FR3126683A1 - Railway analysis device - Google Patents

Abstract

Railway analysis device A railroad analysis device comprises a memory (4) arranged to receive railroad track measurement data each comprising at least one railroad track measurement value and a dated measurement path identifier, an extractor (6) arranged for determining a set of features from the railway track measurement data, which set of features includes at least two features selected from a group comprising deviation features, leveling features, left features, cant features , rolling stock characteristics, and historical characteristics, and a gradient enhancement based machine learning unit (8) trained on rail track measurement data associated with rail track fault values, which track fault railway is chosen from a group comprising at least one defect from among a cant defect, a defect of n ively and a left fault, and arranged to return one or more rail track fault values. Fig.1

Description

Railway analysis device

The invention relates to the field of railway analysis and in particular the predictive maintenance of railway tracks.

The field of predictive maintenance of railway infrastructures has been studied for several decades. Indeed, it is critical to maintain rail infrastructure in good working order in order to guarantee the safety of users and goods. However, even the simplest infrastructure quickly includes several hundred kilometers of rails, making manual maintenance cost-prohibitive.

In order to automate data collection, machines, such as the EM80H locomotive manufactured by Plasser & Theurer (see for example https://www.plassertheurer.com/fr/machines-systemes/voitures-autophériques-de- measure/em80h) have been developed.

The American company RailPod (see https://rail-pod.com/?page_id=2263) offers an autonomous measurement device and algorithms for analyzing the geometry of tracks and their condition (erosion, surface condition, internal structural problem).

The FleetONE solution from the specialist Wabtec Controls Systems (http://trackiq.com.au/FleetONE.html) offers a system for acquiring and managing railway maintenance data, mainly for rolling stock. It offers the ability to search and visualize simple data. No predictive rail maintenance functionality appears to be offered.

None of these solutions offers the possibility of predictive maintenance.

Predictive maintenance consists in trying to determine portions of railway tracks which are the most critical to maintain. This determination is a source of many challenges.

It seems obvious that predictive maintenance requires comparing measurement signals acquired successively in order to analyze the evolution of track sections. This requires synchronizing the measurement signals between different acquisitions because these vary greatly due to the way in which they are determined (variance on the acquisition devices, determination of the portions by odometry, by GPS or by singular point inducing ambiguities , etc.). Numerous patent applications have been filed on this subject, such as document EP 2 305 532 which describes the determination of a cross-correlation on overlapping sections, document US 2009/0070064 which describes a method for determining a trend from vertical position deviations of the track, or the document US 2004/0204882 which maps track measurement data (MTD for "Measured Track Data" in English) on geographical track data (TGD for "Track Geographical Data" in English). Most of these methods are imprecise and/or tedious.

Then, once the signals are synchronized, it is a question of choosing how to identify the evolutions of the problematic portions. Here again, no commercial offer really exists to date, and the state of the art is mainly based on academic work. Iman Soleimanmeigouni's article "Track geometry degradation and maintenance modelling: A review", https://journals.sagepub.com/doi/10.1177/0954409716657849 reviews about a hundred articles, none of which brings real satisfaction . Finally, some prospective works have sought to propose the use of artificial intelligence to carry out predictive maintenance. In general, like the article “Systematic Literature Review on Data-Driven Models for Predictive Maintenance of Railway Track: Implications in Geotechnical Engineering” by Xie et al (https://www.mdpi.com/2076-3263 /10/11/425/pdf) suggests that deep and unsupervised learning is the most promising area. But here again, no method seems to have succeeded in going beyond the academic stage and being sufficiently reliable from an industrial point of view.

The invention improves the situation. To this end, it proposes a railway analysis device comprising a memory arranged to receive measurement data from a railway track each comprising at least one railway track measurement value and a dated measurement route identifier, an extractor arranged for determining a set of features from the railway track measurement data, which set of features includes at least two features selected from a group comprising deviation features, leveling features, left features, cant features , rolling stock characteristics, and historical characteristics, and a gradient enhancement based machine learning unit trained on rail track measurement data associated with rail track fault values, which rail track fault is selected among a group comprising at least one fault among a cant fault, a leveling fault and a left fault, and arranged to return one or more railway track fault values.

