Train wheel health detection method and system based on semi-supervised generative adversarial model

Abstract

The application provides a train wheel health detection method and system based on a semi-supervised generative adversarial model, comprising: acquiring rail vibration data; transforming the rail vibration data to obtain time-frequency pattern samples, and dividing the time-frequency pattern samples into labeled samples and unlabeled samples according to label information; constructing a semi-supervised deep generative adversarial network multi-label classification model, and training and testing on the time-frequency pattern samples to obtain a trained wheel health detection model; and inputting to-be-detected rail vibration data into the trained wheel health detection model to obtain a wheel defect type. The application trains a large number of unlabeled samples by using a generative adversarial network, so that a discriminator fully learns the potential features of a time-frequency pattern, and then a small amount of labeled samples are used for supervised learning, so that subsequent wheel health detection has the advantages of high efficiency, accuracy, saving of operation cost and the like.

Description

Technical Field

[0001] This invention relates to the field of rail transit technology, and in particular to a method and system for detecting train wheel health based on a semi-supervised generative adversarial model. Background Technology

[0002] Railways are the main artery of the national economy and a vital infrastructure essential to people's livelihoods. The wheel-rail contact relationship, as the most fundamental constraint in the vehicle-track coupling system, directly determines the operational safety of trains. With the increasing scale of my country's railway network, rising transport demands, and more complex operating environments, wheel-rail related defects affecting train operation safety are gradually emerging. These include wheel defects such as out-of-roundness, wheel abrasion, wheel spalling, wheel cracks, and wheel scuffing. Once these defects occur, they can range from reducing train running stability and increasing environmental vibration and noise to causing unpredictable and serious accidents such as derailments.

[0003] Currently, there are various methods for detecting wheel defects, typically including manual inspection, magnetic particle testing, and ultrasonic testing. These methods require personnel with considerable professional knowledge and experience, and also necessitate disassembling the train body, making the process complex and inefficient, which contradicts the goal of building smart railways. Additionally, there are methods that use axle box acceleration signals to identify wheel defects. While this allows for real-time online monitoring of wheel conditions, it requires installing acceleration sensors on the axle box of each wheel, increasing maintenance costs. Therefore, achieving rapid perception, intelligent diagnosis, and maintenance management of train wheel service conditions is a primary challenge facing railway operations departments.

[0004] Existing intelligent wheel defect detection models are mostly trained using supervised learning. However, unlike ordinary fault detection, negative samples for wheel defect detection are relatively few and difficult to obtain. This leads to an imbalance between positive and negative samples in the dataset, resulting in poor model performance in practical deployments. Therefore, there is an urgent need to develop a deep learning neural network model based on fewer labeled samples, which can simultaneously detect and identify multiple wheel defects. Summary of the Invention

[0005] The purpose of this invention is to provide a high-precision and high-efficiency train wheel health detection method and system based on a semi-supervised generative adversarial model that can simultaneously identify at least one of the technical problems existing in the background art.

[0006] To achieve the above objectives, the present invention adopts the following technical solution.

[0007] In a first aspect, the present invention provides a train wheel health detection method based on a semi-supervised generative adversarial model, comprising:

[0008] Acquire rail vibration data, which is the vibration data generated by the wheels acting on the rails during train operation;

[0009] The rail vibration data is divided into rail vibration data corresponding to each wheel, and the rail vibration data corresponding to each wheel is transformed to obtain a time-frequency map sample.

[0010] The time-frequency graph samples are divided into labeled samples and unlabeled samples based on the label information;

[0011] A semi-supervised deep generative adversarial network multi-label classification model is constructed and trained and tested on the time-frequency map samples to obtain a trained wheel health detection model.

[0012] Input the vibration data of the rail to be detected into the trained wheel health detection model to obtain the wheel defect type.

[0013] Optionally, the step of segmenting the rail vibration data into rail vibration data corresponding to each wheel includes:

[0014] Wheel-rail force sensors are installed at the beginning and end of the train passage area. The wheel-rail force sensors are used to determine the time of wheel passage. When the train passes through the area, the wheel-rail force sensors are triggered to collect rail strain signals, and the rail vibration acceleration sensors are simultaneously triggered to collect rail vibration signals. The time of passage of each wheel is determined based on the strain signals.

[0015] The rail vibration signals collected by different rail vibration acceleration sensors are divided according to the time of each wheel's passage, and then spliced ​​together on the time axis according to the spatial position of the rail vibration acceleration sensors to obtain the time domain signal of rail vibration acceleration corresponding to each wheel, which is used as the rail vibration data corresponding to each wheel.

[0016] Optionally, the step of transforming the rail vibration data corresponding to each wheel to obtain time-frequency map samples includes:

[0017] The continuous wavelet transform of the rail vibration data corresponding to each wheel is converted to the frequency domain to obtain the continuous wavelet transform of the rail vibration data in the frequency domain. :

[0018]

[0019] In the formula, a and b are the scale factor and time shift, respectively. For the Fourier transform of the vibration signal, ω is the angular frequency and i is the imaginary sign;

[0020] The instantaneous angular frequency of the rail vibration data is calculated based on the continuous wavelet transform of the frequency domain rail vibration data. :

[0021]

[0022] Based on the continuous wavelet transform of the frequency domain rail vibration data and the instantaneous angular frequency of the rail vibration data, the synchronous compressed wavelet transform of the rail vibration data is obtained. :

[0023]

[0024] In the formula, δ is the Dirac function. The center frequency;

[0025] The time-frequency diagram of the rail vibration data is obtained by synchronous compressed wavelet transform of the rail vibration data.

