Wheel flat fault identification method and system, terminal equipment and medium

Abstract

The invention provides a wheel flat fault identification method and system, terminal equipment and a medium. The method comprises the following steps: acquiring a wheel flat fault sample data set; for each wheel flat fault sample, converting the vertical vibration acceleration signal based on the time domain into a spatial domain, and performing time-frequency processing on the converted spatial domain sample to obtain a two-dimensional time-frequency image of the wheel flat fault; constructing a wheel flat fault recognition model used for wheel flat fault recognition, and training the wheel flat fault recognition model by using the two-dimensional time-frequency image until the loss value of the wheel flat fault recognition model is smaller than a preset loss threshold value, thereby obtaining the trained wheel flat fault recognition model; and inputting the collected vertical vibration acceleration signal to be detected into the trained wheel flat fault recognition model to obtain a wheel flat fault recognition result. According to the invention, the accuracy of wheel flat fault identification can be improved.

Description

Technical Field

[0001] The invention belongs to the field of traffic technology, and specifically relates to a wheel flat fault identification method, system, terminal equipment and medium. Background Art

[0002] Wheel flats usually refer to local abrasions on the wheels caused by abnormal braking or wheel idling under low adhesion between the train wheel and rail, which will cause intermittent impacts on the wheel-rail contact interface. The impact load caused by the flats is several times or even dozens of times the normal static load, thereby exacerbating the dynamic response of the wheel and rail, and seriously threatening the safety of subway vehicle operation. Long-term service in this condition will shorten the service life of key vehicle components and rail infrastructure, increase maintenance pressure, and increase operating and maintenance costs. At the same time, wheel flats will also cause wheel-rail impact noise, affecting passenger comfort and satisfaction.

[0003] At present, wheel flat fault identification mainly relies on manual regular inspection, which is cumbersome and inefficient. At the same time, manual inspection requires a lot of experience and is affected by human subjective factors, so the accuracy of fault identification is low. For some vehicles with long operating time and large passenger flow (such as subways and high-speed trains), it is even more inconvenient to conduct regular manual inspection.

[0004] Therefore, how to quickly and accurately identify wheel flat faults is a technical problem that needs to be solved urgently. Summary of the invention

[0005] The technical problem to be solved by the present invention is to provide a wheel flat fault identification method, system, terminal equipment and medium to achieve accurate identification of wheel flat fault.

[0006] In a first aspect, the present invention provides a method for identifying a wheel flat fault, the method comprising the following steps:

[0007] Acquire a wheel flat fault sample data set; the wheel flat fault sample data set includes multiple wheel flat lengths and vertical vibration acceleration signals generated by the vibration of the vehicle frame at each wheel flat length, and the multiple wheel flat lengths correspond to the vertical vibration acceleration signals one by one in the time domain;

[0008] For each wheel flat fault sample, the vertical vibration acceleration signal based on the time domain is converted to the space domain, and the converted space domain samples are subjected to time-frequency processing to obtain a two-dimensional time-frequency image of the wheel flat fault; the space domain samples are used to describe the spatial relationship between the wheel flat and the vertical vibration acceleration signal;

[0009] A wheel flat fault recognition model for wheel flat fault recognition is constructed, and the wheel flat fault recognition model is trained using a two-dimensional time-frequency image until the loss value of the wheel flat fault recognition model is less than a preset loss threshold, thereby obtaining a trained wheel flat fault recognition model; the wheel flat fault recognition model includes a channel attention module for pooling the two-dimensional time-frequency image to generate a channel attention feature map, a spatial feature module for multiplying the channel attention feature map and the two-dimensional time-frequency image element by element to generate a spatial attention feature map, and a recognition module for obtaining a wheel flat fault recognition result according to the spatial attention feature map;

[0010] The collected vertical vibration acceleration signal to be tested is input into the trained wheel flat fault recognition model to obtain the wheel flat fault recognition result.

[0011] Optionally, the channel attention module, the spatial feature module, and the recognition module are connected in sequence;

[0012] The channel attention module includes a first maximum pooling layer, a first global pooling layer, a multi-layer perceptron network and a first activation layer; wherein the input ends of the first maximum pooling layer and the first global pooling layer receive the two-dimensional time-frequency image, the input end of the multi-layer perceptron network is respectively connected to the output end of the first maximum pooling layer and the output end of the first global pooling layer, the output end of the multi-layer perceptron network is connected to the input end of the first activation layer, and the output end of the first activation layer outputs the channel attention feature map;

[0013] The spatial feature module includes a merging layer, a splicing layer, a second maximum pooling layer, a second global pooling layer, a convolution layer and a second activation layer; wherein the merging layer is used to perform element-by-element multiplication of the channel attention feature map and the two-dimensional time-frequency image to obtain a new feature map, the input ends of the second maximum pooling layer and the second global pooling layer receive the new feature map, the input end of the splicing layer is respectively connected to the output end of the second maximum pooling layer and the output end of the second global pooling layer, the output end of the splicing layer is connected to the input end of the convolution layer, the output end of the convolution layer is connected to the input end of the second activation layer, and the output end of the second activation layer outputs the spatial attention feature map.

