A method for fault detection of railway installations

Abstract

This invention provides a method for fault detection in railway facilities. The method includes: acquiring measured wheel-rail noise signals during train operation; obtaining time-domain information of the wheel-rail noise based on the signals; dividing the acquired time-domain information into labeled and unlabeled wheel-rail noise datasets according to a preset ratio; constructing a railway facility fault detection model; training the model using the labeled and unlabeled wheel-rail noise datasets to obtain a trained model; inputting the wheel-rail noise signal to be detected into the trained model for detection; and obtaining the fault type and location corresponding to the detected wheel-rail noise signal. This invention enables the simultaneous detection of multiple railway facility faults using a single acoustic model, thereby improving detection efficiency, reducing detection costs, and ensuring the safe, stable, and efficient operation of railways.

Description

Technical Field

[0001] This invention relates to the field of railway facility fault detection technology, and in particular to a method for fault detection of railway facilities. Background Technology

[0002] Railways are the main artery of the national economy and a vital infrastructure vital to people's livelihoods. Ensuring the safe, stable, and efficient operation of railways is a fundamental requirement and primary prerequisite for building a strong transportation nation. With the increasing scale of my country's railway network, rising transport demands, deteriorating service conditions of railway facilities, and a more complex operating environment, malfunctions in railway facilities (rolling vehicles, tracks, roadbeds, bridges and tunnels, communication and signaling equipment, etc.) are inevitable. However, some malfunctions (rail breaks, vehicle derailments, roadbed collapses, etc.) can lead to incalculable and serious consequences. Furthermore, defects such as rail corrugation and out-of-round wheels, while not affecting operational safety, can reduce train stability and comfort. Currently, there are various methods for detecting and monitoring railway facility malfunctions. Typically, comprehensive inspection vehicles are used to inspect the railway condition on each line once a month (or longer). This method is expensive, and a large amount of data cannot be fully utilized. Furthermore, using static inspection equipment during railway maintenance windows is inefficient and contradicts the goal of building smart railways. Therefore, how to achieve rapid perception, intelligent diagnosis, and maintenance management of railway facility status is the primary issue facing railway operation departments.

[0003] Existing railway facility fault detection models are mostly single-type, meaning that the model architecture often differs significantly for different railway facility faults, while the data sources are basically the same. For example, wheel-rail noise signals are used to detect various types of defects such as out-of-round wheels, rail corrugation, and track slab delamination. Therefore, there is an urgent need to develop a large-scale model technology based on acoustic signals that can simultaneously detect and identify multiple railway facility faults. Summary of the Invention

[0004] Embodiments of the present invention provide a method for fault detection of railway facilities, so as to effectively detect and identify various faults in railway facilities.

[0005] To achieve the above objectives, the present invention adopts the following technical solution.

[0006] A method for fault detection of railway facilities includes:

[0007] Acquire measured wheel-rail noise signals during train operation, and obtain time-domain information of wheel-rail noise based on the wheel-rail noise signals;

[0008] The acquired wheel-rail noise time-domain information is divided into labeled wheel-rail noise datasets and unlabeled wheel-rail noise datasets according to a preset ratio;

[0009] A railway facility fault detection model is constructed, and the labeled wheel-rail noise dataset and the unlabeled wheel-rail noise dataset are used to train the railway facility fault detection model to obtain a trained railway facility fault detection model.

[0010] The wheel-rail noise signal to be detected is input into the trained railway facility fault detection model for detection, and the fault type and location corresponding to the wheel-rail noise signal to be detected are obtained.

[0011] Preferably, the step of acquiring the measured wheel-rail noise signal during train operation and acquiring the time-domain information of the wheel-rail noise based on the wheel-rail noise signal includes:

[0012] A sound pressure sensor is installed on the side of the train frame near the wheel, and the sound pressure sensor is used to collect the acoustic signal generated by the vibration radiation of the wheel and rail.

[0013] A pulse generator is installed on the axle of the train being measured. When the wheel starts to roll, the sound pressure sensor is triggered to sample the acoustic signal, and when the wheel stops rolling, the sound pressure sensor stops sampling the acoustic signal.

[0014] The above operation is repeated on different trains on a certain railway line to obtain a wheel-rail noise dataset for a set duration, and the time-domain information of the wheel-rail noise is obtained based on the wheel-rail noise dataset.

[0015] Preferably, the construction of the railway facility fault detection model involves training the railway facility fault detection model using the labeled wheel-rail noise dataset and the unlabeled wheel-rail noise dataset to obtain a trained railway facility fault detection model.

[0016] The labeled wheel-rail noise dataset is segmented into segmented wheel-rail noise time series according to the preset window length, and a segmented wheel-rail noise time series dataset is constructed.

[0017] Perform discrete wavelet transform on each segment of wheel-rail noise time series in the labeled wheel-rail noise dataset, calculate Lipschitz coefficients, and obtain the fault type label corresponding to each segment of wheel-rail noise time series.

[0018] An acoustic model based on self-supervised learning is established. The unlabeled wheel-rail noise dataset is input into the acoustic model for training. The acoustic representation of wheel-rail noise is learned, and a trained acoustic model is obtained.