[0026] Optionally, the label information includes at least wheel out of round, wheel scratches, wheel peeling, wheel cracks, wheel dents, and normal wheels.

[0027] Optionally, the labeled sample corresponds to at least one label;

[0028] The label is represented by a vector of length S:

[0029]

[0030] The value at each position in the vector is 0 or 1. When the value is 1, it means that there is a labeled sample with the wheel defect type at the corresponding position. When the value is 0, it means that there is a labeled sample without the wheel defect type at the corresponding position. S is the number of defect categories.

[0031] Optionally, the semi-supervised deep generative adversarial network multi-label classification model consists of a generator, an unsupervised discriminator, and a supervised discriminator;

[0032] The generator comprises, in sequence, a random noise input layer, a fully connected layer, an activation layer, a reshape layer, a transposed convolutional layer, an activation layer, a transposed convolutional layer, an activation layer, and another transposed convolutional layer, wherein the activation layer uses LeakyReLU as the activation function.

[0033]

[0034] In the formula, x represents the output of each layer. Leakage coefficient;

[0035] The generator is used to take a random noise signal of a specified dimension as input and output an image of a specified size.

[0036] The supervised discriminator comprises, in sequence, a labeled sample input layer, a convolutional layer, an activation layer, another convolutional layer, an activation layer, a convolutional layer, an activation layer, a flattening layer, a fully connected layer, and an output layer. The activation layer uses LeakyReLU as the activation function, the fully connected layer uses dropout to reduce overfitting, and the output layer uses the sigmoid function to activate each node.

[0037]

[0038] A supervised discriminator takes labeled samples as input and outputs predicted sample labels.

[0039] The unsupervised discriminator comprises an unlabeled sample input layer, a convolutional layer, an activation layer, another convolutional layer, an activation layer, a convolutional layer, an activation layer, a flattening layer, a fully connected layer, and an output layer, connected in sequence. The activation layer uses LeakyReLU as the activation function, the fully connected layer uses dropout to reduce overfitting, and the output layer uses the activation function shown in Equation (7) to activate each node.

[0040]

[0041] In the formula, K represents the number of categories, l k (x) represents the output value predicted by the network for each category;

[0042] The unsupervised discriminator takes as input unlabeled samples and the generator's output image, and outputs either true or false.

[0043] Optionally, the step of training and testing on the time-frequency map samples to obtain a trained wheel health detection model includes:

[0044] The input is labeled samples, and the loss function is binary cross-entropy. The parameters are trained and updated on a supervised discriminator.

[0045]

[0046] In the formula, y is the category label 0 or 1, and p(y) represents the probability that the network predicts that the label is y;

[0047] Input a random noise signal, and the generator outputs an image;

[0048] Input unlabeled samples and images output by the generator, use binary cross-entropy as the loss function, train and update parameters on an unsupervised discriminator;

[0049] With the parameters of the unsupervised discriminator fixed, random noise signal is input, and binary cross-entropy is used as the loss function to train and update the parameters on the generator.

[0050] Training stops when the training loss no longer decreases within a preset number of iterations. The parameters from the last training iteration are saved as the optimal parameters of the model. The trained wheel health detection model is then obtained based on these optimal parameters.

[0051] Secondly, the present invention provides a train wheel health detection system based on a semi-supervised generative adversarial model, the system comprising:

[0052] The data acquisition module is used to acquire rail vibration data, which is the vibration data generated by the wheels acting on the rails during train operation;

[0053] The data segmentation module is used to segment the rail vibration data into rail vibration data corresponding to each wheel, and transform the rail vibration data corresponding to each wheel to obtain a time-frequency diagram sample.

[0054] The sample acquisition module is used to divide the time-frequency graph samples into labeled samples and unlabeled samples according to the label information.

[0055] The model training module is used to construct a semi-supervised deep generative adversarial network multi-label classification model, and to train and test it on the time-frequency map samples to obtain a trained wheel health detection model.

[0056] The detection module is used to input the vibration data of the rail to be detected into the trained wheel health detection model to obtain the wheel defect type.

[0057] Thirdly, the present invention provides a non-transitory computer-readable storage medium for storing computer instructions, which, when executed by a processor, implement the train wheel health detection method based on a semi-supervised generative adversarial model as described above.

[0058] Fourthly, the present invention provides a computer program product, including a computer program that, when run on one or more processors, is used to implement the train wheel health detection method based on a semi-supervised generative adversarial model as described above.

[0059] Fifthly, the present invention provides an electronic device, comprising: a processor, a memory, and a computer program; wherein the processor is connected to the memory, the computer program is stored in the memory, and when the electronic device is running, the processor executes the computer program stored in the memory to cause the electronic device to execute instructions for implementing the train wheel health detection method based on a semi-supervised generative adversarial model as described above.