[0014] Optionally, the expression of the channel attention feature map is as follows:

[0015]

[0016] Among them, M c (F) represents the channel attention feature map, σ(·) represents the sigmoid activation function, W 0 ,W 1represents the shared weight parameter in the multilayer perceptron network, MLP(·) represents the multilayer perceptron network, AvgPool(·) represents global pooling, MaxPool(·) represents maximum pooling, F represents a two-dimensional time-frequency image, represents the local feature map, represents the global eigenvector.

[0017] Optionally, the expression of the spatial attention feature map is as follows:

[0018]

[0019] Among them, M s (F ′ ) represents the spatial attention feature map, f 7×7 represents the convolution operation with a kernel size of 7×7, (·;·) represents concatenation, and F ′ Represents the new feature map obtained by element-by-element multiplication of the channel attention feature map and the two-dimensional time-frequency image. Indicates F ′ The global pooling result is Indicates F ′ The maximum pooling result.

[0020] Optionally, the wheel flat scar fault identification result is a mild flat scar, a moderate flat scar or a severe flat scar; wherein, the flat scar length corresponding to a mild flat scar is less than or equal to 25 mm, the flat scar length corresponding to a moderate flat scar is greater than 25 mm and less than or equal to 40 mm, and the flat scar length corresponding to a severe flat scar is greater than 40 mm and less than or equal to 60 mm.

[0021] Optionally, it also includes a fully connected layer for enhancing the nonlinear feature extraction capability and a dropout layer for overfitting of the wheel flat fault recognition model;

[0022] The input of the fully connected layer is connected to the output of the concatenation layer, and the output of the fully connected layer is connected to the input of the convolution layer;

[0023] The input of the dropout layer is connected to the output of the second activation layer, and the output of the dropout layer is connected to the input of the recognition module.

[0024] Optionally, the loss value is the cross entropy loss value.

[0025] In a second aspect, the present invention provides a wheel flat fault identification system, comprising:

[0026] A data acquisition module is used to acquire a wheel flat fault sample data set; the wheel flat fault sample data set includes a plurality of wheel flat lengths and a vertical vibration acceleration signal generated by the vibration of the vehicle frame at each wheel flat length, and the plurality of wheel flat lengths and the vertical vibration acceleration signal correspond one to one in the time domain;

[0027] The preprocessing module is used to convert the vertical vibration acceleration signal based on the time domain into the space domain for each wheel flat fault sample, and perform time-frequency processing on the converted space domain samples to obtain a two-dimensional time-frequency image of the wheel flat fault; the space domain samples are used to describe the spatial relationship between the wheel flat and the vertical vibration acceleration signal;

[0028] A model training module is used to construct a wheel flat fault recognition model for wheel flat fault recognition, and to train the wheel flat fault recognition model using a two-dimensional time-frequency image until the loss value of the wheel flat fault recognition model is less than a preset loss threshold, thereby obtaining a trained wheel flat fault recognition model; the wheel flat fault recognition model includes a channel attention module for pooling the two-dimensional time-frequency image to generate a channel attention feature map, a spatial feature module for multiplying the channel attention feature map and the two-dimensional time-frequency image element by element to generate a spatial attention feature map, and a recognition module for obtaining a wheel flat fault recognition result according to the spatial attention feature map;

[0029] The wheel flat identification module is used to input the collected vertical vibration acceleration signal to be tested into the trained wheel flat fault identification model to obtain the wheel flat fault identification result.

[0030] In a third aspect, the present invention provides a terminal device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the above-mentioned wheel flat fault identification method when executing the computer program.

[0031] In a fourth aspect, the present invention provides a computer-readable storage medium storing a computer program, which implements the above-mentioned wheel flat fault identification method when executed by a processor.