[0019] By adjusting the trained acoustic model using the fault type labels corresponding to each segment of the wheel-rail noise time series, a trained railway facility fault detection model is obtained.

[0020] Preferably, the step of segmenting the labeled wheel-rail noise dataset into segmented wheel-rail noise time series according to a preset window length, and constructing a segmented wheel-rail noise time series dataset, includes:

[0021] Suppose the wheel-rail noise signals in the labeled wheel-rail noise dataset are x1, x2, ..., x3. n The segmentation window length L is determined based on the train's operating speed. Based on the segmentation window length L, the wheel-rail noise signals x1, x2, ..., x... are analyzed. n The labeled wheel-rail noise dataset is segmented into segmented wheel-rail noise time series.

[0022] Each segment of wheel-rail noise signal x i The i=1,2,…,n signal is divided into N equal-length wheel-rail noise signals y. j For j=1,2,…,N, the wheel-rail noise time series with a length less than L are directly discarded;

[0023] (1).

[0024] Preferably, the step of performing discrete wavelet transform on each segment of the wheel-rail noise time series in the labeled wheel-rail noise dataset, calculating Lipschitz coefficients, and obtaining the fault type label corresponding to each segment of the wheel-rail noise time series includes:

[0025] Discrete wavelet transform is performed on the segmented wheel-rail noise time series according to equations (2) and (3).

[0026] (2)

[0027] (3)

[0028] In the formula, h is the coefficient vector of the scaling filter, m = 2k + n, k is the translation amount of the mother wavelet, n is the integer index of the filter, and h1 is the wavelet filter.

[0029] A four-level multi-resolution analysis was performed on the segmented wheel-rail noise time series to obtain wavelet coefficient information at five scales. The maximum modulus of the wavelet coefficient at each scale was found, and the corresponding average value A was calculated to determine the detection threshold λ, as shown in equations (4) and (5).

[0030] (4)

[0031] (5)

[0032] In the formula, a i K represents the maximum modulus of wavelet coefficients at each scale, and K is a threshold control coefficient with a value ranging from 1.2 to 1.5.

[0033] Wavelet coefficients at each scale and Wavelet coefficients below the threshold λ are directly reset to 0, and other wavelet coefficients above the threshold λ are marked as suspected fault locations u.

[0034] Perform continuous wavelet transform on the segmented wheel-rail noise time series according to equation (6);

[0035] (6)

[0036] In the formula, For the mother wavelet, a is the scale parameter and b is the positioning parameter;

[0037] The Lipschitz coefficient corresponding to the suspected fault location u is calculated according to equation (6), as shown in equation (7);

[0038] (7)

[0039] In the formula, J corresponds to the maximum scale of wavelet decomposition;

[0040] The fault type and location are determined based on the suspected fault location and the corresponding Lipschitz coefficient. The fault types include wheel flattening, rail flattening, and rail corrugation.

[0041] Specifically, signals with a Lipschitz coefficient greater than 0 at suspected fault locations in scales 1-5 of the wheel-rail noise signal decomposition are identified as rail corrugation; signals with a Lipschitz coefficient greater than 1 at suspected fault locations in scales 1-3 are identified as rail flattening; and signals with a Lipschitz coefficient less than 1 at suspected fault locations in scales 1-2 are identified as wheel flattening. Based on the suspected fault locations in the segmented wheel-rail noise time series and the determined fault types and corresponding wheel-rail noise time series ranges, a new labeled wheel-rail noise dataset is constructed, in the format {wheel-rail noise time series, fault type}.

[0042] Preferably, the step of establishing an acoustic model based on self-supervised learning, which involves inputting the unlabeled wheel-rail noise dataset into the acoustic model for training, learning the acoustic representation of wheel-rail noise, and obtaining a trained acoustic model, includes:

[0043] An acoustic model is constructed, consisting of convolutional layers, Transformer layers, fully connected layers, and a pseudo-label generation layer. The acoustic model is based on self-supervised learning.

[0044] The unlabeled wheel-rail noise dataset is divided into a training set, a validation set, and a test set; wherein the ratio of the training set, validation set, and test set is 8:1:1.

[0045] The training set is input into the acoustic model, and the optimal parameters of the model are found by a greedy algorithm. The training set is divided according to a preset window length L to obtain segmented unlabeled wheel-rail noise time series. Each segment of the unlabeled wheel-rail noise time series is downsampled through a convolutional layer, and the output length is determined by equation (8).

[0046] (8)

[0047] In the formula, M is the input length, i.e. the sampling points contained in the window, K is the convolution kernel size, the stride is S, and P zeros are padded at both ends of the input.

[0048] The input raw, unlabeled wheel-rail noise time series is processed by a convolutional layer to produce H vectors v. i Let i = 1, 2, ..., H, and each vector have a length of O. Use a K-means classifier to classify the H vectors v. i The classification process generates a corresponding pseudo-label vector V' for each vector. H vectors are then randomly masked and fed into a Transformer layer to produce H output vectors v. i ’ , i=1,2,…,H;

[0049] v i ’ The output vector with the Mask operation is then input into the fully connected layer to obtain the target output vector V. The distance D between the target output vector V and the corresponding pseudo-label vector V' is calculated, as shown in equation (9).

