Wheel set anomaly detection method and system based on generative adversarial network

Abstract

The invention provides a wheel set anomaly detection method and system based on a generative adversarial network, and the method comprises the steps: obtaining wheel set data: obtaining the wheel set data; a data fusion and classification step of fusing the wheel set data to generate a data matrix and classifying the data matrix; a generative adversarial network construction step: according to the normal wheel set data set, training generator output of the generative adversarial network to obtain a pseudo normal wheel set data set, and inputting the normal wheel set data set and the pseudo normal wheel set data set into a discriminator of the generative adversarial network to train and output a similarity rate; a wheel set abnormity judgment step: inputting the real-time wheel set data into the generative adversarial network, and outputting wheel set data abnormity or wheel set data normality according to a similarity rate calculation and similarity threshold comparison result; the method has the advantages that the abnormal condition of train wheel set eccentric wear can be judged, the dependence on abnormal data samples is reduced, and the accuracy and the detection efficiency of train wheel set anomaly detection are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of rail vehicle component detection, and more specifically to a wheelset anomaly detection method and system based on a generative adversarial network. Background Art

[0002] In the field of rail vehicle component detection technology, wheelset anomaly detection has always been a research focus. Traditional wheelset anomaly detection methods mostly rely on single sensor data, such as vibration frequency, vibration amplitude or wheelset force, etc. These methods have limitations in detection accuracy and sensitivity, lack the method of using multi-source heterogeneous data fusion to improve the comprehensiveness of data, and even if multi-source data fusion technology is used, how to effectively extract key features from massive data and determine whether there is an abnormality in the wheelset is still a problem.

[0003] In addition, since the sample data that can be collected for abnormal conditions of train wheelsets are usually small, existing detection methods often rely on manually set thresholds for judgment, or build unsupervised learning neural networks for detection. The technical means adopted are to learn the characteristics of normal data and perform detection when abnormal data exists. Even if the unsupervised learning neural network does not need to provide abnormal data, it is still prone to overfitting. In addition, this type of detection still has the problems of inaccurate abnormality detection and low stability, which makes it difficult to meet the current safety requirements of wheel set abnormality detection for rail trains. Summary of the invention

[0004] In view of the shortcomings of the prior art, the object of the present invention is to provide a wheelset anomaly detection method and system based on a generative adversarial network.

[0005] To achieve the above object, the present invention provides the following technical solutions: A method for detecting anomalies of wheelsets based on a generative adversarial network comprises the following steps: A wheelset data acquisition step, obtaining a plurality of sensors to collect wheelset vibration frequency, vibration amplitude, wheelset force and wheelset image; a data fusion classification step, fusing the wheelset vibration frequency, vibration amplitude, wheelset force and wheelset image to generate a data matrix, and classifying the data matrix according to whether there is an abnormality to obtain a normal wheelset data set and an abnormal wheelset data set; A generative adversarial network construction step, according to the normal wheel pair data set, a pseudo normal wheel pair data set is obtained by training the generator output of the generative adversarial network, the normal wheel pair data set and the pseudo normal wheel pair data set are input into the discriminator of the generative adversarial network to train the output similarity, and the loss values ​​of the generator and the discriminator are respectively calculated and compared with the loss value threshold, and a generator adjustment instruction or a network construction completion instruction is output according to the comparison result; The step of judging the abnormality of wheelset is to obtain a real-time wheelset data set, input the real-time wheelset data set into a generative adversarial network to obtain a pseudo wheelset data set through the generator, input the real-time wheelset data set and the pseudo wheelset data set into a discriminator to obtain an actual similarity, compare the actual similarity with a similarity threshold, and output the wheelset data abnormality or the wheelset data normal according to the comparison result.

[0006] Furthermore, it also includes a wheelset image fusion step, obtaining a wheelset two-dimensional image and a wheelset point cloud image and inputting them into two branches of the image reconstruction branch fusion network respectively, extracting features of the wheelset point cloud image through a kernel point convolution network to obtain wheelset geometric features, extracting features of the wheelset two-dimensional image through another branch convolutional neural network to obtain wheelset two-dimensional shape features, and fusion of the wheelset geometric features and the wheelset two-dimensional shape features through a fusion layer to output a wheelset image.

