contact line and the pantograph and / or a measurement at each instant of the vertical acceleration of the pantograph. 7. - Method (100, 200) according to claim 2 or claim 4, wherein the processing step (130, 230) of the or each segmented signal consists in applying a statistical processing in the time domain, a statistical processing in the frequency domain, and / or a transformation into wavelets, so as to obtain descriptor values. 8. - Method (100, 200) according to any one of the preceding claims, in which the reference SVM model is multi-class so as to classify a segment in the normal state of wear or the abnormal state of wear, and in the event of classification in the abnormal state of wear, of labeling the segment with a type of defect among a plurality of possible defects, each possible defect corresponding to a class of the reference model SVM. 5 9.- Device for monitoring an overhead contact line, characterized in that it is programmed so as to implement the monitoring method (100, 200) according to any one of claims 1 to 8. 10.- Vehicle for monitoring the state of wear of an overhead contact line, characterized in that it comprises a device according to claim 9.