Method for determining an indicator of structural defectiveness

Abstract

A method is proposed for determining an indicator (I) of structural defectiveness of an element (10) of a running gear (20) of a land vehicle (100), the method comprising the steps of: (i) acquiring during a time window (W) a vibratory signal (V) associated with the element (10) of the running gear (20), (ii) determining a sequence (S) of spectral components (A1, A2, ..., Ap) of the vibratory signal (V) acquired during the time window (W), (iii) iterating steps (i) and (ii) for a set of time windows (W1, ..., Wn) so as to obtain a set of sequences (S1,..., Sn) of spectral components of the vibratory signal (V), each sequence (Sk) of the set of sequences (S1,..., Sn) corresponding to one time window (Wk), (iv) determining an indicator (I) of structural defectiveness of the element (10) of the running gear (20) based on the set of sequences (S1,..., Sn) of spectral components of the vibratory signal (V) and based on a detection model implementing a machine-learning algorithm.

Description

Method for determining a structural defect indicator technical field [1] The present invention relates to the field of methods for detecting mechanical defects in the ground connection elements of vehicles, such as trucks or trains. Previous technique [2] Some vehicles have one or more axles on which the vehicle chassis rests via suspension springs, such as leaf springs. During the vehicle's lifetime, a spring may become damaged due to mechanical stress. Cracks can gradually propagate through one or more of the spring's leaves. The propagation of these cracks can sometimes lead to total spring failure due to fatigue, which can create a dangerous situation and immobilize the vehicle. [3] It is known to monitor the condition of suspension springs during vehicle use to detect damage before it can lead to failure. This is done by measuring the acceleration of the spring itself, or of the axle. Comparing the measured acceleration with a threshold allows for the diagnosis of spring damage. However, due in particular to variations in operating conditions, such as vehicle load, speed, or road profile, known methods are not entirely satisfactory. Indeed, known methods often exhibit excessively high non-detection and false detection rates. Furthermore, most known methods only allow for qualitative, not quantitative, diagnosis. In other words, the presence of a defect can be detected, but the severity of the detected defect remains difficult to assess. [4] The present disclosure aims to improve this situation by proposing an improved method to establish the presence of a defect in a more precise and robust manner, so as to be able to anticipate corrective action. Summary [5] To this end, the invention proposes a method for determining a structural defect indicator for a component of a land vehicle's running gear, the method comprising the following steps: - (i) acquire, during a time window, a vibration signal associated with the train element rolling, - (it) determine a sequence of spectral components of the vibration signal acquired during the time window, - (iii) iterate steps (i) and (ii) for a set of time windows so as to obtain a set of sequences of spectral components of the vibration signal, each sequence in the set of sequences corresponding to a time window, - (iv) determine a structural fault indicator of the running gear element from the set of spectral component sequences of the vibration signal and from a detection model implementing a machine learning algorithm. [6] The features listed in the following paragraphs can be implemented independently of each other, or in any technically possible combination: [7] According to one embodiment of the process, the spectral components of the sequence of spectral components are determined by fast Fourier transform of the samples of the vibration signal acquired during the time window. [8] Each spectral component of a sequence of spectral components of the vibration signal is thus a value calculated over the entire corresponding acquisition time window. [9] According to one example of implementation of the process, the structural defect is a crack in the running gear element.

[0010] According to one example of implementation of the process, the running gear element is an elastic suspension element of the vehicle.

[0011] The elastic suspension element is, for example, a leaf spring.

[0012] The leaf spring is positioned between a vehicle axle and a vehicle chassis.

[0013] Alternatively, the elastic suspension element can be a helical spring.

[0014] Alternatively, the elastic suspension element can also be a torsion bar.

[0015] Alternatively, the elastic suspension element can be an air cushion.

[0016] The vehicle chassis can be a self-supporting vehicle body.

[0017] According to an example of the application of the process, the structural defect is a defect in a wheel bearing connected to the axle.

[0018] A structural defect such as damage to a ball or part of the raceway of one of the bearing rings alters the vibration response of the running gear and can be detected by analyzing the acquired vibration signal.

[0019] According to another example of application of the process, the structural defect is a structural defect of a tire of a wheel connected to the axle.

[0020] As before, a structural defect such as a tire deformation in the running gear changes the vibration response of the running gear and can be detected by analyzing the acquired vibration signal.

[0021] In yet another example of the application of the process, the structural defect is an axle overload.

[0022] Similarly, excessive weight carried by the axle can be detected by analyzing the acquired vibration signal.

[0023] According to yet another application of the process, the structural defect is an excessive reduction in the damping coefficient of the relative movements of the axle with respect to the vehicle chassis. Wear on the axle oscillation dampers can also be detected through analysis of the acquired vibration signal.

[0024] In one example of the application of the process, the vehicle is a truck, or a van.

[0025] In another example of the application of the process, the vehicle is railway rolling stock, such as a train wagon or locomotive.

[0026] According to one embodiment of the process, the vibration signal is an acceleration signal from an acceleration sensor.

[0027] The acceleration sensor, for example, is located on a component of the vehicle's running gear.

[0028] In other words, the acceleration sensor is then placed on the element whose operating state we want to monitor.

[0029] According to another implementation of the process, the acceleration sensor is placed on the axle.

[0030] According to yet another implementation of the process, the acceleration sensor is placed on the vehicle chassis.

[0031] In these two implementation examples, the acceleration sensor is mounted on a different component than the one being monitored, while still being mechanically connected to it. This simplifies the installation of the acceleration sensor by increasing the number of possible mounting locations.

[0032] The process can jointly use the signal from several acceleration sensors arranged on different elements of the running gear or on the chassis.

[0033] According to one embodiment, the acceleration sensor is a three-axis sensor.

[0034] Alternatively, the acceleration sensor can be a single-axis sensor.

[0035] The sampling frequency is greater than 30 Hz. According to one embodiment, the acquisition frequency, also called the sampling frequency, of the vibration signal of the accelerometer is between 25 kHz and 35 kHz, for example equal to 32 kHz.

[0036] According to one embodiment of the process, each sequence of the set of spectral component sequences of the vibration signal is obtained by averaging a plurality of spectral component sequences corresponding to a plurality of elementary acquisition windows of the vibration signal.

[0037] The plurality of elementary acquisition windows can include between 5 acquisition windows and 100 acquisition windows.

[0038] The duration of a vibration signal acquisition window W is configurable. For example, the duration of a vibration signal acquisition window can range from 2 seconds to 200 seconds. It can also be set to 100 seconds.

[0039] The time windows for acquiring the vibration signal can be contiguous.