[0060] The beneficial effects of this invention are as follows: The train wheel health detection method and device based on a semi-supervised generative adversarial model provided by this invention obtains time-frequency map samples by collecting rail vibration data and converting time-frequency maps. These time-frequency map samples are then divided into a small number of labeled samples and a large number of unlabeled samples. A generative adversarial network is used to train on the large number of unlabeled samples, allowing the discriminator to fully learn the latent features of the time-frequency maps. Supervised learning is then performed using a small number of labeled samples. This enables the finally trained wheel health detection model to perform health detection on different wheel conditions corresponding to the labeled information. The wheel health detection method of this invention has advantages such as high efficiency, accuracy, and cost savings, which is beneficial for railway operators to conduct scientific maintenance and promotes the green development of rail transit.

[0061] 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

[0062] 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.

[0063] Figure 1 This is one of the flowcharts illustrating a train wheel health detection method based on a semi-supervised generative adversarial model provided in an embodiment of the present invention;

[0064] Figure 2 This is the second flowchart illustrating a train wheel health detection method based on a semi-supervised generative adversarial model, as provided in an embodiment of the present invention.

[0065] Figure 3 The third flowchart illustrates a train wheel health detection method based on a semi-supervised generative adversarial model, as provided in an embodiment of the present invention.

[0066] Figure 4 A schematic diagram illustrating the acquisition of the final rail vibration data provided in an embodiment of the present invention;

[0067] Figure 5 This is a schematic diagram of a wheel passing over a rail during the acquisition of rail vibration data, as provided in an embodiment of the present invention.

[0068] Figure 6 A schematic diagram illustrating the health status of wheels provided in an embodiment of the present invention;

[0069] Figure 7 This is a schematic diagram of the generator provided in an embodiment of the present invention;

[0070] Figure 8 This is a schematic diagram of the structure of a supervised discriminator provided in an embodiment of the present invention;

[0071] Figure 9 This is a schematic diagram of the structure of the unsupervised discriminator provided in an embodiment of the present invention;

[0072] Figure 10 This is a schematic diagram of detection data transmission provided in an embodiment of the present invention. Detailed Implementation

[0073] Embodiments of the present invention are described in detail below. Examples of these embodiments 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.

[0074] Those skilled in the art will understand that, unless specifically stated otherwise, the singular forms “a,” “an,” “the,” 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 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.

[0075] 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.

[0076] To facilitate understanding of the present invention, the present invention will be further explained and described below with reference to the accompanying drawings and specific embodiments. However, the specific embodiments do not constitute a limitation on the embodiments of the present invention.

[0077] Those skilled in the art should understand that the accompanying drawings are merely schematic diagrams of embodiments, and the components in the drawings are not necessarily essential for implementing the present invention.

[0078] Example 1

[0079] Figure 1 This is one of the flowcharts illustrating a train wheel health detection method based on a semi-supervised generative adversarial model provided in an embodiment of the present invention; Figure 2 This is the second flowchart illustrating a train wheel health detection method based on a semi-supervised generative adversarial model, as provided in an embodiment of the present invention. Figure 3 This is the third flowchart illustrating the train wheel health detection method based on a semi-supervised generative adversarial model provided in this embodiment of the invention. Figure 1-3 As shown, the train wheel health detection method based on a semi-supervised generative adversarial model includes the following steps:

[0080] S101, acquire rail vibration data, the rail vibration data being the vibration data generated by the wheels acting on the rail during train operation.

[0081] Specifically, when a train is running, the wheels exert vibration data on the rails, which is obtained by using an acceleration sensor installed in a target section of the track.

[0082] It should be noted that, in order to minimize the impact of track irregularities, the target section in this embodiment refers to a straight section within the train line where the rails are in good condition, the track structure is simple, and the train is traveling at a constant speed. Furthermore, this embodiment uses high-frequency accelerometers with a range of ±200g, a spacing of 0.2m between each accelerometer, and a sampling frequency of 10000Hz, thereby ensuring the acquisition of complete signals excited by wheel defects.

[0083] S102, the rail vibration data is divided into rail vibration data corresponding to each wheel, and the rail vibration data corresponding to each wheel is transformed to obtain a time-frequency map sample.

[0084] The rail vibration data was obtained in the following way:

[0085] Wheel-rail force sensors are installed at the beginning and end of the train passage area (i.e., the target section). The wheel-rail force sensors are used to determine the time it takes for the wheels to pass through the area. When the train passes through the area, the wheel-rail force sensors are triggered to collect rail strain signals, and the rail vibration acceleration sensors are triggered simultaneously to collect rail vibration signals. The time it takes for each wheel to pass through the area is determined based on the strain signals.

[0086] The rail vibration signals collected by different rail vibration acceleration sensors are divided according to the time of each wheel's passage, and then spliced ​​together on the time axis according to the spatial position of the rail vibration acceleration sensors to obtain the time domain signal of rail vibration acceleration corresponding to each wheel, which is used as the rail vibration data corresponding to each wheel.

[0087] The length L of the target road segment is determined by the wheel radius R of the train, as shown in equation (1);

[0088] (1)

[0089] In the formula, The symbol indicates rounding up, meaning L must be greater than or equal to the circumference of the wheel's rolling circle.