[0007] Furthermore, it also includes a generator sub-network training step, randomly initializing the learning parameters of the generator, inputting the normal wheelset data set into the input layer of the generator sub-network, extracting the local image features of the wheelset image through the convolution layer of the generator sub-network, and obtaining the processed wheelset image feature vector; learning and extracting the global feature vectors of the vibration frequency, vibration amplitude and wheelset force through the fully connected layer of the generator sub-network; inputting the extracted feature vector into the splicing layer for splicing and normalization, and restoring it through the pooling layer and the deconvolution layer to obtain a pseudo-normal wheelset data set, calculating the loss value of the generator sub-network, and setting a generator loss value threshold indicating that the network training is completed, comparing the loss value of the generator sub-network and the generator loss value threshold, and when the loss value of the generator sub-network is less than the generator loss value threshold, outputting a generator sub-network construction completion instruction.

[0008] Furthermore, it also includes a discriminator sub-network training step, inputting the normal wheelset data set and the pseudo-normal wheelset data set into the input layer of the discriminator sub-network, extracting and learning the feature vectors of the wheelset image, vibration frequency, vibration amplitude and wheelset force in the normal wheelset data set and the pseudo-normal wheelset data set through the hidden layer of the discriminator sub-network, weightedly calculating the feature vectors through the output layer to obtain the similarity, and calculating the loss value of the discriminator sub-network, setting a discriminator loss value threshold indicating that the network training is completed, comparing the loss value of the discriminator sub-network with the discriminator loss threshold, and when the loss value of the discriminator sub-network is greater than the discriminator loss value threshold, outputting a discriminator adjustment instruction, otherwise, outputting a network construction completion instruction.

[0009] Furthermore, the similarity calculation formula is configured as: ; in, is the similarity function of the two data sets output by the discriminator, X is the normal data matrix, which represents the real data input to the generator. is a pseudo-normal data matrix, representing the pseudo data output by the generator; i=1, 2, 3, 4 correspond to the wheelset image, vibration frequency, vibration amplitude and wheelset force data, respectively. i is the entropy of different data types, , where p ik is the relative probability of the kth sample in the i-th data type, Represents the weight of each data in the data matrix; and are the norms of the data in the normal data matrix and the pseudo-normal data matrix, respectively, indicating the mapping of this type of data in the matrix space.

[0010] Furthermore, it also includes a generator correction step. When a generator adjustment instruction is received, the generator structure is adjusted according to the similarity calculation result and the attention mechanism is introduced to adjust the feature extraction result, and a weight decay regularization method is introduced to generalize the feature extraction of the data in the generator. The generator and the discriminator are readjusted through an alternating training strategy until the generator sub-network loss value and the discriminator sub-network loss value are respectively less than the generator loss threshold and the discriminator loss threshold.

[0011] Furthermore, the alternating training strategy, after the fixed discriminator is used to train the generator for generator correction, the fixed generator is used to train the discriminator. According to the comparison result of the discriminator subnetwork loss value and the discriminator loss threshold, when the discriminator adjustment instruction is received, the weight adjustment factor is output to adjust the discriminator similarity calculation, and the training of the generator and the discriminator is performed alternately until the construction of the generative adversarial network is completed.

[0012] Furthermore, the weight adjustment factor formula is configured as: ; Among them, α is the weight adjustment factor, noise (0,1) represents Gaussian noise with a mean of 0 and a variance of 1, which is used to balance the randomness in the generation process, S 0 is the similarity threshold, Indicates the influence of similarity calculation results on weight adjustment; by introducing the weight adjustment factor into H i In the calculation of , the adjusted similarity calculation result is obtained.

[0013] Furthermore, it also includes a generative adversarial network verification step. When a network construction completion instruction is received, a pseudo-abnormal wheel pair data set is obtained through an input generator according to the abnormal wheel pair data set, the abnormal wheel pair data set and the pseudo-abnormal wheel pair data set are input into a discriminator to obtain an abnormal data similarity, and the abnormal data similarity is compared with the similarity threshold. If the abnormal data similarity is higher than the similarity threshold, the generator is output to adjust the instruction again. If the abnormal data similarity is lower than the similarity threshold, a verification completion instruction is output.

[0014] A wheelset anomaly detection system based on a generative adversarial network includes a wheelset data acquisition module, which acquires a plurality of sensors to collect wheelset vibration frequency, vibration amplitude, wheelset force and wheelset image; A data fusion classification module fuses the wheelset vibration frequency, vibration amplitude, wheelset force and wheelset image to generate a data matrix, and classifies the data matrix into a normal wheelset data set and an abnormal wheelset data set according to whether there is an abnormality in the data matrix; A generative adversarial network construction module is configured to obtain a pseudo normal wheel pair data set by training a generator of a generative adversarial network according to the normal wheel pair data set, input the normal wheel pair data set and the pseudo normal wheel pair data set into a discriminator of a generative adversarial network to train an output similarity, compare the similarity with a similarity threshold, and output a generator adjustment instruction or a network construction completion instruction according to the comparison result; The wheelset anomaly judgment module obtains a real-time wheelset data set, inputs the real-time wheelset data set into a generative adversarial network to obtain a pseudo wheelset data set through the generator, inputs the real-time wheelset data set and the pseudo wheelset data set into a discriminator to obtain an actual similarity, compares the actual similarity with a similarity threshold, and outputs wheelset data abnormality or wheelset data normality according to the comparison result.