[0050] (9)

[0051] Repeat the above operations, use the Adam optimizer to update the parameters, and the objective function is as shown in equation (10);

[0052] (10)

[0053] In the formula, w and b are the learnable parameters in the neural network model;

[0054] The validation set is input into the acoustic model to validate the acoustic model. The test set is input into the acoustic model to perform random vector matching prediction to evaluate the effect of the acoustic model in learning the acoustic representation of wheel-rail noise. After multiple rounds of iterative training, a trained acoustic model is obtained.

[0055] Preferably, the H vectors are randomly masked and then input into the Transformer layer to obtain H output vectors v. i ’ ,include:

[0056] H vectors are randomly masked and then fed into a Transformer layer to obtain H output vectors v. i ’ (i=1,2,…,H), each time the original unlabeled wheel-rail noise time series is input, after passing through the convolutional layer, H vectors are generated. 20% of the vectors are randomly selected and replaced with specific symbols. <mask>When the selected vector is permuted into <mask>At that time, there is a 60% probability that it will continue to remain unchanged. <mask>It remains unchanged, with a 20% probability of being replaced by another random vector, and a remaining 20% ​​probability that the vector is left unchanged.

[0057] The H processed vectors are input into a Transformer layer, which consists of 12 Transformer blocks, with a hidden layer size of 768 and a header of 12. The Transformer layer outputs H output vectors v. i ’ .

[0058] Preferably, the step of adjusting the trained acoustic model using the fault type labels corresponding to each segment of the wheel-rail noise time series to obtain a trained railway facility fault detection model specifically includes:

[0059] Different multilayer perceptron (MLP) structures were designed for different fault types to obtain fault detection models for railway facilities.

[0060] The new labeled wheel-rail noise dataset in the format {wheel-rail noise time series, fault type} is divided into a training set, a validation set, and a test set, with a ratio of 6:2:2.

[0061] The parameters of the trained acoustic model are used as the initial parameters of the railway facility fault detection model.

[0062] The training set is input into the railway facility fault detection model, and the optimal parameters of the railway facility fault detection model are found by Bayesian search algorithm;

[0063] The validation set is input into the railway facility fault detection model to validate the railway facility fault detection model;

[0064] The test set is input into the railway facility fault detection model for fault diagnosis, and the detection effect of the railway facility fault detection model is tested. After multiple rounds of iterative training, a trained railway facility fault detection model is obtained.

[0065] As can be seen from the technical solutions provided by the embodiments of the present invention described above, the present invention can realize the simultaneous detection of multiple railway facility faults using a single acoustic model, thereby improving detection efficiency, reducing detection costs, and ensuring the safe, stable, and efficient operation of railways.

[0066] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and will become apparent from the description or may be learned by practice of the invention. Attached Figure Description

[0067] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0068] Figure 1 This is a schematic diagram illustrating the types of railway facility failures provided in an embodiment of the present invention.

[0069] Figure 2 An acoustic model flowchart for railway facility health monitoring is provided as an embodiment of the present invention.

[0070] Figure 3 This is a schematic diagram of a railway facility fault detection model based on MLP (Multilayer Perceptron) provided in an embodiment of the present invention.

[0071] Figure 4 This is a schematic diagram of a train wheel-rail noise test provided in an embodiment of the present invention.

[0072] Figure 5 This is a schematic diagram of wheel-rail noise dataset segmentation provided in an embodiment of the present invention.

[0073] Figure 6 This is a flowchart of a wheel-rail noise dataset label extraction method provided in an embodiment of the present invention.

[0074] Figure 7 This is a schematic diagram of an acoustic model training process provided in an embodiment of the present invention.

[0075] Figure 8 This is a flowchart of a pseudo-label vector generation and training process provided in an embodiment of the present invention. Detailed Implementation

[0076] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0077] Those skilled in the art will understand that, unless specifically stated otherwise, the singular forms "a," "an," "described," and "the" used herein may also include the plural forms. It should be further understood that the term "comprising" as used in this specification means the presence of the stated features, integers, steps, operations, elements, and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof. It should be understood that when we say an element is "connected" or "coupled" to another element, it can be directly connected or coupled to the other element, or there may be intermediate elements. Furthermore, "connected" or "coupled" as used herein can include wireless connections or couplings. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0078] It will be understood by those skilled in the art that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. It should also be understood that terms such as those defined in general dictionaries should be understood to have the same meaning as in the context of the prior art, and should not be interpreted in an idealized or overly formal sense unless defined as herein.

[0079] To facilitate understanding of the embodiments of the present invention, the following will provide further explanation and description with reference to the accompanying drawings and several specific embodiments. These embodiments do not constitute a limitation on the embodiments of the present invention.

[0080] Example 1

[0081] Figure 1 This diagram illustrates a type of railway facility fault provided in an embodiment of the present invention. The method of the present invention utilizes collected measured wheel-rail noise data to train an acoustic model based on self-supervised learning, and then fine-tunes the model using labeled railway facility fault data, thereby enabling the detection of different types of faults.