[0090] Figure 4 This is a schematic diagram illustrating the acquisition of the final rail vibration data provided in an embodiment of the present invention; as shown. Figure 4 As shown, the specific segmentation process of the rail vibration data is as follows:

[0091] First, wheel-rail force sensors installed at the start and end points of the target road segment are used to determine the start time t1 and end time t2 of each wheel's passage through the target road segment. That is, the wheel-rail force sensors collect the rail strain signal, and then the time is determined based on the rail strain signal, thereby calculating the time for each wheel to pass through. ;

[0092] (2)

[0093] The initial rail vibration data collected by different acceleration sensors are divided according to the time of each wheel's passage, and then spliced ​​together on the time axis according to the spatial position of the acceleration sensors to obtain the time domain signal of rail vibration acceleration corresponding to each wheel, which is the final rail vibration data.

[0094] In practical applications, a schematic diagram of the rail vibration data acquisition process is shown below. Figure 5 As shown, vibration and rail strain signals are acquired through acceleration sensors and wheel-rail force sensors installed on the rails in the target section (also called the detection area). Additionally, as... Figure 10 As shown, data cables are set up in the detection area to transmit sensor data, and data from each sensor in the entire detection area is collected by a trackside server and then transmitted to the central control console.

[0095] After acquiring the time-domain signal of rail vibration acceleration corresponding to each wheel, the rail vibration data corresponding to each wheel is transformed to obtain a time-frequency map sample, specifically including:

[0096] First, the rail vibration data is high-pass filtered to eliminate interference from low-frequency noise signals such as trend terms;

[0097] Then, the Morlet wavelet is selected as the mother wavelet. According to Parseval's theorem, the continuous wavelet transform of the rail vibration data corresponding to each wheel is converted to the frequency domain, resulting in the continuous wavelet transform of the rail vibration data in the frequency domain. As shown in equation (3);

[0098] (3)

[0099] In the formula, a and b are the scale factor and time shift, respectively. This is the Fourier transform of the filtered rail vibration data, where ω is the angular frequency and i is the imaginary sign.

[0100] Next, based on the continuous wavelet transform of the frequency domain rail vibration data calculated by equation (3), the instantaneous angular frequency of the rail vibration data is calculated. As shown in equation (4);

[0101] (4)

[0102] Next, based on the continuous wavelet transform of the frequency domain rail vibration data and the instantaneous angular frequency of the rail vibration data, the synchronous compressed wavelet transform of the rail vibration data is obtained. That is, the wavelet coefficients are compressed and rearranged according to the instantaneous angular frequency of the time-frequency plane. Here, the frequency resolution is set to 1Hz, thus obtaining the synchronous compressed wavelet transform of the vibration data. The result of this transformation is the time-frequency diagram sample, as shown in equation (5);

[0103] (5)

[0104] In the formula, δ is the Dirac function. The center frequency.

[0105] Finally, the time-frequency diagram of the rail vibration data is obtained by synchronous compressed wavelet transform of the rail vibration data.

[0106] S103, the time-frequency map samples are divided into labeled samples and unlabeled samples according to the label information.

[0107] In this step, the time-frequency map samples are divided into labeled samples and unlabeled samples based on the label information. The label information includes at least the following: wheel out-of-roundness, wheel scratches, wheel peeling, wheel cracks, wheel dents, and normal wheels, specifically as follows: Figure 6 The diagram shows the health status of the wheels.

[0108] The sample classification here involves treating 1% of the time-frequency image samples as labeled samples, with label types including {wheel not round, wheel scratch, wheel peeling, wheel crack, wheel dent, normal wheel}. Each time-frequency image sample can have one or more labels, obtained through annotation. The remaining 99% of the time-frequency image samples are treated as unlabeled samples.

[0109] In this embodiment, the time-frequency image sample size is no less than 60,000 images, the number of labeled samples is no less than 120 images, and the number of labeled samples corresponding to each type of label information is no less than 20 images. The resolution of the time-frequency image sample is 256×256.

[0110] In the actual training process, the labels are converted into vectors of length S as shown in equation (6);

[0111] (6)

[0112] In the formula, the value of each position in the vector is 0 or 1. When the value is 1, it means that there is a labeled sample with the corresponding wheel defect type. When the value is 0, it means that there is a labeled sample without the corresponding wheel defect type. S is the number of defect categories. That is, the number of each position is 0 or 1. 1 represents having the corresponding type and 0 represents not having the corresponding type. S is the number of defect categories.

[0113] S104. Construct a semi-supervised deep generative adversarial network multi-label classification model, and train and test it on the time-frequency map samples to obtain a trained wheel health detection model (i.e., a trained semi-supervised deep generative adversarial network multi-label classification model).

[0114] The semi-supervised deep generative adversarial network multi-label classification model consists of a generator, an unsupervised discriminator, and a supervised discriminator.

[0115] The structural diagram of the generator is shown below. Figure 7 As shown, it includes a random noise input layer, a fully connected layer, an activation layer, a reshape layer, a transposed convolutional layer, an activation layer, a transposed convolutional layer, an activation layer, and a transposed convolutional layer connected in sequence, wherein the activation layer uses LeakyReLU as the activation function, as shown in Equation (7).

[0116] (7)

[0117] Where x represents the output of each layer. This is the leakage coefficient, which is usually taken as 0.01.

[0118] The generator takes a random noise signal of a specified dimension as input and outputs an image of size 256×256.

[0119] The structural schematic diagram of the supervised discriminator is shown below. Figure 8 As shown, it includes a label sample input layer, a convolutional layer, an activation layer, a convolutional layer, an activation layer, a convolutional layer, an activation layer, a flattening layer, a fully connected layer, and an output layer, which are connected in sequence. The activation layer uses LeakyReLU as the activation function, the fully connected layer uses dropout to reduce overfitting, and the output layer uses the sigmoid function to activate each node, as shown in Equation (8).