[0015] The beneficial effects of the present invention are as follows: by fusing multi-source heterogeneous data, data anomalies when there are abnormalities in the wheelset can be considered in multiple dimensions, thereby improving the sensitivity and accuracy of wheelset abnormality monitoring; after the generator of the generative adversarial network extracts features from the normal wheelset data and learns to generate pseudo-normal wheelset data, the discriminator judges the similarity between the normal wheelset data and the pseudo-normal wheelset data input therein, and the generator can be structurally adjusted to output more similar pseudo-normal wheelset data; after the generator is fixed, the discriminator is adjusted according to the similarity, thereby avoiding overfitting of the generative adversarial network; by inputting the actually collected wheelset data into the trained generative adversarial network, if there is an abnormality in the wheelset data input at this time, the pseudo-wheelet data generated by the generator will be significantly different from the input wheelset data; the discriminator judges the similarity between the two, and when the similarity is lower than the set threshold, it can be judged and output that there is an abnormality in the wheelset at this time, thereby helping to timely discover the wheelset abnormality and take measures to avoid the deterioration of the problem, thereby improving the safety and stability of train travel. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 It is the overall flow chart of the present invention; Figure 2 is a flow chart of wheelset image fusion in the present invention; Figure 3 It is a flowchart of generating adversarial network verification in the present invention. DETAILED DESCRIPTION

[0017] The present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. The same parts are represented by the same reference numerals. It should be noted that the words "front", "rear", "left", "right", "upper" and "lower" used in the following description refer to directions in the accompanying drawings, and the words "bottom surface" and "top surface", "inner" and "outer" refer to directions toward or away from the geometric center of a specific component, respectively.

[0018] Since the sample data that can be collected for abnormal situations of train wheelsets are usually relatively small, existing detection methods often rely on manually set thresholds for judgment, or construct unsupervised learning neural networks for detection. The technical means adopted is to learn the characteristics of normal data and perform detection when abnormal data exists. Even if the unsupervised learning neural network does not need to provide abnormal data, there is still a problem of overfitting. In addition, this type of detection still has the problems of inaccurate abnormality detection and low stability, which makes it difficult to meet the current safety requirements of wheelset abnormality detection for rail trains. Therefore, the present invention designs such a wheelset abnormality detection method for rail trains.

[0019] like Figure 1 As shown, the following steps are included: The steps for acquiring wheelset data are as follows: a high-precision acceleration sensor is used to obtain the acceleration change information of the wheelset and analyze it to obtain the wheelset vibration frequency and vibration amplitude data; a high-sensitivity pressure sensor is used to obtain the force value at the contact point between the wheelset and the track; a low-noise camera is used to ensure that clear and complete images of the train track and wheels (wheelset images) can be captured even in complex lighting environments. The collected data is then cleaned and sorted, and pre-processed such as removing noise and outliers to facilitate the accuracy of subsequent data fusion.

[0020] A wheelset database is preset, which stores a number of wheelset vibration frequencies, vibration amplitudes, wheelset forces and wheelset image data.

[0021] In the data fusion classification step, four different types of wheelset data are synthesized into a data matrix: the wheelset image is processed with a CNN network for feature extraction, and then the wheelset vibration frequency, vibration amplitude and wheelset force are normalized to ensure that data of different dimensions can be fused at the same scale. Different data sequences are generated according to their data types, and Transformer is used to extract features from the data sequence. All feature extraction results are then constructed to generate a high-dimensional data matrix according to the collaborative feature representation strategy. Then, according to the abnormal situation of the wheelset data in the data matrix, it is divided into a normal wheelset data set and an abnormal wheelset data set. The normal wheelset data set contains a large amount of data, while the abnormal wheelset data set only needs to contain far less wheelset data than the normal wheelset data.

[0022] Collaborative feature representation strategy: After extracting the features of wheelset image, wheelset vibration frequency, vibration amplitude and wheelset force, all feature vectors are arranged in sequence and concatenated to form a long vector, and the data in the data matrix is ​​reduced in dimension through linear discriminant analysis to reduce the complexity of the data matrix as much as possible while retaining most of the useful information; for wheelset images, the extracted feature vectors include edge, texture, shape and other information in the image; for wheelset vibration frequency, vibration amplitude and force value, the extracted feature vectors include the main frequency, secondary frequency, amplitude peak and force value distribution curve of the vibration.