[0040] According to one embodiment, the process comprises one step: - (iii1) determine the vehicle's forward speed, - (iii2) eliminate from the set of acquired time windows the time windows corresponding to a vehicle forward speed below a predetermined minimum threshold.

[0041] Eliminating these acquisition windows improves the signal-to-noise ratio, as these acquisition windows correspond to vehicle operating phases during which the vibration signal is weak.

[0042] The process may include one step: - (iii3) eliminate from the set of acquired time windows the time windows corresponding to a zero speed of the vehicle.

[0043] Alternatively, or in addition, the process includes one step: - (iii2') eliminate from the set of acquired time windows the time windows corresponding to a vehicle forward speed greater than a predetermined maximum threshold.

[0044] Eliminating these acquisition windows helps to avoid potential saturation due to a high signal level.

[0045] The process may include a data normalization step. This means that the signal scale is reframed so as to have a mean of zero and a standard deviation equal to 1. The data used for the subsequent steps of the process are therefore those of the set of normalized spectral component sequences.

[0046] Step (iv) of determining a process structural defect indicator includes a step (iv1) of reducing the dimensions of the set of spectral component sequences of the vibration signal, so as to obtain a set of reduced sequences of combinations of spectral components of the vibration signal.

[0047] Each reduced sequence has fewer spectral components than the number of spectral components determined in step (ii).

[0048] The dimensionality reduction step reduces the complexity of the calculations performed, while retaining the components that best characterize the acquired data.

[0049] According to one embodiment of the process, the step (iv1) of dimensioning the set of spectral component sequences of the vibration signal implements an unsupervised dimensioning algorithm.

[0050] An unsupervised algorithm allows for easy adaptation to a wide variety of technical definitions. Indeed, an unsupervised machine learning algorithm makes it possible to build mathematical models based on actually observed data, which in this case are the acquired samples of the vibration signal.

[0051] Various unsupervised dimensionality reduction algorithms can be used.

[0052] For example, an auto-encoding network can be used. The step (iv1) of dimension reduction of the set of sequences of spectral components of the vibration signal is thus carried out by encoding a sequence of spectral components of the vibration signal, using an encoder of a pre-trained self-encoding network, in a reduced dimension latent space.

[0053] A principal component analysis algorithm is another example of a usable algorithm.

[0054] The unsupervised dimensionality reduction algorithm generates a dimensionality reduction model.

[0055] The set of reduced sequences of spectral components can thus be determined by a model obtained through machine learning.

[0056] The process can therefore automatically adapt to the mechanical configuration on which the process is used.

[0057] According to one example of implementation of the process, a reduced sequence of spectral components includes, for example, seven elements.

[0058] According to another example of implementation of the process, a reduced sequence of spectral components includes, for example, three elements.

[0059] According to yet another example of implementation of the process, a reduced sequence of spectral components includes, for example, two elements.

[0060] The associated calculations can thus be reduced by a large factor, compared to the initially acquired data.

[0061] The dimensionality reduction model of the set of spectral component sequences of the vibration signal can be obtained by machine learning of the vibration signal from a subset of time windows acquired under reference conditions in which the running gear element is free of structural defects.

[0062] The process includes one step: - (iv2) determine, for each time window, a reconstructed sequence of spectral components of the vibrational signal acquired during the time window from the reduced sequence of combinations of spectral components and from a reconstruction model obtained from the dimensionality reduction model.

[0063] The model for reconstructing the spectral components of the acquired vibration signal is thus based on automatic learning of the vibration signal from a subset of time windows acquired under reference conditions in which the running gear element is free of structural defects.

[0064] According to one aspect of the process, the reconstruction model of spectral components of the acquired vibration signal is obtained by inverting the dimensionality reduction model.

[0065] The reconstruction model can thus be easily determined from the model already established for the dimensionality reduction of the acquired data.

[0066] According to one aspect of the disclosure, the process includes one step: - (iv3) Determine, for each time window, a difference between the sequence of spectral components of the acquired vibration signal and the reconstructed sequence of spectral components of the vibration signal, and the structural defect indicator is based on the determined difference.

[0067] In the absence of a structural defect, the reconstructed signal is close to the initially acquired signal. A relatively larger deviation indicates that the behavior of the mechanical system deviates from the modeled behavior in the absence of a defect, thus indicating the presence of a defect.

[0068] According to one embodiment of the process, the difference between the sequence of spectral components of the acquired vibration signal and the reconstructed sequence of spectral components of the vibration signal is determined for all spectral components of the vibration signal.

[0069] In this embodiment, the spectral components of the signal are reconstructed for all spectral frequencies, and the difference between the original signal and the reconstructed signal is also determined by taking into account all spectral frequencies, i.e. all the frequency information of the signal.

[0070] According to one embodiment of the process, the difference between the sequence of spectral components of the acquired vibration signal and the reconstructed sequence of spectral components of the vibration signal is determined for a subset of spectral components. The subset of spectral components is included in the sequence of spectral components of the acquired vibration signal.

[0071] In other words, in this case the spectral components of the signal are reconstructed for all spectral frequencies, but the difference between the original signal and the signal The reconstructed value is determined by considering only certain selected frequency ranges. The deviation can thus be quantified only within the frequency ranges most conducive to highlighting a structural defect.

[0072] According to one example of the process implementation, the determined gap is the Euclidean distance between the sequence of spectral components of the acquired vibration signal and the reconstructed sequence of spectral components of the vibration signal.

[0073] According to one aspect of the disclosure, the process includes one step: - (iv4) Determine a detection threshold, the detection threshold being equal to the maximum value of the deviation determined for a subset of time windows acquired under reference conditions in which the running gear element is free from structural defects.

[0074] Learning performed under reference conditions allows the defect threshold to be defined. The process can then be automatically adapted to any mechanical configuration. Furthermore, the defect threshold can be periodically updated as new data is collected under reference conditions. This improves the robustness of the process.

[0075] According to one example of implementation, the process includes one step: - (iv5-a) if the determined deviation is less than or equal to the detection threshold, the structural defect indicator takes an initial value corresponding to an absence of structural defect, - (iv5-b) if the determined deviation is greater than the detection threshold, the structural defect indicator takes a second value corresponding to the presence of a structural defect.

[0076] A significant difference between the acquired vibration signal and the reconstructed vibration signal tends to indicate the presence of a structural defect in the mechanical system. A small difference indicates that the system's response is consistent with its behavior in the absence of a defect.

[0077] The structural defect indicator can take discrete values, a first value corresponding to an absence of structural defect and a second value corresponding to the presence of a structural defect.

[0078] According to one embodiment, the process comprises one step: - (v) emit an alert signal in response based on the structural fault indicator.

[0079] The warning signal emitted allows for a thorough analysis of the vehicle's component and helps anticipate a failure.