[0082] This invention provides a method for fault detection of railway facilities, with the following processing flow: Figure 2 As shown, the processing steps include the following:

[0083] Step S10: Obtain the measured wheel-rail noise signal during train operation, and obtain the time-domain information of the wheel-rail noise based on the wheel-rail noise signal;

[0084] Step S20: Divide the acquired wheel-rail noise time-domain information into labeled wheel-rail noise datasets and unlabeled wheel-rail noise datasets according to a preset ratio;

[0085] Step S30: Divide the labeled wheel-rail noise dataset into segmented wheel-rail noise time series according to the preset window length, and construct a segmented wheel-rail noise time series dataset;

[0086] Step S40: Perform discrete wavelet transform on each time series in the labeled wheel-rail noise dataset, calculate the Lipschitz coefficients, and extract the fault type labels corresponding to the segmented wheel-rail noise time series;

[0087] Step S50: Establish an acoustic model based on self-supervised learning, input an unlabeled wheel-rail noise dataset for training, and learn the acoustic representation of wheel-rail noise.

[0088] Step S60: Fine-tune the trained acoustic model using the labeled wheel-rail noise dataset to obtain the railway facility fault detection model; Figure 3 This is a schematic diagram of a railway facility fault detection model based on MLP, provided as an embodiment of the present invention.

[0089] Step S70: Input the wheel-rail noise signal to be detected into the railway facility fault detection model for detection, thereby identifying the type and location of the fault.

[0090] The acquisition of the measured wheel-rail noise signal during train operation in step S10 above specifically includes:

[0091] A high-precision IEPE acoustic pressure sensor is installed on the side of the train frame near the wheel, and the acoustic pressure sensor is used to collect the acoustic signal generated by the vibration radiation of the wheel and rail.

[0092] Acoustic signals were acquired using a 24AD high-precision data acquisition system, with the sampling frequency set to 40000Hz.

[0093] A pulse generator is installed on the axle being tested. When the wheel starts to roll, the sound pressure sensor is triggered to sample the acoustic signal, and when the wheel stops rolling, the sound pressure sensor stops sampling the acoustic signal.

[0094] Repeat the above operation on different commercial trains on a certain railway line to obtain a large dataset of wheel-rail noise of no less than 100,000 hours.

[0095] In this example, step S30, which involves segmenting the labeled wheel-rail noise dataset into a segmented wheel-rail noise time series according to a preset window length, specifically includes:

[0096] Suppose the wheel-rail noise signals in the labeled wheel-rail noise dataset are x1, x2, ..., x3. n (x) i (The lengths are not equal). The segmentation window length L is determined based on the train speed. The wheel-rail noise signals x1, x2, ..., x are then analyzed based on the segmentation window length L. n Segment the signal, and obtain each segment of the wheel-rail noise signal x. i (i=1,2,…,n) can be divided into N equal-length wheel-rail noise signals y j (j=1,2,…,N), the wheel-rail noise time series with a length less than L are directly discarded.

[0097] (1)

[0098] In this example, step S40 involves performing discrete wavelet transform on each time series segment in the labeled wheel-rail noise dataset, calculating Lipschitz coefficients, and extracting fault type labels corresponding to the segmented wheel-rail noise time series.

[0099] First, perform discrete wavelet transform on the segmented wheel-rail noise time series according to equations (2) and (3).

[0100] (2)

[0101] (3)

[0102] In the formula, h is the coefficient vector of the scaling filter, m = 2k + n, k is the translation amount of the mother wavelet, n is the integer index of the filter, and h1 is the wavelet filter.

[0103] A four-level multi-resolution analysis was performed on the segmented wheel-rail noise time series to obtain wavelet coefficient information at five scales. The maximum modulus of the wavelet coefficient at each scale was found, and the corresponding average value A was calculated. The detection threshold λ was then determined, as shown in equations (4) and (5).

[0104] (4)

[0105] (5)

[0106] In the formula, a i K represents the maximum modulus of wavelet coefficients at each scale, and K is a threshold control coefficient with a value ranging from 1.2 to 1.5.

[0107] If the wavelet coefficients at each scale are below the threshold λ, they are directly reset to 0. Then, the positions where the wavelet coefficients are above the threshold λ are marked as suspected fault locations.

[0108] Perform continuous wavelet transform on the segmented wheel-rail noise time series according to equation (6);

[0109] (6)

[0110] In the formula, For the mother wavelet, a is the scale parameter and b is the positioning parameter;

[0111] The Lipschitz coefficient corresponding to the suspected fault location u is calculated according to equation (6), as shown in equation (7);

[0112] (7)

[0113] In the formula, J corresponds to the maximum scale of wavelet decomposition;

[0114] Determine the type and location of the fault (wheel flattening, rail flattening, rail corrugation, etc.) based on the suspected fault location and the corresponding Lipschitz coefficient.

[0115] Specifically, signals with a Lipschitz coefficient greater than 0 at suspected fault locations in scales 1-5 of the wheel-rail noise signal decomposition are identified as rail corrugation; signals with a Lipschitz coefficient greater than 1 at suspected fault locations in scales 1-3 are identified as rail flattening; and signals with a Lipschitz coefficient less than 1 at suspected fault locations in scales 1-2 are identified as wheel flattening.

[0116] Based on the suspected fault locations and the determined fault types and corresponding wheel-rail noise time series ranges in the segmented wheel-rail noise time series, a new labeled wheel-rail noise dataset is constructed, with the format {wheel-rail noise time series, fault type}.