[0023] The generative adversarial network construction step is to obtain a pseudo normal wheel pair data set by training the generator output of the generative adversarial network according to the normal wheel pair data set, input the normal wheel pair data set and the pseudo normal wheel pair data set into the discriminator training output similarity of the generative adversarial network, and respectively calculate the loss values ​​of the generator and the discriminator and compare them with the loss value threshold, and output a generator adjustment instruction or a network construction completion instruction according to the comparison result; wherein when the generator sub-network loss value is higher than the generator loss value threshold, the generator adjustment instruction is output according to the comparison result, and the generator is adjusted so that its output data is closer to the real input data, and after a certain number of rounds (i.e., a preset training round) of generator adjustment or the generator sub-network loss value is lower than the generator loss value threshold, the generator is fixed and the discriminator is adjusted according to the loss value calculation result, so that the discriminator calculates the similarity more in line with the data situation, and the discriminator parameters are fixed again, and the generator is adjusted again according to the comparison result of the discriminator sub-network loss value and the discriminator loss value threshold until the generative adversarial network training is completed.

[0024] Specifically, the generator subnetwork construction and training steps are as follows: design the generator subnetwork network structure, select ReLU as the activation function, introduce nonlinear characteristics, and set the dimension and format of the output layer to be consistent with the wheelset data matrix in the present invention. The weights of the parameters of each data type in the pre-trained generator are set, the generator learning rate and the number of training rounds are set, the discriminator parameters are fixed, the generator parameters are updated by optimizing the generator loss function, the normal wheelset data set is input into the input layer of the generator sub-network, the local image features of the wheelset image are extracted through the convolution layer of the generator sub-network, and the processed wheelset image feature vector is obtained; the global feature vectors of the vibration frequency, vibration amplitude and wheelset force are learned and extracted through the fully connected layer of the generator sub-network; the extracted feature vectors are input into the splicing layer for splicing and normalization, and restored through the pooling layer and the deconvolution layer to obtain a pseudo-normal wheelset data set, the loss value of the generator sub-network is calculated, and the generator loss value threshold indicating the completion of network training is set, the loss value of the generator sub-network is compared with the generator loss value threshold, and when the loss value of the generator sub-network is less than the generator loss value threshold, the generator sub-network construction completion instruction is output.

[0025] Specifically, the discriminator sub-network construction and training steps are as follows: design the discriminator sub-network network structure, use the Sigmoid function in the discriminator output layer, limit the output value between 0 and 1, use the binary cross entropy loss function as the measure of the discriminator prediction accuracy, when the generator reaches the training round or the similarity calculation result reaches the similarity threshold condition, fix the generator parameters, update the discriminator parameters by optimizing the discriminator loss function, input the normal wheel pair data set and the pseudo-normal wheel pair data set into the input layer of the discriminator sub-network, and use the hidden layer of the discriminator sub-network to calculate the discriminator parameters. The hidden layer extracts and learns the feature vectors of the wheelset image, vibration frequency, vibration amplitude and wheelset force in the normal wheelset data set and the pseudo-normal wheelset data set, and the feature vectors are weightedly calculated through the output layer to obtain the similarity, and the loss value of the discriminator sub-network is calculated, and a discriminator loss value threshold indicating that the network training is completed is set, and the loss value of the discriminator sub-network is compared with the discriminator loss threshold. When the loss value of the discriminator sub-network is greater than the discriminator loss value threshold, a discriminator adjustment instruction is output, otherwise, a network construction completion instruction is output.

[0026] The step of judging the abnormality of wheelset is to obtain a real-time wheelset data set, input the real-time wheelset data set into a generative adversarial network, obtain a pseudo wheelset data set through the generator, and input the real-time wheelset data set and the pseudo wheelset data set into a discriminator to obtain an actual similarity rate. Specifically, when the real-time wheelset data set is obtained, the real-time wheelset data matrix is ​​input into the generator of the generative adversarial network, and four different data types in the wheelset data matrix are extracted respectively. The extracted features are judged according to the data type and a pseudo wheelset data matrix is ​​generated. The real input wheelset data matrix and the pseudo wheelset data matrix are input into the branch of the discriminator subnetwork, and the data features in the wheelset data matrix are calculated by setting parameters through the discriminator subnetwork. Similarity, compare the calculated similarity with the similarity threshold, and output the wheel pair data abnormality or normal wheel pair data according to the comparison result. The setting of the similarity threshold represents the minimum similarity that the pseudo wheel pair data matrix generated after the normal wheel pair data matrix is ​​input into the generator needs to achieve. Since the generator is constructed and optimized according to the characteristics of the normal wheel pair data matrix, the pseudo data output by the generator is closer to the real data only when normal data is input, that is, the similarity is higher than the similarity threshold. If there is abnormal data in the input wheel pair data matrix at this time, there is a large gap between the pseudo data matrix generated by the generator and the real abnormal data input. At this time, the similarity is lower than the similarity threshold, and it is judged that the wheel pair data is abnormal.