[0080] The process can therefore include the following steps: - (v1) determine an average value for the structural defect indicator, - (v2) issue an alert signal if a difference between the determined average value and a reference value is greater than a predetermined threshold.

[0081] A single change in the value of the structural fault indicator does not trigger an alert signal. Repeated changes in value do trigger an alert signal.

[0082] The warning signal may be a fault code stored in an electronic control unit.

[0083] In one variant or as a complement, the warning signal may be the lighting of a warning light.

[0084] According to another variant or also in a complementary way, the warning signal can be a message sent to a vehicle manager.

[0085] The disclosure also relates to a structural fault detection device in a component of a land vehicle's running gear, comprising: - at least one accelerometer, - a device for acquiring the signal from the accelerometer, - an electronic control unit configured to implement the process described above.

[0086] According to one embodiment of the detection device, the electronic control unit is external to the vehicle, and the detection device includes a communication means configured to transmit the acquisitions or spectral components of the accelerometer signal to the electronic control unit.

[0087] The detection system includes an onboard computing unit configured to determine the complete sequence of spectral components of the vibration signal. An onboard computing unit is defined as a computing unit located on the vehicle. The signal transmitted to the external control unit can be either a frequency signal or a raw time signal. Transmitting a frequency signal limits the amount of data exchanged. Alternatively, both the raw time signal and the frequency signal can be transmitted. Brief description of the drawings

[0088] Other features, details, and advantages will become apparent upon reading the detailed description below and analyzing the attached drawings, on which:

[0089] Figure 1 is a schematic representation of a vehicle incorporating a device for detecting structural defects in a component of a running gear.

[0090] Figure 2 is a detail of the running gear of the vehicle in Figure 1.

[0091] Figure 3 is a schematic time curve illustrating the vibration signal acquisition stage,

[0092] Figure 4 is a curve illustrating the temporal evolution of the acquired vibration signal.

[0093] Figure 5 is a view illustrating a particular step in the process,

[0094] Figure 6 is a diagram illustrating some steps of the process,

[0095] Figure 7 is a view illustrating another specific step in the process,

[0096] Figure 8 is a view illustrating another specific step in the process,

[0097] Figure 9 is a block diagram illustrating different stages of the process according to the invention. Description of the implementation methods

[0098] To facilitate the reading of the figures, the different elements are not necessarily drawn to scale. In these figures, identical elements bear the same reference numbers. Some elements or parameters may be indexed, that is, designated, for example, as first element or second element, or first parameter and second parameter, etc. This indexing aims to differentiate similar, but not identical, elements or parameters. This indexing does not imply any priority of one element or parameter over another, and the designations can be interchanged. When it is specified that a device includes a given element, this does not exclude the presence of other elements in that device.

[0099] Figure 1 shows a land vehicle 100 comprising a running gear 20. The running gear 20 designates all the mechanical components involved in the connection between the chassis 2 of the vehicle and the ground on which the vehicle moves.

[0100] The vehicle 100 shown here is a truck. The running gear 20 comprises an axle 1 connected to the chassis 2 by an elastic element 10. The elastic element is a leaf spring, or more precisely, a leaf spring positioned near each end of the axle 1. The axle 1 has a wheel 4 at each end, connected to the axle by a bearing 3. The wheel 4 is fitted with a tire 5.

[0101] It is common for a component of the chassis (20) to develop a structural defect during the vehicle's lifespan. A structural defect is defined as a mechanical anomaly affecting the mechanical properties of the component in question. Such a structural defect can, over time, lead to chassis failure, which can impair vehicle performance and eventually render it inoperable. Even though chassis components are generally inspected during vehicle maintenance, it is advisable to be able to detect any structural defect in a chassis component during vehicle operation, in order to address the defect and anticipate potential failures.

[0102] A method is thus proposed for determining a structural defect indicator I of an element 10 of a running gear 20 of a land vehicle 100, the method comprising the following steps: - (i) acquire, during a time window W, a vibration signal V associated with element 10 of the running gear 20, - (ii) determine a sequence S of spectral components (A1, A2, ..., Ap) of the vibration signal V acquired during the time window W, - (iii) iterate steps (i) and (ii) for a set of time windows Wi, ..., W n so as to obtain a set of sequences Si, ..., S n of spectral components of the vibrational signal V, each sequence Sk of the set of sequences Si, ..., S n corresponding to a Wk time window, - (iv) determine a structural defect indicator I of element 10 of the running gear 20 from the set of sequences Si, ..., S n spectral components of the vibration signal V and from a detection model implementing a machine learning algorithm.

[0103] According to an example of implementation of the process, the structural defect is a crack in element 10 of the running gear 20.

[0104] Element 10 of the running gear 20 is here an elastic suspension element of vehicle 100.

[0105] More specifically, the elastic suspension element 10 is a leaf spring. The leaf spring 10 is located between an axle 1 of the vehicle and a chassis 2 of the vehicle.

[0106] According to an unillustrated example, the elastic suspension element 10 can be a helical spring. According to yet another unillustrated example, the elastic element of Suspension element 10 is a torsion bar. According to another, unillustrated example, the elastic element of suspension element 10 is an air spring.

[0107] In the illustrated example, chassis 2 comprises two parallel longitudinal members connected by transverse cross members, and a cab rests on chassis 2. According to an example not illustrated, chassis 2 of vehicle 100 may be a self-supporting body of vehicle 100.

[0108] Various types of structural defects can be detected by the proposed method. A structural defect may, in particular, be a defect in a bearing 3 of a wheel 4 connected to axle 1. A structural defect may also be a structural defect in a tire 5 of a wheel 4 connected to axle 1. A structural defect may also be an overload of axle 1. A structural defect may also be wear of the axle oscillation dampers.

[0109] When one of these structural defects is present, the vibration response of the running gear 20 is altered. It is therefore possible to detect the presence of a structural defect by analyzing the acquired vibration signal. The frequency range in which the signal is used varies depending on the nature of the defect and the intended application.

[0110] Vehicle 100 here is a truck. The vehicle could also be a van. In another application of the process, vehicle 100 is railway rolling stock. Vehicle 100 could thus be a train car or a train locomotive. In this case, the cab of the car, or the locomotive, rests on a series of bogies suspended by a set of elastic elements such as leaf springs.

[0111] The vibration signal V is an acceleration signal from an acceleration sensor 7.

[0112] The acceleration sensor 7 can be installed on various components or parts. The acceleration sensor 7 is rigidly fixed, for example by gluing, to the outer surface of a chassis-related component 20.

[0113] The acceleration sensor 7 can be positioned on the element 10 of the vehicle's running gear 20. In the example shown in Figure 2, an acceleration sensor 7 is positioned on the leaf spring 10. The acceleration sensor 7 is thus fixed to the element whose operating state is to be monitored.