[0117] In this example, the acoustic model in step S50 consists of convolutional layers, Transformer layers, fully connected layers, and a pseudo-label generation layer. Based on self-supervised learning, it does not require a labeled dataset.

[0118] The unlabeled wheel-rail noise time series is input into the acoustic model. The unlabeled wheel-rail noise time series is segmented according to the preset window length L. Each segment of the time series is downsampled through a convolutional layer. The output length is then determined by equation (8).

[0119] (8)

[0120] In the formula, M is the input length, i.e. the sampling points contained in the window, K is the convolution kernel size, the stride is S, and P zeros are padded at both ends of the input.

[0121] The wheel-rail noise time series is passed through a convolutional layer in the acoustic model to produce H vectors v. i (i=1,2,…,H), each vector has a length of O. H vectors are randomly masked and then fed into a Transformer layer to obtain H output vectors v. i ’ (i=1,2,…,H);

[0122] v i ’ The output vector with the Mask operation is input into the fully connected layer to obtain the target output vector V. Then, the distance D between the target output vector V and the corresponding pseudo-label vector V' is calculated, as shown in equation (9).

[0123] (9)

[0124] Repeat the above operations, use the Adam optimizer to update the parameters, and the objective function is as shown in equation (10);

[0125] (10)

[0126] In the formula, w and b are the learnable parameters in the neural network model.

[0127] In this example, H vectors are randomly masked and then input into the Transformer layer to obtain H output vectors v. i ’ (i=1,2,…,H), specifically, each time the original wheel-track noise time series is input, after passing through the convolutional layer, H vectors are generated, and 20% of these vectors are randomly selected and replaced with specific symbols. <mask>When the selected vector is permuted into <mask>At that time, there is a 60% probability that it will continue to remain unchanged. <mask>It remains unchanged, with a 20% probability of being replaced by another random vector, and a remaining 20% ​​probability that the vector is left unchanged.

[0128] The Transformer layer consists of 12 Transformer blocks, the hidden layer size is 768, and the header number is 12.

[0129] In this example, the distance D between the target output vector V and the corresponding pseudo-label vector V' is calculated, where the pseudo-label vector V' is determined by the pseudo-label generation layer.

[0130] First, all wheel-rail noise time series are segmented according to a preset window length. Then, each time series is subjected to a 4-level discrete wavelet transform to obtain wavelet coefficients of 5 scales. These coefficients are combined into a feature vector, and each time series can be represented by a feature vector.

[0131] By using the K-means algorithm to perform cluster analysis on the dataset composed of all feature vectors, a classifier with M categories can be obtained;

[0132] The H vectors v generated after the convolutional layer in the acoustic model are... i Before performing the Mask operation (i=1,2,…,H), the K-means classifier is first used to classify the vectors. Each vector will then have a corresponding class vector, which is the pseudo-label vector.

[0133] In this example, the input unlabeled wheel-rail noise dataset is used for training to learn the acoustic representation of wheel-rail noise, specifically including:

[0134] The unlabeled wheel-rail noise dataset is divided into a training set, a validation set, and a test set; wherein the ratio of the training set, validation set, and test set is 8:1:1.

[0135] The training set is input into the acoustic model, and the optimal parameters of the model are found through a greedy algorithm.

[0136] The validation set is input into the acoustic model to validate the model;

[0137] The test set is input into the acoustic model for random vector matching prediction to evaluate the effectiveness of wheel-rail noise representation learning.

[0138] In this example, step S60, which involves fine-tuning the trained acoustic model using a labeled wheel-rail noise dataset to obtain a railway facility fault detection model, specifically includes:

[0139] Different MLP structures are designed for different fault types to obtain fault detection models for railway facilities.

[0140] The dataset {wheel-rail noise time series, fault type} is divided into a training set, a validation set, and a test set; wherein the ratio of the training set, validation set, and test set is 6:2:2.

[0141] The parameters of the trained acoustic model are used as the initial parameters of the railway facility fault detection model.

[0142] The training set is input into the railway facility fault detection model, and the optimal parameters of the model are found by Bayesian search algorithm;

[0143] The validation set is input into the railway facility fault detection model to validate the model;

[0144] The test set is input into the railway facility fault detection model for fault diagnosis, and the detection effect of the model is tested.

[0145] The working principle of this invention is based on the fact that wheel-rail noise contains rich information about defects in wheels, tracks, and the substructure, thus enabling defect detection. Large model techniques can be used to learn acoustic representations of wheel-rail noise; that is, the acoustic model can learn features about various defects, providing a foundational model for defect detection. However, obtaining sample pairs of measured defects and their corresponding wheel-rail noise signals is extremely difficult. Wavelet decomposition and Lipschitz coefficients can be used to extract category labels corresponding to faults from wheel-rail noise samples, allowing for fine-tuning of the acoustic model using a labeled wheel-rail noise dataset. Therefore, this invention uses measured wheel-rail noise signals to train an acoustic model for railway facility health monitoring, enabling the simultaneous detection of multiple railway facility faults.

[0146] Compared with existing railway facility fault detection models, this invention uses a large amount of wheel-rail noise data to train an acoustic model based on self-supervised learning, and then uses a labeled wheel-rail noise dataset for fine-tuning, thereby enabling the simultaneous detection of multiple defects.