[0027] Specifically, the similarity calculation formula is configured as: ; in, is the similarity function of the two data sets output by the discriminator, X is the normal data matrix, which represents the real data input to the generator. is a pseudo-normal data matrix, representing the pseudo data output by the generator; i=1, 2, 3, 4 correspond to the wheelset image, vibration frequency, vibration amplitude and wheelset force data, respectively. i is the entropy of different data types, , where p ik is the relative probability of the kth sample in the i-th data type, Represents the weight of each data in the data matrix; and are the norms of the data in the normal data matrix and the pseudo-normal data matrix, respectively, indicating the mapping of this type of data in the matrix space.

[0028] X is the real data matrix of the input generator, including the wheel pair image X 1 , vibration frequency X 2 , vibration amplitude X 3 and wheelset force X 4 In the generative adversarial network construction step, X represents the normal data matrix of the input generator; in the generative adversarial network verification step, X can represent the input abnormal data matrix; in the actual wheelset abnormality detection step, X can represent the actual collected wheelset data matrix to be tested; Pseudo data matrix generated for the generator, including pseudo wheelset images , pseudo vibration frequency , pseudo vibration amplitude and pseudo wheelset force . Normally, when the train wheelset has abnormal wear, it is partial wear. At this time, the train tilts. Therefore, the partial wear of the wheelset can be more directly reflected through the wheelset image. At this time, the vibration frequency and amplitude of the train wheelset are significantly different from the normal train vibration. The force value at the contact point between the wheelset and the train track is unbalanced. Each data has different performance in the wheelset abnormality. The weight of the above data in the wheelset abnormality gradually decreases. Therefore, different similarity calculation weights are set for each data. When calculating the similarity, more complex functions are involved. The basic function of similarity calculation is based on cosine similarity. The exponential function is introduced to represent the exponential growth of the data type similarity when it is close to the truth (that is, the similarity is 1). The angle of cosine similarity is the angle between the real data matrix and the pseudo data matrix. It is obtained by comparing the product of the two data matrices with the product of the norms of the two data matrices. Each different data type has different weight parameters. The final similarity calculation formula is obtained by multiplying the weight and the basic function of similarity calculation.

[0029] Specifically, it also includes a generator correction step. When a generator adjustment instruction is received, the generator structure is adjusted according to the similarity calculation result and the attention mechanism is introduced to adjust the feature extraction result. The weight decay regularization method is introduced to generalize the feature extraction of the data in the generator, and the generator and discriminator are readjusted through an alternating training strategy until the generator sub-network loss value and the discriminator sub-network loss value are respectively less than the generator loss threshold and the discriminator loss threshold. The specific steps are: simulating the neural network of visual attention through the attention mechanism, optimizing the feature extraction process, enabling the generator to focus on the key information of the input data matrix, and improving the quality of generated data. The weight decay regularization method increases the complexity of the parameters by adding the square of the weight to the loss function, which can increase the network stability in the process of generalized feature extraction and prevent the network from overfitting.

[0030] Specifically, the alternating training strategy trains the generator with a fixed discriminator to perform generator correction, and then fixes the generator to train the discriminator. The discriminator weight adjustment factor is output by comparing the discriminator subnetwork loss value and the discriminator loss threshold to adjust the discriminator similarity calculation. The training of the generator and the discriminator is performed alternately until the construction of the generative adversarial network is completed. By fixing the parameters of the generator and the discriminator in turn and optimizing the parameters of the other branch network, the data generation ability and data authenticity judgment ability of the generative adversarial network can be optimized.

[0031] Specifically, the weight adjustment factor formula is configured as follows:

[0032] Among them, α is the weight adjustment factor, noise (0,1) represents Gaussian noise with a mean of 0 and a variance of 1, which is used to balance the randomness in the generation process, S 0 is the similarity threshold, Indicates the influence of similarity calculation results on weight adjustment; by introducing the weight adjustment factor into H i In the calculation of , the adjusted similarity calculation result is obtained.