[0114] According to the schematic example in Figure 1, the acceleration sensor 7 is located on axle 1. According to an example not shown, the acceleration sensor 7 is located on the chassis 2 of the vehicle 100.

[0115] In both configurations, the acceleration sensor 7 is located on a different component than the one being monitored. This simplifies the installation of the acceleration sensor 7, as it can then be installed in a more accessible location less exposed to potential wheel spray, such as the vehicle's chassis.

[0116] The method can use in parallel several signals acquired simultaneously by multiple acceleration sensors located on different components of the running gear or on the chassis. For example, it is possible to use an acceleration sensor located on a leaf spring and another acceleration sensor located on the chassis. It is also possible to use an acceleration sensor located on a leaf spring and an acceleration sensor located on the corresponding axle. A third acceleration sensor located on the chassis can also be used in conjunction with the spring.

[0117] The vibration signal V is said to be associated with element 10 of the running gear 20, because the vibration signal V accounts for the vibrations within element 10 of the running gear 20, even when fixed on a mechanical component other than element 10, such as the axle or chassis of the vehicle.

[0118] Acceleration sensor 7, for example, is a three-axis sensor. The sensor outputs a signal that depends on the acceleration along the three spatial directions. In Figures 1 and 2, the Z direction represents the vertical axis, the X direction represents the longitudinal axis of the vehicle, and the Y direction represents the transverse direction of the vehicle. A three-axis sensor is therefore sensitive to accelerations along each of the X, Y, and Z axes.

[0119] Alternatively, the acceleration sensor 7 can be a single-axis sensor. In this case, the sensor 7 outputs a signal related to acceleration in only one direction. Preferably, the sensor is installed so as to be preferentially sensitive to acceleration along the vertical Z-axis.

[0120] Acceleration sensor 7 outputs an analog signal. The bandwidth of acceleration sensor 7 is greater than 30 Hz. For example, the bandwidth of acceleration sensor 7 is between 25 kHz and 35 kHz.

[0121] An acquisition device 8 acquires the signal V from the accelerometer 7. During an acquisition time window W, the vibration signal V is sampled at a predetermined acquisition frequency Fa. The acquisition frequency Fa of the vibration signal V is preferably constant. For the purposes of this disclosure, acquisition frequency and sampling frequency are equivalent terms.

[0122] The terms "time window", "acquisition time window" and "signal acquisition time window" are equivalent for the purposes of this application.

[0123] The acquisition frequency Fa, also called the sampling frequency, of the vibration signal V from the accelerometer 7 is greater than 30 Hz. The acquisition frequency Fa can be between 25 kHz and 35 kHz. This sampling frequency is, for example, equal to 32 kHz.

[0124] Figure 3 schematically illustrates the acquisition time windows. A plurality of successive acquisition windows, designated by increasing indices from Wk to Wk+6, are shown. The arrows marked with the symbol g schematically illustrate the sampling times. The sampling times are separated by a duration Ta equal to the inverse of the sampling frequency Fa. To simplify the figure, the sampling times have been represented only for the Wk time acquisition window of index k, and have not been represented for the other windows. Each acquisition time window generates a plurality of measurement samples, each measurement sample corresponding to a value of the vibration signal V sampled at a given instant. The number of samples contained within an acquisition window depends on the duration of that acquisition window and the sampling frequency used. The measurement samples contained within an acquisition time window can be referred to as the "acquisition window content."

[0125] In the example of Figure 3, the time windows for acquiring the vibration signal V are contiguous. In other words, the beginning of a time window for acquiring Wj corresponds to the end of the previous acquisition window Wj-i.

[0126] According to an unillustrated example, the time windows for acquiring the vibrational signal V can be disjoint, that is, an acquisition time window Wj starts later than the end of the previous acquisition window Wj.-i.

[0127] Figure 4 illustrates the shape of the vibration signal V over a total acquisition time of 30 minutes. The peaks between times t1 and t2, for example, correspond to operating phases in which the accelerations are highest. Numerous other peaks are visible, such as around the time designated by ta.

[0128] The proposed method uses the characteristics of the acceleration signal in the frequency domain, and thus determines the spectral components of the acquired vibration signal V. The set of spectral components (A1, A2, ..., Ap) of the signal acquired during A given acquisition window is designated by the term 'sequence' S. For a time acquisition window of index k, designated by Wk, the corresponding sequence is designated by Sk, that is to say with the same index k.

[0129] The spectral components (A1, A2, ..., Ap) of the sequence S of spectral components are for example determined by fast Fourier transform of the samples of the vibration signal V acquired during the time window W.

[0130] Once the vibration signal acquisitions have been completed, the remaining steps of the process can be carried out in real time, or with delayed processing.

[0131] Each spectral component (A1, A2, ..., Ap) of a sequence Sk of spectral components of the vibration signal V is thus a value calculated over the entire corresponding acquisition time window Wk. The components (A1, A2, ..., Ap) of a sequence Sk of spectral components are, for example, calculated in decibels.

[0132] Each component Ai of a sequence Sk of spectral components corresponds to a frequency Fi. The component designated by Ai is the i ème spectral component of the Sk sequence. The frequency range Fi extends, for example, from 0.5 Hz to 16 kHz.

[0133] The set of sequences Si, ..., S n The spectral components are represented as a two-dimensional matrix. The columns can correspond to the different spectral components (A1, A2, ..., Ap), arranged in ascending order. The matrix rows then correspond to the different acquisition time windows Wi, ..., W narranged in chronological order of acquisition. It is of course possible to arrange the different spectral frequencies in rows, and the different time samples in columns, that is to say to transpose the matrix.

[0134] The duration of a time window W for acquiring the vibration signal V is configurable. The duration of a time window W for acquiring the vibration signal V can be between 2 seconds and 200 seconds. For example, the duration of a time window W for acquiring the vibration signal V is equal to 100 seconds.

[0135] According to one embodiment of the process, each sequence Sk of the set of sequences Si, ..., S n spectral components of the vibration signal V is obtained by averaging a plurality of sequences of spectral components corresponding to a plurality of elementary acquisition windows of the vibration signal V.

[0136] The plurality of elementary acquisition windows can include, for example, between 5 acquisition windows and 100 acquisition windows.

[0137] In other words, rather than using a single acquisition window W of a given duration, it is also possible to use several acquisition windows of shorter duration and average the results obtained over each of these windows. For example, instead of acquiring the vibration signal V during a 100-second time window, it is possible to acquire 50 successive windows, each lasting 2 seconds, and average the results obtained over these 50 windows, also representing 100 seconds of acquisition in total. This averaging step is optional, and the subsequent data processing is identical.