[0147] Example 2

[0148] like Figures 1 to 1 As shown in Figure 0, the difference between this embodiment and Embodiment 1 is that this embodiment is a real case simulating a railway site, and it further explains and applies the technical details in Embodiment 1.

[0149] The implementation is as follows:

[0150] The process is divided into the following ten steps: 1) Measure the wheel-rail noise signal generated during the actual train operation to form a wheel-rail noise dataset; 2) Divide the wheel-rail noise dataset into a labeled wheel-rail noise dataset and an unlabeled wheel-rail noise dataset according to a preset ratio; 3) Divide the labeled wheel-rail noise dataset into segmented wheel-rail noise time series of equal length according to a preset window length; 4) Perform a four-level discrete wavelet transform on each segment of the wheel-rail noise time series in the labeled wheel-rail noise dataset to obtain the maximum modulus and corresponding average value of the wavelet coefficients at each scale, and further determine the detection threshold; 5) Determine the suspected fault location based on the detection threshold; 6) Perform a continuous wavelet transform on each segment of the wheel-rail noise time series in the labeled wheel-rail noise dataset to solve for the Lipschitz coefficients at the suspected fault location and classify the fault type; 7) Determine the fault type corresponding to each segment of the wheel-rail noise time series in the labeled wheel-rail noise dataset based on the wavelet transform and Lipschitz coefficients, and generate a dataset of {wheel-rail noise time series, fault type}; 8) 9) Establish an acoustic model based on self-supervised learning and train it on an unlabeled wheel-rail noise dataset; 10) Fine-tune the trained acoustic model using a labeled wheel-rail noise dataset to obtain a railway facility fault detection model; 11) Input the wheel-rail noise signal to be detected into the railway facility fault detection model for inspection and identify the corresponding fault category.

[0151] The equipment used in the method of this invention includes:

[0152] A sound pressure sensor, which is positioned at the axle box of the train, is used to measure wheel-rail noise signals;

[0153] A pulse generator, which is arranged on the train axle, is used to control the sampling of wheel-rail noise. Sampling begins when the wheel is rolling and ends when the wheel stops.

[0154] The data acquisition module is connected to the sound pressure sensor and is used to acquire time-domain information of wheel-rail noise from the sound pressure signal.

[0155] The data processing module is connected to the data acquisition module and is used to process the data acquired by the data acquisition module into a format that can be used by a computer for fault identification.

[0156] The sound pressure sensor is positioned near the wheels at the axle box of the train to ensure that the relevant sound pressure signal generated by the wheel-rail interaction is acquired, thereby achieving the best recognition effect.

[0157] The following are examples of implementation methods:

[0158] (1) Data Description

[0159] A field test was conducted on a high-speed railway in China to obtain wheel-rail noise signals generated during train operation, such as... Figure 4 As shown.

[0160] (2) Data partitioning

[0161] The raw wheel-rail noise measurement dataset from the field is divided into labeled wheel-rail noise datasets and unlabeled wheel-rail noise datasets according to a preset ratio, such as... Figure 5 As shown.

[0162] (3) Tag extraction

[0163] Discrete wavelet transform is performed on segmented wheel-rail noise time series in a labeled wheel-rail noise dataset to determine suspected fault locations. Then, continuous wavelet transform is performed to solve for the Lipschitz coefficients corresponding to the suspected fault locations, determining the fault type and extent. Figure 6 As shown.

[0164] (4) Pre-trained model dataset

[0165] The unlabeled wheel-rail noise dataset is randomly divided into a training set, a validation set, and a test set; wherein the ratio of the training set, validation set, and test set is 8:1:1.

[0166] (5) Pre-trained model

[0167] An acoustic model is built based on self-supervised learning. It is trained using an unlabeled wheel-rail noise dataset. A greedy algorithm is used to optimize the parameters, yielding the optimal model parameters. Figure 7 As shown.

[0168] (6) Pseudo-label vector generation

[0169] The unlabeled wheel-rail noise dataset is input into the pseudo-label generation layer to obtain pseudo-labels for the corresponding vectors, such as... Figure 8 As shown.

[0170] (7) Model fine-tuning dataset

[0171] The labeled {wheel-rail noise time series, fault type} dataset is divided into training set, validation set and test set; the ratio of training set, validation set and test set is 6:2:2.

[0172] (8) Model fine-tuning

[0173] The labeled wheel-rail noise dataset is input into different MLPs for training to obtain a railway facility fault detection model. The optimal parameters of the model are then found using a Bayesian search algorithm.

[0174] (9) Output of detection results

[0175] The input wheel-rail noise signal to be detected is first processed into a specific format input vector, and then detected using a railway facility fault detection model to identify the type and location of the fault.

[0176] In summary, compared to traditional methods of detecting railway facility faults using specialized equipment such as integrated inspection vehicles, track inspection vehicles, and corrugated trolleys, this invention utilizes wheel-rail noise signals generated during commercial train operation to detect multiple defects, offering advantages such as low detection cost and high detection efficiency. Compared to existing railway facility fault detection models, this invention trains a self-supervised learning-based acoustic model using a large amount of wheel-rail noise data, and then fine-tunes it using labeled wheel-rail noise datasets. This enables simultaneous detection of multiple defects, facilitating the shift of railway operations from "planned maintenance" to "condition-based maintenance," and ensuring the safe, stable, and efficient operation of railways.