[0033] when When the discriminator outputs the similarity function of the two data sets, the similarity calculation result adjusted by the weight adjustment factor is: At this time, by introducing the weight adjustment factor, when the similarity is lower than the similarity threshold, the weight adjustment factor is adjusted. Since the data matrix input to train the generative adversarial network is a normal wheel pair data matrix, the discriminator judges that the similarity is lower than the threshold and the value of the calculated similarity should be increased. Gradually adjusting the discriminator can more accurately identify the similarity between the data generated by the generator and the real data.

[0034] The advantages of the present invention are: by fusing multi-source heterogeneous data, data anomalies when there are abnormalities in the wheelset can be considered in multiple dimensions, the sensitivity and accuracy of wheelset abnormality monitoring can be improved, after the generator of the generative adversarial network extracts features from the normal wheelset data and learns to generate pseudo-normal wheelset data, the similarity between the normal wheelset data and the pseudo-normal wheelset data input therein is judged by the discriminator, the generator can be structurally adjusted to output more similar pseudo-normal wheelset data, and the discriminator is adjusted according to the similarity after the generator is fixed, so that overfitting of the generative adversarial network can be avoided, and the wheelset data actually collected is input into the trained generative adversarial network. If there is an abnormality in the wheelset data input at this time, the pseudo-wheelet data generated by the generator will be greatly different from the input wheelset data, and the similarity between the two is judged by the discriminator. When the similarity is lower than the set threshold, it can be judged and output that there is an abnormality in the wheelset at this time, which is helpful to timely discover the wheelset abnormality and take measures to avoid the deterioration of the problem, thereby improving the safety and stability of train driving.

[0035] like Figure 2 As shown, it also includes a wheelset image fusion step, obtaining a wheelset two-dimensional image and a wheelset point cloud image and inputting them into two branch convolutional neural networks of an image reconstruction branch fusion network respectively, passing the wheelset point cloud image through a branch convolutional neural network through structures such as a convolutional layer, a pooling layer, and a fully connected layer, extracting features such as geometric shapes, surface normals, and depth information in the wheelset point cloud image, passing the wheelset two-dimensional image through another branch convolutional neural network through structures such as a convolutional layer, a pooling layer, and a fully connected layer, extracting features such as shape, texture, and edges in the wheelset two-dimensional image, and merging the wheelset three-dimensional shape features and the wheelset two-dimensional shape features through a fusion layer and outputting the wheelset image at the dry network output.

[0036] Since the generative adversarial network is constructed by training the normal wheel pair data matrix, in order to verify that the generative adversarial network can generate and judge different performances for abnormal wheel pair data, such as Figure 3As shown, it also includes a generative adversarial network verification step. When a network construction completion instruction is received, a trained generator is used to take an abnormal wheel pair data set as input. At this time, the basic function of the generator to generate data is to simulate the distribution of real normal data. Therefore, when the input is an abnormal data matrix, the simulation ability of the generator is poor, and there is a significant difference between the generated data and the real input data. According to the abnormal wheel pair data set, a pseudo abnormal wheel pair data set is obtained by inputting the generator, and the abnormal wheel pair data set and the pseudo abnormal wheel pair data set are input into the discriminator to obtain the abnormal data similarity. The abnormal data similarity is compared with the similarity threshold. If the If the similarity rate of abnormal data is higher than the similarity threshold, it means that the generated pseudo-abnormal wheel pair data set is similar to the real data set. However, since the goal of the training generator is to simulate the normal real data characteristics, it is impossible to perform a high-similarity simulation when the abnormal data is input. Therefore, it is considered that the generator of the generative adversarial network needs to be re-optimized to distinguish the data characteristics of the normal data matrix and the abnormal data matrix. At this time, the generator is output to adjust the instruction again. If the similarity rate of the abnormal data is lower than the similarity threshold, it means that there is a large difference between the generated pseudo-abnormal wheel pair data set and the real data set. At this time, the generator of the generative adversarial network can achieve the training goal, and the verification completion instruction is output.