[0032] The beneficial effects of the present invention are:

[0033] The present invention discloses a wheel flat fault identification method. The wheel flat fault identification model constructed by the method includes a channel attention module and a spatial feature module, which can extract and combine the feature information of the vertical vibration acceleration signal and the wheel flat in two dimensions of channel and space, accurately mine the correlation features between the vertical vibration acceleration signal and the wheel flat, which is beneficial to improving the accuracy of wheel flat fault identification; in the spatial feature module, the channel attention feature map and the two-dimensional time-frequency image are multiplied element by element, which can realize adaptive feature optimization and dynamically adjust the feature weight, so that the wheel flat fault identification model can mine the important features of the wheel flat affecting the vertical vibration acceleration signal, reduce noise interference, and improve the accuracy of wheel flat fault identification; in addition, the method does not rely on people's subjective experience judgment, can reduce the influence of human factors, and further improve the accuracy of wheel flat fault identification. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 This is a flow chart of a method for identifying a wheel flat fault in one embodiment of the present application;

[0035] Figure 2 This is a structural diagram of a wheel flat fault identification model in one embodiment of the present application;

[0036] Figure 3 This is a structural diagram of a channel attention module in one embodiment of the present application;

[0037] Figure 4 This is a structural diagram of a spatial feature module in another embodiment of the present application;

[0038] Figure 5 This is a comparison diagram of recognition effects under different noise conditions in one embodiment of the present application;

[0039] Figure 6a This is a comparison diagram of the recognition effect of the LeNet-5 model under the noise condition of SNR=2dB in one implementation mode of the present application;

[0040] Figure 6b This is a comparison chart of the recognition effect of the VGG-16 model under the noise condition of SNR=2dB in one implementation mode of the present application;

[0041] Figure 6c This is a comparison diagram of the recognition effect of the ResNet model under the noise condition of SNR=2dB in one implementation mode of the present application;

[0042] Figure 6d This is a comparison diagram of the recognition effect of the wheel flat fault recognition method provided by the present invention under the noise condition of SNR=2dB in one embodiment of the present application;

[0043] Figure 7a This is the confusion matrix result of the LeNet-5 model under the noise condition of SNR=2dB in another implementation manner of the present application;

[0044] Figure 7b This is the confusion matrix result of the VGG-16 model under the noise condition of SNR=2dB in another implementation manner of the present application;

[0045] Figure 7c This is the confusion matrix result of the ResNet model under the noise condition of SNR=2dB in another implementation manner of the present application;

[0046] Figure 7d This is the confusion matrix result of the wheel flat fault identification method provided by the present invention under the noise condition of SNR=2dB in another embodiment of the present application;

[0047] Figure 8 This is a structural diagram of a wheel flat fault identification system in one embodiment of the present application;

[0048] Fig. 9 This is a structural diagram of a terminal device in one of the implementation modes of the present application. DETAILED DESCRIPTION

[0049] In view of the problem of low accuracy of traditional wheel flat fault identification methods, the present invention provides a wheel flat fault identification method, system, terminal device and medium, wherein the wheel flat fault identification model constructed by the method includes a channel attention module and a spatial feature module, which can extract and combine the feature information of the vertical vibration acceleration signal and the wheel flat in the two dimensions of channel and space, accurately mine the correlation features between the vertical vibration acceleration signal and the wheel flat, which is beneficial to improving the accuracy of wheel flat fault identification; in the spatial feature module, the channel attention feature map and the two-dimensional time-frequency image are multiplied element by element, which can realize adaptive feature optimization and dynamically adjust the feature weights, so that the wheel flat fault identification model can mine the important features of the wheel flat affecting the vertical vibration acceleration signal, reduce noise interference, and improve the accuracy of wheel flat fault identification; in addition, the method does not rely on people's subjective experience and judgment, can reduce the influence of human factors, and further improve the accuracy of wheel flat fault identification.

[0050] It should be noted that the wheel flat fault identification method provided by the present invention is widely applicable to wheel flat fault identification of vehicles, including but not limited to automobiles, trains, high-speed trains, etc.

[0051] The wheel flat fault identification method provided by the present invention is described below.

[0052] like Figure 1As shown, the wheel flat fault identification method includes the following steps:

[0053] Step 11, obtaining a wheel flat fault sample data set.

[0054] The wheel flat fault sample data set includes multiple wheel flat lengths and vertical vibration acceleration signals generated by the vibration of the vehicle frame at each wheel flat length. The multiple wheel flat lengths correspond to the vertical vibration acceleration signals one by one in the time domain.

[0055] It should be understood that the vertical vibration acceleration signal refers to the vibration response of the vehicle in the vertical direction. This signal is affected by the flat condition of the wheel and can be used for wheel flat fault identification, comfort evaluation and vehicle design optimization.

[0056] In some embodiments of the present invention, a subway vehicle dynamics model can be constructed to obtain multiple wheel flat lengths and vertical vibration acceleration signals generated by the vibration of the vehicle frame at each wheel flat length through simulation.

[0057] Exemplarily, taking a subway vehicle as an example, a multi-rigid body dynamics model of the subway vehicle is constructed based on the dynamics software (Wheel / Rail) according to the parameter table of the subway vehicle; three-dimensional finite element models of the subway train wheelset and track are established separately through finite element analysis, and the degree of freedom of the model is reduced by analyzing its substructure; the operating parameters of the subway train (such as speed, acceleration and track characteristics) are set; the dynamics software and the finite element model are combined to simulate different wheel flat lengths and the vertical vibration acceleration signals generated by the vibration of the vehicle frame at each wheel flat length.