[0177] Those skilled in the art will understand that the accompanying drawings are merely schematic diagrams of one embodiment, and the modules or processes shown in the drawings are not necessarily essential for implementing the present invention.

[0178] As can be seen from the above description of the embodiments, those skilled in the art can clearly understand that the present invention can be implemented by means of software plus necessary general-purpose hardware platforms. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in various embodiments or some parts of the embodiments of the present invention.

[0179] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, for apparatus or system embodiments, since they are basically similar to method embodiments, the description is relatively simple; relevant parts can be referred to the descriptions in the method embodiments. The apparatus and system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0180] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.< / mask> < / mask> < / mask> < / mask> < / mask> < / mask>

Claims

1. A method for fault detection of railway facilities, characterized in that, include: Acquire measured wheel-rail noise signals during train operation, and obtain time-domain information of wheel-rail noise based on the measured wheel-rail noise signals; The acquired wheel-rail noise time-domain information is divided into labeled wheel-rail noise datasets and unlabeled wheel-rail noise datasets according to a preset ratio; A railway facility fault detection model is constructed, and the labeled wheel-rail noise dataset and the unlabeled wheel-rail noise dataset are used to train the railway facility fault detection model to obtain a trained railway facility fault detection model. The wheel-rail noise signal to be detected is input into the trained railway facility fault detection model for detection, and the fault type and location corresponding to the wheel-rail noise signal to be detected are obtained. The railway facility fault detection model is constructed by training the railway facility fault detection model using the labeled wheel-rail noise dataset and the unlabeled wheel-rail noise dataset to obtain a trained railway facility fault detection model. The labeled wheel-rail noise dataset is segmented into segmented wheel-rail noise time series according to the preset window length, and a segmented wheel-rail noise time series dataset is constructed. Perform discrete wavelet transform on each segment of wheel-rail noise time series in the labeled wheel-rail noise dataset, calculate Lipschitz coefficients, and obtain the fault type label corresponding to each segment of wheel-rail noise time series. An acoustic model based on self-supervised learning is established. The unlabeled wheel-rail noise dataset is input into the acoustic model for training. The acoustic representation of wheel-rail noise is learned, and a trained acoustic model is obtained. The fault type label corresponding to each segment of wheel-rail noise time series is used to adjust the trained acoustic model to obtain a trained railway facility fault detection model. The step of performing discrete wavelet transform on each segment of the wheel-rail noise time series in the labeled wheel-rail noise dataset, calculating Lipschitz coefficients, and obtaining the fault type label corresponding to each segment of the wheel-rail noise time series includes: Discrete wavelet transform is performed on the segmented wheel-rail noise time series according to equations (2) and (3). (2) (3) In the formula, h is the coefficient vector of the scaling filter, m = 2k + n, k is the translation amount of the mother wavelet, n is the integer index of the filter, and h1 is the wavelet filter. A four-level multi-resolution analysis was performed on the segmented wheel-rail noise time series to obtain wavelet coefficient information at five scales. The maximum modulus of the wavelet coefficient at each scale was found, and the corresponding average value A was calculated to determine the detection threshold λ, as shown in equations (4) and (5). (4) (5) In the formula, a i K represents the maximum modulus of wavelet coefficients at each scale, and K is a threshold control coefficient with a value ranging from 1.2 to 1.

5. Wavelet coefficients at each scale and Wavelet coefficients below the threshold λ are directly reset to 0, and other wavelet coefficients above the threshold λ are marked as suspected fault locations u. Perform continuous wavelet transform on the segmented wheel-rail noise time series according to equation (6); (6) In the formula, For the mother wavelet, a is the scale parameter and b is the positioning parameter; The Lipschitz coefficient corresponding to the suspected fault location u is calculated according to equation (6), as shown in equation (7); (7) In the formula, J corresponds to the maximum scale of wavelet decomposition; The fault type and location are determined based on the suspected fault location and the corresponding Lipschitz coefficient. The fault types include wheel flattening, rail flattening, and rail corrugation. Signals with a Lipschitz coefficient greater than 0 at suspected fault locations in scales 1-5 of the wheel-rail noise signal decomposition are identified as rail corrugation; signals with a Lipschitz coefficient greater than 1 at suspected fault locations in scales 1-3 are identified as rail flattening; and signals with a Lipschitz coefficient less than 1 at suspected fault locations in scales 1-2 are identified as wheel flattening. Based on the suspected fault locations and the determined fault types and corresponding wheel-rail noise time series ranges in the segmented wheel-rail noise time series, a new labeled wheel-rail noise dataset is constructed, with the format {wheel-rail noise time series, fault type}.