This device is particularly advantageous because it makes it possible to offer a systematic, reliable and automated analysis of railway track faults, which was not possible previously.

According to various embodiments, the invention may have one or more of the following characteristics:

- the device further comprises a corrector arranged to receive current rail track measurement data associated with the same measurement route identifier, the measurement values of which indicate a cant measurement, to analyze these rail track measurement data in order to 'identify beginnings and ends of curves, and to modify the current rail track measurement data by comparing the beginnings and ends of curves with the beginnings and ends of reference path measurement data curves,

- the extractor determines a set of characteristics which includes at least one left characteristic, at least one deviation characteristic, at least one leveling characteristic, at least one cant characteristic, at least one rolling stock characteristic and at least a historical feature,

- the measurement data of a railway track further comprises a section identifier, and the extractor is arranged to process together the measurement data of a railway track associated with the same dated measurement route identifier and with the same section,

- the extractor is arranged to determine a set of characteristics comprising a left characteristic, a deviation characteristic, a leveling characteristic and a cant characteristic of the maximum value type, minimum value, value at percentile 90, value at percentile 10 in the railway track measurement data associated with the same identifier of previous dated measurement route and with the same section,

- the extractor is arranged to determine a set of characteristics comprising at least one historical characteristic which is a leveling characteristic, a cant characteristic, or a rolling stock characteristic determined for the measurement data of a railway track associated with a same section and to an earlier dated measurement route identifier,

- the extractor is arranged to determine a set of characteristics comprising a characteristic of rolling stock of the type number of trains passed in the last 7, 15 or 30 days, or of the type tonnage of rolling stock passed in the last 7, 15 or 30 days, and

- the machine learning unit based on gradient enhancement uses a binary classification with a logistic loss function.

The invention also relates to a method for predicting a railway track fault value, comprising the following operations:
a) receiving rail track measurement data each comprising at least one rail track measurement value and a dated measurement route identifier,
b) determining a feature set from the rail track measurement data, which feature set includes at least two features selected from a group comprising deviation features, leveling features, left features, cant, rolling stock characteristics, and historical characteristics,
c) transmitting the feature set determined in step b) to a gradient enhancement based machine learning unit trained on rail track measurement data associated with rail track fault values, which rail track fault is chosen from a group comprising at least one fault from among a cant fault, a leveling fault and a left fault, and arranged to return one or more railway track fault values,
d) returning the railway track fault value(s) determined for the set of characteristics determined in step b).

The invention finally relates to a computer program product comprising instructions for implementing the method when it is executed on a computer, and to a storage medium on which the computer program product is recorded.

Other characteristics and advantages of the invention will appear better on reading the following description, taken from examples given by way of illustration and not limitation, taken from the drawings in which:

There represents a generic diagram of a device according to the invention,

There shows an example of a left fault for a railway track,

There shows an example of a cant defect for a railway track,

There shows an example of a leveling defect for a railway track,

There shows an example of a deviation fault for a railway track,

There represents an example of a cant signal received at the input, and

There shows an example of a signal synchronization function of the with a theoretical signal.

The drawings and the description below contain, for the most part, certain elements. They may therefore not only be used to better understand the present invention, but also contribute to its definition, if necessary.

There represents a generic diagram of a device according to the invention. The railway analysis device 2 comprises a memory 4, an extractor 6, a machine learning unit based on gradient enhancement 8 and an optional corrector 10.

The memory 4 can be any type of data storage capable of receiving digital data: hard disk, hard disk with flash memory, flash memory in any form, random access memory, magnetic disk, storage distributed locally or in the cloud, etc. The data calculated by the device can be stored on any type of memory similar to memory 4, or on the latter. This data can be erased after the device has performed its tasks or retained.

In the example described here, memory 4 receives measurement data from a railway track. These measurement data are in the example described here for the most part received from an EM80H locomotive, called “Plasser”. This type of locomotive is specifically dedicated to the measurement of railway track in order to ensure its monitoring and maintenance. The outputs of this type of locomotive are typically signals as shown in .

In the example described here, the memory 4 thus receives four measurement signals in the form of samples: a left measurement signal, a tilt measurement signal, a leveling measurement signal, and a leveling measurement signal. gap.