[0138] According to one embodiment, the process comprises one step: - (iii1) determine the forward speed C of vehicle 100, - (iii2) eliminate from the set of time windows Wi, ..., W n acquired the time windows corresponding to a vehicle forward speed C 100 lower than a predetermined minimum threshold Cmin.

[0139] Eliminating acquisition windows acquired when the vehicle speed is below the minimum threshold Cmin improves the signal-to-noise ratio. These acquisition windows correspond to vehicle operating phases during which the vibration signal V is weak and therefore contains little relevant information. In the example shown in Figure 4, the signal is very weak during the operating periods between times t and t5, as well as between times t and t7. These periods correspond to phases during which the vehicle is moving at very low speed or is stationary. It is therefore possible to eliminate the data points corresponding to these time periods from the acquired results. The samples of the vibration signal V acquired during time windows corresponding to a vehicle advancement speed C 100 lower than the predetermined minimum threshold Cmin are thus eliminated. In other words, when an acquisition window corresponds to a vehicle speed that is too low, the contents of that acquisition window—that is, the measurement samples acquired during that time window—are not taken into account for the rest of the proposed process. Only the samples from acquisition windows corresponding to a sufficient vehicle speed are retained.

[0140] For a road vehicle, the threshold Cmin is, for example, equal to 10 km / h. When the forward speed C of the vehicle 100 is less than the predetermined threshold Cmin for the entire duration of an acquisition window Wj of the vibration signal V, this window The data acquired during the acquisition window Wj is eliminated from the set of acquired data. In other words, the data acquired during this acquisition window Wi is not used by the subsequent steps of the process.

[0141] For example, the process may include one step: - (iii3) eliminate from the set of time windows Wi, ..., W n acquired the time windows corresponding to a zero speed of the vehicle 100. In other words, only the time windows corresponding to a moving vehicle are retained.

[0142] Given the frequency and duration of acquisition, the matrix formed by the set of sequences Si, ..., S nThe spectral components (A1, A2, ..., Ap) of the vibration signal V present a large number of rows and columns. This matrix can include several thousand spectral components and several hundred time windows. However, most of the variance in the calculated data can be retained by using significantly fewer independent spectral frequencies. The calculations can thus be significantly simplified while maintaining sufficient accuracy.

[0143] Alternatively, or in addition, the process includes one step: - (iii2') eliminate from the set of time windows (Wi, ..., W n ) acquired the time windows corresponding to a vehicle speed C 100 greater than a predetermined maximum threshold Cmax. This step can help eliminate acquisition windows where signal saturation is likely to occur.

[0144] The process may include a data normalization step. This means that the signal scale is reframed to have a mean of zero and a standard deviation of 1. The reframing transformation here is an affine transformation. To do this, the mean and standard deviation of the spectral components (A1, A2, ..., Ap) are determined for each sequence Sk of the set of sequences Si, ..., S n . Then, each sample is normalized by subtracting the previously determined mean and dividing the result of this subtraction by the previously determined standard deviation.

[0145] The normalized value Xj norm of a sample Xj of a variable X having a mean Moy and a standard deviation Ec is thus: [Math. 1]

[0147] The data used for the subsequent steps of the process are therefore the normalized values ​​of the set of sequences (Si, S n) of spectral components (A1, A2, Ap). In the example shown in Figure 5, dimensionality reduction to 2 dimensions was performed on a normalized dataset. The two principal components thus have values ​​distributed on either side of zero. The normalization step is optional.

[0148] Step (iv) of determining a structural defect indicator I of the proposed process thus includes a step (iv1) of dimensionality reduction of the set of sequences Si, ..., S n of spectral components (A1, A2, ..., Ap) of the vibration signal V, so as to obtain a set of reduced sequences SRi, ..., SR n of combinations of spectral components of the vibration signal V. Each reduced sequence SRk has a number q of spectral components less than the number p of spectral components determined in step (ii). All reduced sequences have the same number q of spectral frequencies.

[0149] The dimensionality reduction step decreases the complexity of the calculations performed. The number of columns in the matrix forming the set of spectral component sequences is thus reduced compared to the initial data. The number of rows in the matrix remains unchanged, as all acquisition windows are preserved.

[0150] According to one embodiment of the process, the step (iv1) of dimensionality reduction of the set of sequences Si, ..., S n The spectral components of the vibration signal V are analyzed using an unsupervised dimensionality reduction algorithm. An unsupervised algorithm allows for simple adaptation to a wide variety of technical definitions.

[0151] The algorithm implemented for dimensionality reduction can be based on an autoencoder network, commonly referred to by the English acronym "DAE" for "Deep Auto Encoder". The dimensionality reduction step (iv1) of the set of spectral component sequences of the vibration signal is thus performed by encoding a sequence of spectral components of the vibration signal, using an encoder of a pre-trained autoencoder network, in a reduced-dimensional latent space.

[0152] A principal component analysis algorithm, commonly referred to by the English acronym "PCA" for "Principal Component Analysis", is another example of an algorithm that can be implemented to perform the dimensionality reduction step.

[0153] The unsupervised dimensionality reduction algorithm generates a dimensionality reduction model M.

[0154] The set of reduced sequences SRi, SR nThe spectral components are thus determined by a model M obtained by machine learning.

[0155] The process can therefore automatically adapt to the mechanical configuration on which the process is used.

[0156] According to one example of implementation of the process, a reduced SRk sequence of spectral components comprises, for example, seven elements. In this case, a reduced SRk sequence of spectral components comprises seven frequency elements. The number q is 7. According to another example of implementing the process, a reduced SRk sequence of spectral components comprises, for example, three elements. The number q is thus 3. According to yet another example of implementing the process, a reduced SRk sequence of spectral components comprises, for example, two elements. The number q is then 2.

[0157] In other words, the matrix forming the set of spectral component sequences then comprises only 7 columns, 3 columns, and 2 columns, respectively. The associated calculations can thus be reduced by a large factor compared to the initially acquired data. The matrix includes as many rows as there are time windows retained after the elimination steps for windows acquired under unfavorable conditions, for example, for vehicle speeds that were too low or too high, as seen previously.

[0158] The dimensionality reduction model M of the set of sequences (Si, ..., S n ) of spectral components of the vibration signal V is obtained by automatic learning of the vibration signal V from a subset of time windows acquired under reference conditions in which the element 10 of the running gear 20 is free from structural defects. For example, the auto-encoding network is trained under conditions in which a structural fault is not present.

[0159] It is understood that the samples of the vibration signal acquired during time windows of acquisition corresponding to reference conditions are used to perform the machine learning enabling the M model of dimension reduction to be obtained. In other words, the content of the time windows acquired under the reference conditions serves as the basis for training the dimensionality reduction model M. The time windows corresponding to the reference conditions form a subset of the complete set (Wi, W n ) acquisition windows. Reference conditions are conditions in which no structural defects are present.