[0120] (8)

[0121] A supervised discriminator takes labeled samples as input and outputs predicted sample labels.

[0122] The structural schematic diagram of the unsupervised discriminator is shown below. Figure 9 As shown, it includes an unlabeled sample input layer, a convolutional layer, an activation layer, a convolutional layer, an activation layer, a convolutional layer, an activation layer, a flattening layer, a fully connected layer, and an output layer connected in sequence; wherein the activation layer uses LeakyReLU as the activation function, the fully connected layer uses dropout to reduce overfitting, and the output layer uses the activation function shown in Equation (9) to activate each node.

[0123] (9)

[0124] In the formula, K represents the number of categories, l k (x) represents the output value predicted by the network for each category.

[0125] The unsupervised discriminator takes as input unlabeled samples and the generator's output image, and outputs either true or false.

[0126] In this step, training and testing on the time-frequency map samples to obtain a trained wheel health detection model includes:

[0127] The input is labeled samples, and the loss function is binary cross-entropy. The parameters are trained and updated on a supervised discriminator.

[0128]

[0129] In the formula, y is the category label 0 or 1, and p(y) represents the probability that the network predicts that the label is y.

[0130] Input a random noise signal, and the generator outputs an image.

[0131] Input unlabeled samples and images output by the generator, use binary cross-entropy as the loss function, train and update parameters on an unsupervised discriminator.

[0132] With the parameters of the unsupervised discriminator fixed, random noise signal is input, and binary cross-entropy is used as the loss function to train and update the parameters on the generator.

[0133] Training stops when the training loss no longer decreases within a preset number of iterations. The parameters from the last training iteration are saved as the optimal parameters of the model. The trained wheel health detection model is then obtained based on these optimal parameters.

[0134] Specifically, in the actual training process, the following settings can be made: the number of training epochs is 100, the size of the labeled samples is 120, the dimension of the random noise signal input to the generator is 100, the batch size is 120, the Adam optimizer is used when updating the network parameters in each iteration, and the learning rate is 0.0002.

[0135] The specific training details include: First, input labeled images into the supervised discriminator, use binary cross-entropy as the loss function Loss, as shown in Equation (10), and train and update the parameters on the supervised discriminator; Second, input random noise signals into the generator, and the generator outputs images; Third, input unlabeled images and images output by the generator into the unsupervised discriminator, use binary cross-entropy as the loss function, and train and update the parameters on the unsupervised discriminator; Fourth, fix the parameters of the unsupervised discriminator, input random noise signals, use binary cross-entropy as the loss function, and train and update the parameters on the generator.

[0136] (10)

[0137] In the formula, y is the category label 0 or 1, p(y) represents the probability that the network predicts that the sample belongs to label y, N is the number of samples, and i is the i-th sample.

[0138] S105, input the vibration data of the rail to be detected into the wheel health detection model to obtain the wheel defect type.

[0139] Obtain the vibration data of the rail to be tested. That is, acquire the initial vibration data of the rail to be tested in the target section through the aforementioned accelerometer. Then, in the same way as in the training process, segment the vibration data and convert it into a time-frequency map to obtain the time-frequency map to be tested. This time-frequency map to be tested is the final vibration data to be tested, which will not be described in detail here.

[0140] The vibration data to be detected is input into a pre-trained wheel health detection model to obtain the wheel defect type. The wheel defect type includes one or more of the following: wheel out of round, wheel abrasion, wheel peeling, wheel crack, wheel dent, and normal wheel.

[0141] The train wheel health detection method based on a semi-supervised generative adversarial model provided by the present invention obtains time-frequency map samples by collecting rail vibration data and converting time-frequency maps. These time-frequency map samples are then divided into a small number of labeled samples and a large number of unlabeled samples. A generative adversarial network is used to train the discriminator on the large number of unlabeled samples, allowing it to fully learn the latent features of the time-frequency maps. Supervised learning is then performed using a small number of labeled samples. This enables the finally trained wheel health detection model to perform health detection on different wheel conditions corresponding to the labeled information. The wheel health detection method of the present invention has advantages such as high efficiency, accuracy, and cost savings, which is beneficial for railway operators to conduct scientific maintenance and promotes the green development of rail transit.

[0142] Example 2

[0143] Based on Example 1, Example 2 provides a train wheel health detection system based on a semi-supervised generative adversarial model. This train wheel health detection system corresponds to the above-mentioned train wheel health detection method based on a semi-supervised generative adversarial model, and specifically includes:

[0144] The data acquisition module is used to acquire rail vibration data, which is the vibration data generated by the wheels acting on the rails during train operation;

[0145] The data segmentation module is used to segment the rail vibration data into rail vibration data corresponding to each wheel, and transform the rail vibration data corresponding to each wheel to obtain a time-frequency diagram sample.

[0146] The sample acquisition module is used to divide the time-frequency graph samples into labeled samples and unlabeled samples according to the label information.

[0147] The model training module is used to construct a semi-supervised deep generative adversarial network multi-label classification model, and to train and test it on the time-frequency map samples to obtain a trained wheel health detection model.