[0037] Correspondingly, a system for monitoring the abnormality of wheelsets of rail trains is provided, including a wheelset data acquisition module, which acquires historical wheelset vibration frequency, vibration amplitude, wheelset force and wheelset images collected by several sensors; A data fusion classification module fuses the historical wheelset vibration frequency, vibration amplitude, wheelset force and wheelset image to generate a data matrix, and classifies the data matrix into a normal wheelset data set and an abnormal wheelset data set according to whether there is an abnormality in the data matrix; A generative adversarial network construction module is configured to obtain a pseudo normal wheel pair data set by training a generator of a generative adversarial network according to the normal wheel pair data set, input the normal wheel pair data set and the pseudo normal wheel pair data set into a discriminator of a generative adversarial network to train an output similarity, compare the similarity with a similarity threshold, and output a generator adjustment instruction or a network construction completion instruction according to the comparison result; The wheelset anomaly judgment module obtains a real-time wheelset data set, inputs the real-time wheelset data set into a generative adversarial network to obtain a pseudo wheelset data set through the generator, inputs the real-time wheelset data set and the pseudo wheelset data set into a discriminator to obtain an actual similarity, compares the actual similarity with a similarity threshold, and outputs wheelset data abnormality or wheelset data normality according to the comparison result.

[0038] The above are only preferred embodiments of the present invention. The protection scope of the present invention is not limited to the above embodiments. All technical solutions under the concept of the present invention belong to the protection scope of the present invention. It should be pointed out that for ordinary technicians in this technical field, some improvements and modifications without departing from the principle of the present invention should also be regarded as the protection scope of the present invention.

Claims

1. A wheelset anomaly detection method based on a generative adversarial network, characterized in that: The steps include: A wheelset data acquisition step, obtaining a plurality of sensors to collect wheelset vibration frequency, vibration amplitude, wheelset force and wheelset image; a data fusion classification step, fusing the wheelset vibration frequency, vibration amplitude, wheelset force and wheelset image to generate a data matrix, and classifying the data matrix according to whether there is an abnormality to obtain a normal wheelset data set and an abnormal wheelset data set; A generative adversarial network construction step, according to the normal wheel pair data set, a pseudo normal wheel pair data set is obtained by training the generator output of the generative adversarial network, the normal wheel pair data set and the pseudo normal wheel pair data set are input into the discriminator of the generative adversarial network to train the output similarity, and the loss values ​​of the generator and the discriminator are respectively calculated and compared with the loss value threshold, and a generative adversarial network adjustment instruction or a network construction completion instruction is output according to the comparison result; The step of judging the abnormality of wheelset is to obtain a real-time wheelset data set, input the real-time wheelset data set into a generative adversarial network to obtain a pseudo wheelset data set through the generator, input the real-time wheelset data set and the pseudo wheelset data set into a discriminator to obtain an actual similarity, compare the actual similarity with a similarity threshold, and output the wheelset data abnormality or the wheelset data normal according to the comparison result.

2. According to claim 1, a method for detecting anomalies of wheelsets based on a generative adversarial network is characterized in that: The method also includes a wheelset image fusion step, obtaining a wheelset two-dimensional image and a wheelset point cloud image and inputting them into two branches of an image reconstruction branch fusion network respectively, extracting features of the wheelset point cloud image through a kernel point convolution network to obtain wheelset geometric features, extracting features of the wheelset two-dimensional image through another branch convolutional neural network to obtain wheelset two-dimensional shape features, and fusion of the wheelset geometric features and the wheelset two-dimensional shape features through a fusion layer to output a wheelset image.

3. According to claim 1, a method for detecting anomalies of wheelsets based on a generative adversarial network is characterized in that: It also includes a generator sub-network training step, randomly initializing the learning parameters of the generator, inputting the normal wheelset data set into the input layer of the generator sub-network, extracting the local image features of the wheelset image through the convolution layer of the generator sub-network, and obtaining the processed wheelset image feature vector; learning and extracting the global feature vectors of the vibration frequency, vibration amplitude and wheelset force through the fully connected layer of the generator sub-network; inputting the extracted feature vector into the splicing layer for splicing and normalization, and restoring it through the pooling layer and the deconvolution layer to obtain a pseudo-normal wheelset data set, calculating the loss value of the generator sub-network, and setting a generator loss value threshold indicating that the network training is completed, comparing the loss value of the generator sub-network with the generator loss value threshold, and when the loss value of the generator sub-network is less than the generator loss value threshold, outputting a generator sub-network construction completion instruction, otherwise outputting a generator adjustment instruction.

4. According to claim 1, a method for detecting anomalies of wheelsets based on a generative adversarial network is characterized in that: The method also includes a discriminator sub-network training step, wherein the normal wheelset data set and the pseudo-normal wheelset data set are input into the input layer of the discriminator sub-network, and the characteristic vectors of the wheelset image, vibration frequency, vibration amplitude and wheelset force in the normal wheelset data set and the pseudo-normal wheelset data set are extracted and learned through the hidden layer of the discriminator sub-network, and the characteristic vectors are weightedly calculated through the output layer to obtain the similarity, and the loss value of the discriminator sub-network is calculated, and a discriminator loss value threshold indicating that the network training is completed is set, and the loss value of the discriminator sub-network is compared with the discriminator loss threshold. When the loss value of the discriminator sub-network is greater than the discriminator loss value threshold, a discriminator adjustment instruction is output, otherwise, a network construction completion instruction is output.