2. The method according to claim 1, characterized in that, The acquisition of measured wheel-rail noise signals during train operation, and the acquisition of time-domain information of wheel-rail noise based on the wheel-rail noise signals, includes: A sound pressure sensor is installed on the side of the train frame near the wheel, and the sound pressure sensor is used to collect the acoustic signal generated by the vibration radiation of the wheel and rail. A pulse generator is installed on the axle of the train. When the wheel starts to roll, the sound pressure sensor is triggered to sample the acoustic signal. When the wheel stops rolling, the sound pressure sensor stops sampling the acoustic signal. The above operation is repeated on different trains on a certain railway line to obtain a wheel-rail noise dataset for a set duration, and the time-domain information of the wheel-rail noise is obtained based on the wheel-rail noise dataset.

3. The method according to claim 1, characterized in that, The step of segmenting the labeled wheel-rail noise dataset into segmented wheel-rail noise time series according to a preset window length, and constructing a segmented wheel-rail noise time series dataset, includes: Suppose the wheel-rail noise signals in the labeled wheel-rail noise dataset are x1, x2, ..., x3. n The segmentation window length L is determined based on the train's operating speed. Based on the segmentation window length L, the wheel-rail noise signals x1, x2, ..., x... are analyzed. n The labeled wheel-rail noise dataset is segmented into segmented wheel-rail noise time series. Each segment of wheel-rail noise signal x i The i=1,2,…,n signal is divided into N equal-length wheel-rail noise signals y. j For j=1,2,…,N, the wheel-rail noise time series with a length less than L are directly discarded; (1)。 4. The method according to claim 3, characterized in that, The process of establishing a self-supervised learning-based acoustic model involves inputting the unlabeled wheel-rail noise dataset into the acoustic model for training, learning the acoustic representation of wheel-rail noise, and obtaining a trained acoustic model. An acoustic model is constructed, consisting of convolutional layers, Transformer layers, fully connected layers, and a pseudo-label generation layer. The acoustic model is based on self-supervised learning. The unlabeled wheel-rail noise dataset is divided into a training set, a validation set, and a test set; wherein the ratio of the training set, validation set, and test set is 8:1:

1. The training set is input into the acoustic model, and the optimal parameters of the model are found by a greedy algorithm. The training set is divided according to a preset window length L to obtain segmented unlabeled wheel-rail noise time series. Each segment of the unlabeled wheel-rail noise time series is downsampled through a convolutional layer, and the output length is determined by equation (8). (8) In the formula, M is the input length, i.e. the sampling points contained in the window, K is the convolution kernel size, the stride is S, and P zeros are padded at both ends of the input. The input raw, unlabeled wheel-rail noise time series is processed by a convolutional layer to produce H vectors v. i Let i = 1, 2, ..., H, and each vector have a length of O. Use a K-means classifier to classify the H vectors v. i The classification process generates a corresponding pseudo-label vector V' for each vector. H vectors are then randomly masked and fed into a Transformer layer to produce H output vectors v. i ’ , i=1,2,…,H; v i ’ The output vector with the Mask operation is then input into the fully connected layer to obtain the target output vector V. The distance D between the target output vector V and the corresponding pseudo-label vector V' is calculated, as shown in equation (9). (9) Repeat the above operations, use the Adam optimizer to update the parameters, and the objective function is as shown in equation (10); (10) In the formula, w and b are the learnable parameters in the neural network model; The validation set is input into the acoustic model to validate the acoustic model. The test set is input into the acoustic model to perform random vector matching prediction to evaluate the effect of the acoustic model in learning the acoustic representation of wheel-rail noise. After multiple rounds of iterative training, a trained acoustic model is obtained.

5. The method according to claim 4, characterized in that, The process involves randomly masking H vectors and then inputting them into a Transformer layer to obtain H output vectors v. i ’ ,include: H vectors are randomly masked and then fed into a Transformer layer to obtain H output vectors v. i ’ (i=1,2,…,H), each time the original unlabeled wheel-rail noise time series is input, after passing through the convolutional layer, H vectors are generated. 20% of the vectors are randomly selected and replaced with specific symbols. <mask>When the selected vector is permuted into <mask>At that time, there is a 60% probability that it will continue to remain unchanged. <mask> It remains unchanged, with a 20% probability of being replaced by another random vector, and a remaining 20% ​​probability that the vector is left unchanged.< / mask> < / mask> < / mask> The H processed vectors are input into a Transformer layer, which consists of 12 Transformer blocks, with a hidden layer size of 768 and a header of 12. The Transformer layer outputs H output vectors v. i ’ .

6. The method according to claim 1, characterized in that, The method of adjusting the trained acoustic model using fault type labels corresponding to each segment of wheel-rail noise time series to obtain a trained railway facility fault detection model specifically includes: Different multilayer perceptron (MLP) structures were designed for different fault types to obtain fault detection models for railway facilities. The new labeled wheel-rail noise dataset in the format {wheel-rail noise time series, fault type} is divided into a training set, a validation set, and a test set, with a ratio of 6:2:

2. The parameters of the trained acoustic model are used as the initial parameters of the railway facility fault detection model. The training set is input into the railway facility fault detection model, and the optimal parameters of the railway facility fault detection model are found by Bayesian search algorithm; The validation set is input into the railway facility fault detection model to validate the railway facility fault detection model; The test set is input into the railway facility fault detection model for fault diagnosis, and the detection effect of the railway facility fault detection model is tested. After multiple rounds of iterative training, a trained railway facility fault detection model is obtained.

Images