Figures 2 to 5 represent the respective nature of the left, cant, leveling and deviation faults. Thus, a left-hand fault is observed when a portion of a rail is raised relative to the other rail, a cant fault is observed when a rail is generally raised relative to the other rail, a leveling defect is observed when the two rails have a raised portion when they should remain flat, and a deviation defect is observed when the rails deviate axially with respect to the direction they should have.

In order to be able to use and compare these signals, each of the data composing them is associated on the one hand with a dated route identifier, that is to say an identifier making it possible to know which railway track was measured, and on which date this measurement corresponds, and on the other hand, to a section identifier. Indeed, as we will see below, the measurement data are analyzed in the example described here by sections of 50 m in length.

As a variant, the sections could be shorter, for example by 25 m, when sufficient precision is available on the synchronization of the signals, or longer, for example by 100 m. The shorter the length, the lower the risk of the presence of non-homogeneous sections of track between them in the same section. By non-homogeneous portions of track between them in the same section, it is meant the fact that, by the nature of the course, portions of the same section behave differently with respect to the appearance of faults and that the section is therefore “incoherent” from this point of view. This may be the case if the length of the section is large, typically 200 m, and part of the section is close to a place subject to a lot of vibration, or a place where the soil composition is very different. the rest of the section. In the case of railway tracks, the characteristic lengths (bends, stability of the terrain, etc.) are quite significant, and the length of 50 m therefore appears naturally advantageous.

Thus the measurement data is formed by a measurement value of a type of defect, a dated route identifier and a section identifier.

Alternatively, device 2 could receive raw channel measurement data, and calculate the measurement signals described above. Still as a variant, the dated route identifier could be a simple date, the link between the section identifiers of the same route could then be implicit.

In addition to these signals, memory 4 receives other information, such as the number of train passages on the railway track between two measurements, the tonnage of rolling stock between two measurements, these data forming railway stock data.

The measurement data of a given section and of a given date can also be associated with the measurement data of this same section of the previous measurements. These "past" measurement data form historical data and can be replicated as historical data associated with the measurement data or simply associated on the basis of their dated measurement route identifier and their section identifier.

As a variant, the measurement data could be simplified so that a section identifier is associated with only one measurement data item per type of measurement. This amounts to making a section of very small size, or for example to averaging all the measurement data of a given section. The result obtained will be less precise but calculation savings can be made.

The measurement data received in the memory 4 is used directly or indirectly by the extractor 6 in order to produce a set of characteristics which is supplied to the automatic learning unit based on gradient enhancement 8 to carry out the calculation of the values of predictive maintenance. The terms " gradient boosting " and " characteristic " should be interpreted with the meaning given to them in the field of machine learning, i.e. " gradient boosting " and " feature " respectively in English.

As described above, the measurement data may relate to characteristic measurements of faults that a railway track may have. Thus, the extractor 6 is arranged to determine, for the measurement data associated with a given dated route identifier, a plurality of characteristics which will be supplied to the input layer of the reinforcement-based automatic learning unit of the gradient 8.

In the example described here, for each section identifier, each type of fault is associated with four characteristics: the maximum value among the measurement values concerned, the minimum value among the measurement values concerned, the measurement value of the 90th percentile among the measurement values concerned, and the measurement value of the 10th percentile among the measurement values concerned. As a variant, only some of these values could be retained, such as the measurement value of the 90th percentile among the measurement values concerned.

The extractor 6, the machine learning unit based on gradient enhancement 8 and the corrector 10 are elements directly or indirectly accessing the memory 4. They can be implemented in the form of an appropriate computer code executed on a or more processors. By processors, it must be understood any processor suitable for the calculations described below. Such a processor can be produced in any known manner, in the form of a microprocessor for a personal computer, a dedicated chip of the FPGA or SoC type, a computing resource on a grid or in the cloud, a cluster of graphics processors (GPUs), a microcontroller, or any other form capable of providing the computing power necessary for the implementation described below. One or more of these elements can also be made in the form of specialized electronic circuits such as an ASIC. A combination of processor and electronic circuits can also be envisaged. In the case of the machine learning unit based on gradient reinforcement 8, processors dedicated to machine learning can also be considered.