5. The method for detecting anomalies of wheelsets based on a generative adversarial network according to claim 1, characterized in that: The similarity calculation formula is configured as: ； in, is the similarity function of the two data sets output by the discriminator, X is the normal data matrix, which represents the real data input to the generator. is a pseudo-normal data matrix, representing the pseudo data output by the generator; i=1, 2, 3, 4 correspond to the wheelset image, vibration frequency, vibration amplitude and wheelset force data, respectively. i is the entropy of different data types, , where p ik is the relative probability of the kth sample in the i-th data type, Represents the weight of each data in the data matrix; and are the norms of the data in the normal data matrix and the pseudo-normal data matrix, respectively, indicating the mapping of this type of data in the matrix space.

6. The method for detecting anomalies of wheelsets based on a generative adversarial network according to claim 3, characterized in that: It also includes a generator correction step. When a generator adjustment instruction is received, the generator structure is adjusted according to the similarity calculation result and the attention mechanism is introduced to adjust the feature extraction result. The weight decay regularization method is introduced to generalize the feature extraction of the data in the generator, and the generator and discriminator are readjusted through an alternating training strategy until the generator sub-network loss value and the discriminator sub-network loss value are respectively less than the generator loss threshold and the discriminator loss threshold.

7. The method for detecting anomalies of wheelsets based on a generative adversarial network according to claim 6, characterized in that: The alternating training strategy is to train the generator with a fixed discriminator to perform generator correction, and then train the discriminator with a fixed generator. According to the comparison result between the discriminator subnetwork loss value and the discriminator loss threshold, when a discriminator adjustment instruction is received, a weight adjustment factor is output to adjust the discriminator similarity calculation, and the training of the generator and the discriminator is performed alternately until the construction of the generative adversarial network is completed.

8. The method for detecting anomalies of wheelsets based on a generative adversarial network according to claim 7, characterized in that: The weight adjustment factor formula is configured as follows: ； Among them, α is the weight adjustment factor, noise (0,1) represents Gaussian noise with a mean of 0 and a variance of 1, which is used to balance the randomness in the generation process, S0 is the similarity threshold, Indicates the influence of similarity calculation results on weight adjustment; by introducing the weight adjustment factor into H i In the calculation of , the adjusted similarity calculation result is obtained.

9. The method for detecting wheelset anomalies based on a generative adversarial network according to claim 1, characterized in that: The method also includes a generative adversarial network verification step. When a network construction completion instruction is received, a pseudo-abnormal wheel pair data set is obtained through an input generator according to the abnormal wheel pair data set, the abnormal wheel pair data set and the pseudo-abnormal wheel pair data set are input into a discriminator to obtain an abnormal data similarity, and the abnormal data similarity is compared with the similarity threshold. If the abnormal data similarity is higher than the similarity threshold, an output generator readjustment instruction is given; if the abnormal data similarity is lower than the similarity threshold, a verification completion instruction is output.

10. A wheelset anomaly detection system based on a generative adversarial network, characterized in that: include The wheelset data acquisition module acquires the wheelset vibration frequency, vibration amplitude, wheelset force and wheelset image collected by several sensors; A data fusion classification module fuses the wheelset vibration frequency, vibration amplitude, wheelset force and wheelset image to generate a data matrix, and classifies the data matrix into a normal wheelset data set and an abnormal wheelset data set according to whether there is an abnormality in the data matrix; A generative adversarial network construction module is configured to obtain a pseudo normal wheel pair data set by training a generator of a generative adversarial network according to the normal wheel pair data set, input the normal wheel pair data set and the pseudo normal wheel pair data set into a discriminator of a generative adversarial network to train an output similarity, compare the similarity with a similarity threshold, and output a generator adjustment instruction or a network construction completion instruction according to the comparison result; The wheelset anomaly judgment module obtains a real-time wheelset data set, inputs the real-time wheelset data set into a generative adversarial network to obtain a pseudo wheelset data set through the generator, inputs the real-time wheelset data set and the pseudo wheelset data set into a discriminator to obtain an actual similarity, compares the actual similarity with a similarity threshold, and outputs wheelset data abnormality or wheelset data normality according to the comparison result.

Images