[0160] Detailed checks of the mechanical condition of the various components of the running gear 20 are preferably carried out by one or more maintenance operators before the baseline data acquisition phase. When a detailed check is not desired because it is too resource-intensive, the vehicle's new condition can be considered representative of the baseline conditions, free of structural defects and defining the vehicle's nominal operating condition.

[0161] Figure 5 illustrates results obtained after dimensionality reduction to 2 dimensions. The results can thus be plotted on a plane, with each axis representing a linear combination of spectral frequencies. Each point in the point cloud corresponds to a time acquisition window. The horizontal axis C1 corresponds to a first combination of spectral frequencies, and the vertical axis C2 corresponds to a second combination of spectral frequencies. As previously stated, each acquisition window may come from an average of a plurality of elementary acquisition windows, and some elementary windows may have been eliminated because they correspond to vehicle speeds that are too low. For convenience, examples where dimension reduction is done down to three dimensions or seven dimensions have not been shown.

[0162] The next step in the process is to reconstruct the initially acquired vibration signal, or more precisely the spectral components of the acquired vibration signal, from the reduced sequence of determined spectral components.

[0163] To achieve this, the process includes one step: - (iv2) determine, for each time window Wk, a reconstructed sequence S'k of spectral components A'1 , A'2, ..., A'p of the vibrational signal V acquired during the time window Wk from the reduced sequence SRk of combinations of spectral components and from a reconstruction model P obtained from the dimension reduction model M.

[0164] The reconstruction step of the acquired vibration signal is the inverse of the dimensionality reduction step. Figure 6 schematically illustrates the organization of the different steps of the process, with a series of data processing steps aimed at achieving a reduction of the dimensions of the data from the acquired signal and a series of inverse processing, allowing comparison of the acquired signal and a reconstructed signal.

[0165] The reconstruction model P of the spectral components A'1, A'2, A'p of the acquired vibration signal V is thus based on an automatic learning of the vibration signal V from a subset of time windows acquired under reference conditions in which the element 10 of the running gear 20 is free of structural defects.

[0166] The reconstruction model P of the acquired vibration signal V is obtained by inversion of the dimensionality reduction model M.

[0167] The reconstruction model can thus be easily determined from the model already established for the dimensionality reduction of the acquired data.

[0168] When the algorithm for reducing the dimensionality of the original matrix is ​​of the principal component analysis type, the dimensionality reduction step amounts to projecting the multidimensional space formed by the original data onto principal directions of a reduced-dimensional space. To do this, the original data matrix Y is decomposed into a product of three matrices of the form B x C x D T Since matrices B and D are orthogonal matrices, they are therefore invertible. T denotes the transpose matrix of D. The matrix D contains the eigenvectors of the covariance matrix of Y. Dimensionality reduction is achieved by multiplying the original matrix Y by a projection matrix D' containing a selection of columns from matrix D corresponding to the largest eigenvalues ​​of Y. To obtain matrix D', the columns of matrix D are ordered in descending order, and only the columns containing the highest values ​​are retained. The number of retained columns corresponds to the new dimension, reduced compared to the original dimension. The greater the number of retained columns, the greater the proportion of the signal's variance that is preserved. Multiplying the matrix D' by its inverse allows us to perform the inverse transformation of the dimensionality reduction, that is, the reconstruction of the spectral components of the original signal from the reduced spectral components. The dimensionality reduction step (iv1) and the spectral component reconstruction step (iv2) are performed using matrix calculus, i.e., linear algebra.

[0169] When the dimensionality reduction algorithm is based on an autoencoding network, the reconstruction step of the spectral components A1, A'2, ..., A'p of the acquired vibration signal V is ensured by the passage from the latent space to the output layer of The autoencoder. The autoencoder network ensures both the step (Iv1) of dimensionality reduction of the spectral components and the step (iv2) of reconstruction of the spectral components.

[0170] Figure 7 illustrates two examples of comparison between a sequence of spectral components of the acquired signal and a reconstructed sequence S'k of spectral components A'1, A'2, A'p. Part A corresponds to a first time acquisition window starting at time a, and part B corresponds to another time acquisition window starting at time b. a and Sb denote the spectral components of the acquired signal V, shown in black. S' a S'b and S'b denote the spectral components of the reconstructed signal, shown in lighter lines. For each of the two time windows, the reconstructed signal is similar to the acquired signal, but smoother, i.e., less noisy. This is a consequence of the dimensionality reduction step, which preserved most of the variance in the data.

[0171] The next step in the process involves comparing the reconstructed vibration signal with the initially acquired vibration signal. The spectral components of the acquired signal and the spectral components reconstructed after dimensionality reduction are thus compared.

[0172] To achieve this, the process includes one step: - (iv3) Determine, for each time window Wk, a gap E between the sequence Sk of spectral components (A1 , A2, ..., Ap) of the acquired vibration signal V and the reconstructed sequence S'k of spectral components (A'1 , A'2, ..., A'p) of the vibration signal V, and the structural defect indicator I is based on the gap E determined.

[0173] In the absence of a structural defect, the reconstructed spectral components of the signal are close to the spectral components of the initially acquired signal. A relatively larger discrepancy indicates that the behavior of the mechanical system deviates from the modeled behavior in the absence of a defect, which tends to indicate the presence of a defect.

[0174] According to one embodiment of the process, the difference E between the sequence Sk of spectral components (A1, A2, ..., Ap) of the acquired vibration signal V and the reconstructed sequence S'k of spectral components (A'1, A'2, ..., A'p) of the vibration signal V is determined for the set of spectral components (A1, A2, ..., Ap) of the vibration signal V.

[0175] In other words, the spectral components of the signal are reconstructed for the entire spectral frequency range, and the difference between the original signal and the reconstructed signal is calculated. is also determined by taking into account the total spectral frequencies, that is to say all the frequency information of the signal.

[0176] According to a variant of the process, the difference E between the sequence Sk of spectral components (A1, A2, ..., Ap) of the acquired vibration signal V and the reconstructed sequence S'k of spectral components (A'1, A'2, ..., Ap) of the vibration signal V is determined for a subset of spectral components (Ai, ..., Aj). The subset of spectral components (Ai, ..., Aj) is a subset of the sequence (S) of spectral components (A1 , A2, ..., Ap) of the acquired vibration signal V, comprising strictly fewer elements than the complete set of the sequence S of spectral components.