[0148] The detection module is used to input the vibration data of the rail to be detected into the trained wheel health detection model to obtain the wheel defect type.

[0149] For specific details, please refer to the description in the section on train wheel health detection method based on semi-supervised generative adversarial model, which will not be repeated here.

[0150] Example 3

[0151] Embodiment 3 of the present invention provides an electronic device, including a memory and a processor, which communicate with each other. The memory stores program instructions that can be executed by the processor. The processor calls the program instructions to execute a train wheel health detection method based on a semi-supervised generative adversarial model. The train wheel health detection method based on a semi-supervised generative adversarial model includes the following steps:

[0152] Acquire rail vibration data, which is the vibration data generated by the wheels acting on the rails during train operation;

[0153] The rail vibration data is divided into rail vibration data corresponding to each wheel, and the rail vibration data corresponding to each wheel is transformed to obtain a time-frequency map sample.

[0154] The time-frequency graph samples are divided into labeled samples and unlabeled samples based on the label information;

[0155] A semi-supervised deep generative adversarial network multi-label classification model is constructed and trained and tested on the time-frequency map samples to obtain a trained wheel health detection model.

[0156] Input the vibration data of the rail to be detected into the trained wheel health detection model to obtain the wheel defect type.

[0157] Example 4

[0158] Embodiment 4 of the present invention provides a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, it implements a train wheel health detection method based on a semi-supervised generative adversarial model. The method includes the following steps:

[0159] Acquire rail vibration data, which is the vibration data generated by the wheels acting on the rails during train operation;

[0160] The rail vibration data is divided into rail vibration data corresponding to each wheel, and the rail vibration data corresponding to each wheel is transformed to obtain a time-frequency map sample.

[0161] The time-frequency graph samples are divided into labeled samples and unlabeled samples based on the label information;

[0162] A semi-supervised deep generative adversarial network multi-label classification model is constructed and trained and tested on the time-frequency map samples to obtain a trained wheel health detection model.

[0163] Input the vibration data of the rail to be detected into the trained wheel health detection model to obtain the wheel defect type.

[0164] Example 5

[0165] Embodiment 5 of the present invention provides a computer program product, including a computer program that, when run on one or more processors, implements the train wheel health detection method based on a semi-supervised generative adversarial model as described above. The method includes the following steps:

[0166] Acquire rail vibration data, which is the vibration data generated by the wheels acting on the rails during train operation;

[0167] The rail vibration data is divided into rail vibration data corresponding to each wheel, and the rail vibration data corresponding to each wheel is transformed to obtain a time-frequency map sample.

[0168] The time-frequency graph samples are divided into labeled samples and unlabeled samples based on the label information;

[0169] A semi-supervised deep generative adversarial network multi-label classification model is constructed and trained and tested on the time-frequency map samples to obtain a trained wheel health detection model.

[0170] Input the vibration data of the rail to be detected into the trained wheel health detection model to obtain the wheel defect type.

[0171] In summary, this invention obtains time-frequency map samples by collecting rail vibration data and converting them into time-frequency maps. These samples are then divided into a small number of labeled samples and a large number of unlabeled samples. A supervised discriminator is trained using the small number of labeled samples, and an unsupervised discriminator is trained using the unlabeled samples and samples generated by the generator. In other words, a generative adversarial network is used to train the discriminator on a large number of unlabeled samples, allowing it to fully learn the latent features of the time-frequency maps. Supervised learning is then performed using a small number of labeled samples. This enables the finally trained wheel health detection model to perform health detection on different wheel conditions corresponding to the labeled information. This invention's wheel health detection method has advantages such as high efficiency, accuracy, and cost savings, which is beneficial for railway operators to conduct scientific maintenance and contributes to the green development of rail transit.

[0172] 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.

[0173] 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 method or system embodiments, since they are basically similar to method embodiments, the description is relatively simple; relevant parts can be referred to the description of the method embodiments. The method 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.

[0174] The above are merely preferred embodiments 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 scope of the technology 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.