The extractor 6 is also arranged to establish characteristics in the characteristic set on the basis of the railway material data and on the basis of the historical data.

Thus, for the railway equipment data, the extractor 6 can determine the characteristics of railway equipment from the number of train passages and the tonnage of rolling stock having circulated in the last 7 and/or 15 and/or 30 days preceding the date associated with the route identifier dated on the associated route. Similarly, historical features may include gap features, grade features, left features, superelevation features for each leg identifier that corresponds to the first and/or second and/or third and/or fourth previous dated route identifier, for each section identifier, each time with the number of days elapsed. The historical characteristics may also include field readings by operators (geometry faults, general condition of tracks and rails), topography (field models) or even maintenance work.

The set of characteristics thus established for each section identifier is then supplied to the input of the automatic learning unit based on gradient reinforcement 8. In the example described, this is a tree with gradient reinforcement. In the example described here, the LightGBM version 3.0.0 library, maintained by Microsoft, is used from Python with the 'LightGBMClassifier' function and the following hyperparameters.

- Type of models (‘boosting_type’): ‘gbdt’ – the usual gradient boosted tree

- Number of terminal leaves ('num_leaves'): 50,

- Maximum depth of the tree ('max_depth'): no limit,

- Boosting learning rate ('learning_rate'): 0.1,

- Number of boosted trees in the model ('n_estimators'): 1000,

- Objective function (‘objective’): binary classification (‘binary’ with logistic loss function – log-loss),

- Random sampling rate (without replacement) on the data set to build a tree ('subsample'): 0.7 or 70%,

- Random sampling rate (without replacement) on the input variables to build a tree ('colsample_bytree'): 50%,

- Weighting of each class ('class_weight'): balanced for binary classifications.

The other hyperparameters are less important for the configuration of the algorithm and are the default ones of the 'LightGBMClassifier' function.

The Applicant has identified that this type of machine learning is by far the most effective, ahead of logistic regression methods, SVM (Support Vector Machine), or Random Forest methods. , but also compared to methods based on deep learning, which are more suitable for image processing.

The training of the machine learning unit based on gradient reinforcement 8 was carried out from the existing maintenance history for the railway lines of the Gabonese railway, with as a label the measurement of the faults which were actually measures. Ideally, the gradient-enhancement based machine learning unit 8 could be trained specifically for each railway track to which it is to be applied.

The most favorable results were obtained with the use of all the characteristics described above. Alternatively, the machine learning unit based on gradient enhancement 8 could use only one feature per feature category, or even just one feature at all.

Corrector 10 is optional. Indeed, in the foregoing, it was considered that the measurement data were already synchronized with each other. Nevertheless, the Applicant has found that the history of the measurements taken by the EM80H locomotive is difficult to reconcile between two passages. This is due to: the lack of accuracy of the starting point of a measurement campaign, and the wear and slippage of the wheels of the locomotive (the measurement location being determined by odometry). It is therefore desirable to synchronize the measurement data with the measurement data already present in the memory 4. The Applicant has discovered that it is particularly advantageous to use the cant measurement data for this purpose.

There represents an example of implementation of a function making it possible to synchronize the measurement data comprising a new dated route identifier with the measurement data already present in the memory 4 for this route.

In a first operation 700, this function receives the measurement data from the measurement data to be synchronized via an Inp() function. As an option, the Inp() function can also retrieve the reference data associated with the course identifier dated from the measurement data received by the Inp() function. Then, in an operation 710, this data is low-pass filtered by a Filt() function. This makes it possible in particular to eliminate certain disturbances linked to the precision of the locomotive, as well as to the state of the railway track.

Once the signal has been smoothed, the resulting signal is analyzed by a Detect() function in an operation 720 in order to determine the points for which the cant measurement exceeds 15 mm. These points physically represent the start and end of each curve. Alternatively, this threshold could be 5 mm or less, or the start and end points of the curve could be determined, for example based on a derivative of the signal.