[0177] In other words, the spectral components of the signal are also reconstructed for the entire spectral range, but the difference between the original signal and the reconstructed signal is determined by considering only certain selected frequency ranges. The difference between the spectral components of the original signal and the spectral components of the reconstructed signal can thus be quantified only within the frequency ranges most conducive to detecting a structural defect. For example, low-frequency ranges are generally better suited for identifying a defect such as overload or reduced damping efficiency. High-frequency ranges are better suited for identifying a defect such as a crack in a leaf spring.

[0178] The difference between the initially acquired signal and the reconstructed signal can be calculated using different mathematical methods.

[0179] For example, the determined gap E is the Euclidean distance between the sequence Sk of spectral components (A1, A2, ..., Ap) of the acquired vibration signal V and the reconstructed sequence S'k of spectral components (A'1, A'2, ..., A'p) of the vibration signal V.

[0180] By definition, the Euclidean distance between a vector X with components Xi, ..., X q and a vector Y with components (Yi, ..., Y q ) is equal to the square root of the sum, over the set of vector components, of the squares of the differences between the components Xk and Yk of the two vectors X and Y.

[0181] Thus we have: [Math. 2]

[0183] In the next step, the detection threshold Th is determined, this threshold allowing us to differentiate between a state free from structural defects and a state where a structural defect is present.

[0184] The process thus comprises one step: - (iv4) Determine a detection threshold Th, the detection threshold Th being equal to the maximum value of the deviation determined E for a subset of time windows acquired under reference conditions in which element 10 of the rolling stock 20 is free from structural defects.

[0185] As before, it is understood that the samples of the vibration signal V from time windows of acquisition acquired under the reference conditions are used to determine the detection threshold Th. In other words, the content of the time windows acquired under the reference conditions serves as the basis for determining the detection threshold Th. The detection threshold Th is thus determined on the basis of measurement samples from a rolling stock element with no structural defects.

[0186] The training performed under reference conditions allows the detection threshold Th to be defined, thus establishing the presence of a structural defect. The process can therefore be automatically adapted to any mechanical configuration. Furthermore, the threshold used to define a defect can be periodically updated as data acquired under reference conditions is collected. This improves the robustness of the process.

[0187] The presence of a defect is determined by comparison between the previously calculated deviation E and the previously determined defect threshold Th.

[0188] The process thus comprises one step: - (iv5-a) if the determined deviation E is less than or equal to the detection threshold Th, the structural defect indicator I takes a first value I0 corresponding to an absence of structural defect, - (iv5-b) if the determined deviation E is greater than the detection threshold Th, the structural defect indicator I takes a second value 11 corresponding to the presence of a structural defect.

[0189] A significant difference between the acquired vibration signal V and the reconstructed vibration signal V' indicates the presence of a structural defect in the mechanical system. A small difference indicates that the system's response conforms to the modeled behavior in the absence of a defect.

[0190] The structural defect indicator I can take discrete values, with a first value I0 corresponding to the absence of a structural defect and a second value I1 corresponding to the presence of a structural defect. More precisely, the structural defect indicator I takes binary values ​​here.

[0191] Figure 8 illustrates the behavior of the deviation variable E for a first vehicle without a defect and a second vehicle with a defect. More precisely, the vertical axis represents the deviation E divided by the detection threshold Th. Each point in the figure corresponds to the processing performed for a time acquisition window. The horizontal dashed line therefore corresponds to the boundary between the area corresponding to the absence of a defect, below this line, and the area corresponding to a present defect, above the line. The first 75 measurement points correspond to operation under reference conditions without a defect. It can be noted that point P20 is the point used to determine the defect threshold, since this point corresponds to the maximum value in the absence of a defect. The value E / Th is, by definition, 1 for this measurement point.In the area below the horizontal dotted line, the structural defect indicator is 0. Measurement points from 75 onwards correspond to measurements taken on a vehicle where a leaf spring had a structural defect, in the form of a crack. For all these measurement points, the E / Th ratio is greater than 1, meaning that the deviation E is greater than the detection threshold Th. The structural defect indicator therefore takes the value 1, indicating the presence of a structural defect. The area above the horizontal dotted line corresponds to the area in which a structural defect is detected.

[0192] The primary goal of structural defect detection is to enable the vehicle operator or manager to take corrective action. Therefore, it is preferable to inform the operator or manager when the proposed procedure detects the presence of a defect.

[0193] The process thus comprises one step: - (v) emit an alert signal in response based on the structural fault indicator I.

[0194] The warning signal emitted allows for a thorough analysis of the vehicle's component and helps anticipate a failure.

[0195] An alert can be issued as soon as an event corresponding to a structural defect is detected. It is also possible to issue an alert only if Several events corresponding to a fault are detected, in order to confirm the presence of the fault.

[0196] The process can therefore include the following steps: - (v1) determine an average value Imoy of the structural defect indicator I, - (v2) issue an alert signal if a difference between the determined average value Imoy and a reference value is greater than a predetermined threshold Th2.

[0197] A single change in the value of the structural fault indicator does not trigger an alarm. Repeated changes in value trigger an alarm. No alarm is triggered when the determined average value Imoy is less than or equal to the predetermined threshold Th2. The reference value is the value corresponding to a constant absence of faults. It can be arbitrarily chosen that a fault-free state corresponds to the value 0 of indicator I. In this case, a faulty state corresponds to the value 1 of indicator I. A different convention can be chosen; that is, it can also be arbitrarily chosen to assign the value 1 to the indicator for operation without structural faults, and the value 0 for operation with structural faults.

[0198] The warning signal can be a fault code stored in an electronic control unit. Alternatively, or in addition, the warning signal can be an illuminated warning light. This light might, for example, be visible on the vehicle's dashboard. Another alternative, or also in addition, is the warning signal that can be a message sent to a vehicle management system.

[0199] The disclosure also relates to a structural fault detection device 50 in a component 10 of a running gear 20 of a land vehicle 100, comprising: - at least one accelerometer 7, - a signal acquisition device 8 for the accelerometer 7 signal, - an electronic control unit 15 configured to implement the process described above.

[0200] One or more 50 detection devices can be fitted to a given vehicle. The number of signal acquisition devices depends on the capabilities of the devices themselves, as well as the number of accelerometers used. The number of elements to be monitored in the case of a train is naturally greater than for a two-axle van.

[0201] According to one embodiment of the detection device 50, the electronic control unit 15 is external to the vehicle 100, and the detection device 50 includes a communication means 9 configured to transmit the signal acquisitions from the accelerometer 7 to the electronic control unit 15. Alternatively or in addition, the communication means 9 can be configured to transmit the spectral components of the signal from the accelerometer 7 to the electronic control unit 15.