Claims

1. A method for detecting train wheel health based on a semi-supervised generative adversarial model, characterized in that, include: The rail vibration data is obtained, which is the vibration data generated by the wheels acting on the rails during train operation; wherein, the area traversed by the train is the target section, the length L of the target section is determined by the wheel radius R of the train, and the length L is greater than or equal to the wheel rolling circumference; The rail vibration data is divided into rail vibration data corresponding to each wheel, and the rail vibration data corresponding to each wheel is transformed to obtain a time-frequency map sample. The time-frequency graph samples are divided into labeled samples and unlabeled samples based on the label information; A semi-supervised deep generative adversarial network multi-label classification model is constructed and trained and tested on the time-frequency map samples to obtain a trained wheel health detection model. Input the vibration data of the rail to be detected into the trained wheel health detection model to obtain the wheel defect type; The step of inputting the rail vibration data to be detected into the trained wheel health detection model includes: using the same vibration data segmentation and time-frequency map conversion steps as in the training process to obtain the time-frequency map to be detected, and inputting the time-frequency map to be detected into the trained wheel health detection model. The semi-supervised deep generative adversarial network multi-label classification model consists of a generator, an unsupervised discriminator, and a supervised discriminator. The generator comprises, in sequence, a random noise input layer, a fully connected layer, an activation layer, a reshape layer, a transposed convolutional layer, an activation layer, a transposed convolutional layer, an activation layer, and another transposed convolutional layer, wherein the activation layer uses LeakyReLU as the activation function. ； In the formula, x represents the output of each layer. Leakage coefficient; The generator is used to take a random noise signal of a specified dimension as input and output an image of a specified size. The supervised discriminator comprises, in sequence, a labeled sample input layer, a convolutional layer, an activation layer, another convolutional layer, an activation layer, a convolutional layer, an activation layer, a flattening layer, a fully connected layer, and an output layer. The activation layer uses LeakyReLU as the activation function, the fully connected layer uses dropout to reduce overfitting, and the output layer uses the sigmoid function to activate each node. ； The supervised discriminator takes labeled samples as input and outputs predicted sample labels; The unsupervised discriminator comprises an unlabeled sample input layer, a convolutional layer, an activation layer, another convolutional layer, an activation layer, a convolutional layer, an activation layer, a flattening layer, a fully connected layer, and an output layer, connected in sequence. The activation layer uses LeakyReLU as the activation function, the fully connected layer uses dropout to reduce overfitting, and the output layer uses the activation function shown to activate each node. ； In the formula, K represents the number of categories, l k (x) represents the output value predicted by the network for each category; The unsupervised discriminator takes as input unlabeled samples and the image output by the generator, and outputs either true or false. The step of dividing the rail vibration data into rail vibration data corresponding to each wheel includes: Wheel-rail force sensors are installed at the start and end of the train passage area. These sensors are used to determine the time it takes for a wheel to pass through the area. When a train passes through the area, the wheel-rail force sensors are triggered to collect rail strain signals, and simultaneously, rail vibration acceleration sensors are triggered to collect rail vibration signals. The time it takes for each wheel to pass through the area is determined based on the strain signals. Specifically, the wheel-rail force sensors installed at the start and end of the train passage area are used to determine the start time t1 and end time t2 of each wheel passing through the train passage area, and the time Δt = t2 - t1 of each wheel is determined. The rail vibration signals collected by different rail vibration acceleration sensors are divided according to the time of each wheel's passage, and then spliced ​​together on the time axis according to the spatial position of the rail vibration acceleration sensors to obtain the time domain signal of rail vibration acceleration corresponding to each wheel, which is used as the rail vibration data corresponding to each wheel. The process of training and testing on the time-frequency map samples to obtain a trained wheel health detection model includes: The input is labeled samples, and the loss function is binary cross-entropy. The parameters are trained and updated on a supervised discriminator. ； In the formula, y is the category label 0 or 1, and p(y) represents the probability that the network predicts that the label is y; Input a random noise signal, and the generator outputs an image; Input unlabeled samples and images output by the generator, use binary cross-entropy as the loss function, train and update parameters on an unsupervised discriminator; With the parameters of the unsupervised discriminator fixed, random noise signal is input, and binary cross-entropy is used as the loss function to train and update the parameters on the generator. Training stops when the training loss no longer decreases within a preset number of iterations. The parameters from the last training iteration are saved as the optimal parameters of the model. The trained wheel health detection model is then obtained based on these optimal parameters. The process of transforming the rail vibration data corresponding to each wheel to obtain a time-frequency map sample includes: The continuous wavelet transform of the rail vibration data corresponding to each wheel is converted to the frequency domain to obtain the continuous wavelet transform of the rail vibration data in the frequency domain. : ； In the formula, a and b are the scale factor and time shift, respectively. For the Fourier transform of the vibration signal, ω is the angular frequency and i is the imaginary sign; The instantaneous angular frequency of the rail vibration data is calculated based on the continuous wavelet transform of the frequency domain rail vibration data. : ； Based on the continuous wavelet transform of the frequency domain rail vibration data and the instantaneous angular frequency of the rail vibration data, the synchronous compressed wavelet transform of the rail vibration data is obtained. : ； In the formula, δ is the Dirac function. The center frequency; The time-frequency diagram of the rail vibration data is obtained by synchronous compressed wavelet transform of the rail vibration data.

2. The train wheel health detection method based on a semi-supervised generative adversarial model according to claim 1, characterized in that, The label information includes at least wheel out of round, wheel scratches, wheel peeling off pieces, wheel cracks, wheel dents, and normal wheels.

3. The train wheel health detection method based on a semi-supervised generative adversarial model according to claim 1, characterized in that, The labeled sample must have at least one label; The label is represented by a vector of length S: ； The value at each position in the vector is 0 or 1. When the value is 1, it means that there is a labeled sample with the wheel defect type at the corresponding position. When the value is 0, it means that there is a labeled sample without the wheel defect type at the corresponding position. S is the number of defect categories.