These start and end points are then used by a Match() function in an operation 730 to compare them to reference data. These reference data may correspond to measurement data in which absolute confidence is granted, or even a plan of the railway track during its construction. The association by the Match() function between the characteristic points of the smoothed signal and those of the reference data can be made by analyzing the similarity of the resulting curves, for example on the basis of the order of these points and the direction and the amplitude of the corresponding cant signal. This pairing is particularly advantageous because it is quite simple to implement and makes it possible to take into account the fact that the characteristic points and their sequence are unique to each railway track. Thus, it is possible to achieve a precision of the order of 50 centimeters for the measurement data processed by the corrector 10, as opposed to the precisions of the order of about ten meters of the solutions of the state of the 'art.

Finally, a Sync() function modifies in an operation 740 the measurement data on the basis of the readjustments of the characteristic points identified in the operation 730. For this, the beginning and the end of each part of the corresponding signal is aligned on the basis of this readjustment. This induces a translation and/or a compression/dilation of the signal concerned. This same transformation is applied to all the signals associated with the same dated route identifier.

Claims (11)

Railway analysis device comprising a memory (4) arranged to receive measurement data from a railway track each comprising at least one railway track measurement value and a dated measurement route identifier, an extractor (6) arranged to determining a set of features from the railway track measurement data, which set of features includes at least two features selected from a group comprising deviation features, leveling features, left features, cant features, rolling stock characteristics, and historical characteristics, and a gradient enhancement based machine learning unit (8) trained on rail track measurement data associated with rail track fault values, which rail track fault is chosen from a group comprising at least one fault from among a cant fault, a leveling fault and a left fault, and arranged to return one or more railway track fault values. Railway analysis device according to claim 1, further comprising a corrector (10) arranged to receive current railway track measurement data associated with the same measurement route identifier, the measurement values of which indicate a cant measurement, for analyzing the track measurement data to identify curve starts and ends, and to modify the current track measurement data by comparing the curve starts and ends to the curve starts and ends of measurement data reference path. Device according to one of the preceding claims, in which the extractor (6) determines a feature set which comprises at least one left feature, at least one deviation feature, at least one leveling feature, at least one cant, at least one rolling stock characteristic and at least one historical characteristic. Device according to one of the preceding claims, in which the measurement data of a railway track further comprises a section identifier, and in which the extractor (6) is arranged to process the measurement data of a track together associated with the same dated measurement route identifier and the same section. Device according to Claim 4, in which the extractor (6) is arranged to determine a set of characteristics comprising a left characteristic, a deviation characteristic, a leveling characteristic and a cant characteristic of the type maximum value, minimum value , value at percentile 90, value at percentile 10 in the rail track measurement data associated with the same identifier of the previous dated measurement route and with the same section. Device according to Claim 4 or 5, in which the extractor (6) is arranged to determine a set of characteristics comprising at least one historical characteristic which is a leveling characteristic, a cant characteristic, or a rolling stock characteristic determined for the measurement data of a railway track associated with the same section and with a previous dated measurement route identifier. Device according to one of the preceding claims, in which the extractor (6) is arranged to determine a set of characteristics comprising a rolling stock characteristic of the type number of trains passed in the last 7, 15 or 30 days, or of the type tonnage of rolling stock passed in the last 7, 15 or 30 days. Device according to one of the preceding claims, in which the gradient-enhanced machine learning unit (8) uses a binary classification with a logistic loss function. A method of predicting a railway track fault value, comprising the following steps:
a) receiving rail track measurement data each comprising at least one rail track measurement value and a dated measurement route identifier,
b) determining a feature set from the rail track measurement data, which feature set includes at least two features selected from a group comprising deviation features, leveling features, left features, cant, rolling stock characteristics, and historical characteristics,
c) transmitting the feature set determined in step b) to a gradient enhancement based machine learning unit (8) trained on rail track measurement data associated with rail track fault values, which fault railway track is chosen from a group comprising at least one fault from among a cant fault, a leveling fault and a left fault, and arranged to return one or more railway track fault values,
d) returning the railway track fault value(s) determined for the set of characteristics determined in step b). Computer program product comprising portions of program code for implementing the device according to one of Claims 1 to 8 or the method according to Claim 9 when said program is executed on a computer. Data storage medium on which the program according to claim 10 is recorded.