[0202] The data acquired during the vibration signal measurement is sent to a remote cloud-based server. The vehicle only contains the electronic equipment necessary for data acquisition and transmission to the remote server(s). Data processing is performed by computer systems located away from the vehicle. This limits the instrumentation required on the vehicle. There is no limit to the number of vehicles on which the process can be implemented. The process can therefore be implemented on part or all of a vehicle fleet. Alerts can, for example, be sent to a fleet manager. The manager can then organize an inspection of each vehicle for which an alert signal has been issued, and, if necessary, a repair. Monitoring can continue until the repair is carried out.

Claims

Claims

1. Method for determining an indicator (I) of a structural defect of an element (10) of a running gear (20) of a land vehicle (100), the method comprising the steps: (i) acquiring during an acquisition time window (W) a vibration signal (V) associated with the element (10) of the running gear (20), (ii) determine a sequence (S) of spectral components (A1, A2, ..., Ap) of the vibration signal (V) acquired during the acquisition time window (W), (iii) iterate steps (i) and (ii) for a set of time windows (Wi, ..., W n ) acquisition in order to obtain a set of sequences (Si, ..., S n ) of spectral components of the vibration signal (V), each sequence (Sk) of the set of sequences (Si, ..., S n ) corresponding to a time window (Wk) of acquisition, (iv) determining an indicator (I) of structural defect of the element (10) of the running gear (20) from the set of sequences (Si, ..., S n ) of spectral components of the vibration signal (V) and from a detection model implementing a machine learning algorithm.

2. A method according to claim 1, wherein the structural defect is a crack in the element (10) of the running gear (20), wherein the element (10) of the running gear (20) is an elastic suspension element of the vehicle (100), wherein the elastic suspension element (10) is a leaf spring, arranged between an axle (1) of the vehicle and a chassis (2) of the vehicle.

3. Method according to claim 2, wherein the vibration signal (V) is an acceleration signal from an acceleration sensor (7) arranged on the element (10) of the running gear (20) of the vehicle, or arranged on the axle (1), or arranged on the chassis (2) of the vehicle (100).

4. Method according to one of the preceding claims, in which the vibration signal (V) is sampled during each acquisition time window (W) at a predetermined acquisition frequency (Fa), the method comprising a step: (iii1) determine the forward speed (C) of the vehicle (100), (iii2) eliminate from the set of samples of the vibration signal (V) time windows (Wi, ..., W n ) acquisition of the samples of the vibration signal (V) from the windows time corresponding to a forward speed (C) of the vehicle (100) lower than a predetermined minimum threshold (Cmin).

5. Method according to one of the preceding claims, in which step (iv) of determining an indicator (I) of structural defect comprises a step (iv1) of reducing the dimensions of the set of sequences (Si, S n ) of spectral components (A1, A2, Ap) of the vibration signal (V), so as to obtain a set of reduced sequences (SRi, SR n ) of combinations of spectral components of the vibration signal (V), each reduced sequence (SRk) comprising a number of spectral components less than the number p of spectral components determined in step (ii).

6. Method according to the preceding claim, in which the step (iv1) of reducing the dimensions of the set of sequences (Si, ..., S n) of spectral components of the vibration signal (V) implements an unsupervised dimension reduction algorithm, the unsupervised dimension reduction algorithm generating a dimension reduction model (M).

7. Method according to the preceding claim, in which the vibration signal (V) is sampled during each acquisition time window (W) at a predetermined acquisition frequency (Fa), and in which the dimension reduction model (M) of the set of sequences (Si, ..., S n ) of spectral components of the vibration signal (V) is obtained by automatic learning of the vibration signal (V) from the samples of the vibration signal (V) of a subset of acquisition time windows acquired under reference conditions in which the element (10) of the running gear (20) is free from structural defects.

8. Method according to the preceding claim, comprising a step: (iv2) determining, for each time window (Wk), a reconstructed sequence (S'k) of spectral components (A'1, A'2, ..., A'p) of the vibration signal (V) acquired during the time window (Wk) from the reduced sequence (SRk) of combinations of spectral components and from a reconstruction model (P) obtained from the dimension reduction model (M).

9. Method according to the preceding claim, in which the reconstruction model (P) of the spectral components of the acquired vibration signal (V) is obtained by inversion of the dimension reduction model (M).

10. A method according to claim 8 or 9, comprising a step: (iv3) Determine, for each time window (Wk), a deviation (E) between the sequence (Sk) of spectral components (A1, A2, Ap) of the acquired vibration signal (V) and the reconstructed sequence (S'k) of spectral components (A'1, A'2, A'p) of the vibration signal (V), in which method the structural defect indicator (I) is based on the determined deviation (E).

11. Method according to the preceding claim, in which the difference (E) between the sequence (Sk) of spectral components (A1, A2, ..., Ap) of the acquired vibration signal (V) and the reconstructed sequence (S'k) of spectral components (A'1, A'2, ..., A'p) of the vibration signal (V) is determined for a subset of spectral components (Ai, ..., Aj).

12. Method according to claim 10 or 11, in which the determined deviation (E) is the Euclidean distance between the sequence (Sk) of spectral components (A1, A2, ..., Ap) of the acquired vibration signal (V) and the reconstructed sequence (S'k) of spectral components (A'1, A'2, ..., A'p) of the vibration signal (V).

13. Method according to one of claims 10 to 12, in which the vibration signal (V) is sampled during each acquisition time window (W) at a predetermined acquisition frequency (Fa), the method comprising the steps: (iv4) Determine a detection threshold (Th), the detection threshold (Th) being equal to the maximum value of the determined deviation (E) for the samples of the vibration signal (V) of a subset of acquisition time windows acquired under reference conditions in which the element (10) of the running gear (20) is free from structural defects, (iv5-a) if the determined deviation (E) is less than or equal to the detection threshold (Th), the structural defect indicator (I) takes a first value (I0) corresponding to an absence of structural defect, (iv5-b) if the determined deviation (E) is greater than the detection threshold (Th), the structural defect indicator (I) takes a second value (11) corresponding to the presence of a structural defect.

14. Method according to the preceding claim, comprising the steps: (v1) determining an average value (Imoy) of the structural defect indicator (I), (v2) emit an alert signal if a difference between the determined average value (Imoy) and a reference value is greater than a predetermined threshold (Th2).

15. Device (50) for detecting a structural defect in an element (10) of a running gear (20) of a land vehicle (100), comprising: - at least one accelerometer (7), - a device (8) for acquiring the signal from the accelerometer (7), - an electronic control unit (15) configured to implement the method according to one of the preceding claims, in which the electronic control unit (15) is external to the vehicle (100), and in which the detection device (50) comprises a communication means (9) configured to transmit the acquisitions or the spectral components of the signal from the accelerometer (7) to the electronic control unit (15).

Images