4. A train wheel health detection system based on a semi-supervised generative adversarial model, characterized in that, The system includes: The data acquisition module is used to acquire rail vibration data, which is the vibration data generated by the wheels acting on the rails during train operation; wherein, the area traversed by the train is the target section, the length L of the target section is determined by the wheel radius R of the train, and the length L is greater than or equal to the wheel rolling circumference; The data segmentation module is used to segment the rail vibration data into rail vibration data corresponding to each wheel, and transform the rail vibration data corresponding to each wheel to obtain a time-frequency diagram sample. The sample acquisition module is used to divide the time-frequency graph samples into labeled samples and unlabeled samples according to the label information. The model training module is used to construct a semi-supervised deep generative adversarial network multi-label classification model, and to train and test it on the time-frequency map samples to obtain a trained wheel health detection model. The detection module is used to input the vibration data of the rail to be detected into the trained wheel health detection model to obtain the wheel defect type; wherein, the detection module is also used to: use the same vibration data segmentation and time-frequency map conversion steps as in the training process to obtain the time-frequency map to be detected, and input the time-frequency map to be detected into the trained wheel health detection model; The semi-supervised deep generative adversarial network multi-label classification model consists of a generator, an unsupervised discriminator, and a supervised discriminator. The generator comprises, in sequence, a random noise input layer, a fully connected layer, an activation layer, a reshape layer, a transposed convolutional layer, an activation layer, a transposed convolutional layer, an activation layer, and another transposed convolutional layer, wherein the activation layer uses LeakyReLU as the activation function. ； In the formula, x represents the output of each layer. Leakage coefficient; The generator is used to take a random noise signal of a specified dimension as input and output an image of a specified size. The supervised discriminator comprises, in sequence, a labeled sample input layer, a convolutional layer, an activation layer, another convolutional layer, an activation layer, a convolutional layer, an activation layer, a flattening layer, a fully connected layer, and an output layer. The activation layer uses LeakyReLU as the activation function, the fully connected layer uses dropout to reduce overfitting, and the output layer uses the sigmoid function to activate each node. ； The supervised discriminator takes labeled samples as input and outputs predicted sample labels; The unsupervised discriminator comprises an unlabeled sample input layer, a convolutional layer, an activation layer, another convolutional layer, an activation layer, a convolutional layer, an activation layer, a flattening layer, a fully connected layer, and an output layer, connected in sequence. The activation layer uses LeakyReLU as the activation function, the fully connected layer uses dropout to reduce overfitting, and the output layer uses the activation function shown to activate each node. ； In the formula, K represents the number of categories, l k (x) represents the output value predicted by the network for each category; The unsupervised discriminator takes as input unlabeled samples and the image output by the generator, and outputs either true or false. The step of dividing the rail vibration data into rail vibration data corresponding to each wheel includes: Wheel-rail force sensors are installed at the start and end of the train passage area. These sensors are used to determine the time it takes for a wheel to pass through the area. When a train passes through the area, the wheel-rail force sensors are triggered to collect rail strain signals, and simultaneously, rail vibration acceleration sensors are triggered to collect rail vibration signals. The time it takes for each wheel to pass through the area is determined based on the strain signals. Specifically, the wheel-rail force sensors installed at the start and end of the train passage area are used to determine the start time t1 and end time t2 of each wheel passing through the train passage area, and the time Δt = t2 - t1 of each wheel is determined. The rail vibration signals collected by different rail vibration acceleration sensors are divided according to the time of each wheel's passage, and then spliced ​​together on the time axis according to the spatial position of the rail vibration acceleration sensors to obtain the time domain signal of rail vibration acceleration corresponding to each wheel, which is used as the rail vibration data corresponding to each wheel. The process of training and testing on the time-frequency map samples to obtain a trained wheel health detection model includes: The input is labeled samples, and the loss function is binary cross-entropy. The parameters are trained and updated on a supervised discriminator. ； In the formula, y is the category label 0 or 1, and p(y) represents the probability that the network predicts that the label is y; Input a random noise signal, and the generator outputs an image; Input unlabeled samples and images output by the generator, use binary cross-entropy as the loss function, train and update parameters on an unsupervised discriminator; With the parameters of the unsupervised discriminator fixed, random noise signal is input, and binary cross-entropy is used as the loss function to train and update the parameters on the generator. Training stops when the training loss no longer decreases within a preset number of iterations. The parameters from the last training iteration are saved as the optimal parameters of the model. The trained wheel health detection model is then obtained based on these optimal parameters. The process of transforming the rail vibration data corresponding to each wheel to obtain a time-frequency map sample includes: The continuous wavelet transform of the rail vibration data corresponding to each wheel is converted to the frequency domain to obtain the continuous wavelet transform of the rail vibration data in the frequency domain. : ； In the formula, a and b are the scale factor and time shift, respectively. For the Fourier transform of the vibration signal, ω is the angular frequency and i is the imaginary sign; The instantaneous angular frequency of the rail vibration data is calculated based on the continuous wavelet transform of the frequency domain rail vibration data. : ； Based on the continuous wavelet transform of the frequency domain rail vibration data and the instantaneous angular frequency of the rail vibration data, the synchronous compressed wavelet transform of the rail vibration data is obtained. : ； In the formula, δ is the Dirac function. The center frequency; The time-frequency diagram of the rail vibration data is obtained by synchronous compressed wavelet transform of the rail vibration data.

5. A non-transitory computer-readable storage medium, characterized in that, The non-transitory computer-readable storage medium is used to store computer instructions, which, when executed by a processor, implement the train wheel health detection method based on a semi-supervised generative adversarial model as described in any one of claims 1-3.

6. An electronic device, characterized in that, include: The device includes a processor, a memory, and a computer program; wherein the processor is connected to the memory, the computer program is stored in the memory, and when the electronic device is running, the processor executes the computer program stored in the memory to cause the electronic device to execute instructions for implementing the train wheel health detection method based on a semi-supervised generative adversarial model as described in any one of claims 1-3.

Images