[0058] Step 12, for each wheel flat fault sample, convert the vertical vibration acceleration signal based on the time domain into the space domain, and perform time-frequency processing on the converted space domain samples to obtain a two-dimensional time-frequency image of the wheel flat fault.

[0059] The above spatial domain samples are used to describe the spatial relationship between the wheel flat and the vertical vibration acceleration signal.

[0060] In some embodiments of the present invention, the above step 12 includes steps 12.1 to 12.4.

[0061] Step 12.1, data preprocessing.

[0062] A filter (such as a low-pass filter) is used to denoise the acquired wheel flat fault sample data set to eliminate environmental noise and interference.

[0063] Step 12.2, spatial domain cutting.

[0064] The window size is determined, and the vertical vibration acceleration signal is cut into multiple windows in the spatial domain. Each window represents a segment of the vertical vibration acceleration signal at different wheel flat lengths.

[0065] Step 12.3, spatial domain interpolation.

[0066] The cut wheel flat fault samples are interpolated to improve the resolution of the data.

[0067] Step 12.4, time-frequency analysis.

[0068] Perform short-time Fourier transform on the spatial domain samples in each window to obtain the distribution of the vertical vibration acceleration signal in time and frequency; convert the short-time Fourier transform results in each window into a two-dimensional image to obtain a two-dimensional time-frequency image.

[0069] Step 13, construct a wheel flat fault recognition model for wheel flat fault recognition, and use the two-dimensional time-frequency image to train the wheel flat fault recognition model until the loss value of the wheel flat fault recognition model is less than a preset loss threshold, thereby obtaining a trained wheel flat fault recognition model.

[0070] In an embodiment of the present invention, the above-mentioned wheel flat fault recognition model includes a channel attention module for pooling the two-dimensional time-frequency image to generate a channel attention feature map, a spatial feature module for multiplying the channel attention feature map and the two-dimensional time-frequency image element by element to generate a spatial attention feature map, and a recognition module for obtaining the wheel flat fault recognition result according to the spatial attention feature map.

[0071] like Figure 2 As shown, the channel attention module 201, the spatial feature module 202 and the recognition module 203 are connected in sequence.

[0072] The channel attention module is described in detail below.

[0073] like Figure 3 As shown, the channel attention module includes a first maximum pooling layer 301, a first global pooling layer 302, a multi-layer perceptron network 303 and a first activation layer 304.

[0074] Among them, the input ends of the first maximum pooling layer 301 and the first global pooling layer 302 receive two-dimensional time-frequency images, the input end of the multi-layer perceptron network 303 is respectively connected to the output end of the first maximum pooling layer 301 and the output end of the first global pooling layer 302, the output end of the multi-layer perceptron network 303 is connected to the input end of the first activation layer 304, and the output end of the first activation layer 304 outputs a channel attention feature map.

[0075] The spatial feature module is described in detail below.

[0076] like Figure 4 As shown, the spatial feature module includes a merging layer 401, a splicing layer 402, a second maximum pooling layer 403, a second global pooling layer 404, a convolution layer 405 and a second activation layer 406.

[0077] Among them, the merging layer 401 is used to multiply the channel attention feature map and the two-dimensional time-frequency image element by element to obtain a new feature map, the input ends of the second maximum pooling layer 403 and the second global pooling layer 404 receive the new feature map, the input end of the splicing layer 402 is connected to the output end of the second maximum pooling layer 403 and the output end of the second global pooling layer 404 respectively, the output end of the splicing layer 402 is connected to the input end of the convolution layer 405, the output end of the convolution layer 405 is connected to the input end of the second activation layer 406, and the output end of the second activation layer 406 outputs the spatial attention feature map.

[0078] In another embodiment of the present invention, in order to further improve the accuracy of the model, the model is also provided with a fully connected layer 407 for enhancing the nonlinear feature extraction capability and a dropout layer 408 for preventing overfitting of the wheel flat fault identification model.

[0079] Among them, the input end of the fully connected layer 407 is connected to the output end of the concatenated layer 402, and the output end of the fully connected layer 407 is connected to the input end of the convolutional layer 405;

[0080] The input end of the dropout layer 408 is connected to the output end of the second activation layer 406, and the output end of the dropout layer 408 is connected to the input end of the recognition module.

[0081] The identification module is described in detail below.

[0082] In the recognition module, the final features used for wheel flat fault recognition can be obtained by performing element-by-element multiplication of the spatial attention feature map and the two-dimensional time-frequency image.

[0083] It should be noted that, in the embodiment of the present invention, the above-mentioned final feature is the wheel flat length. In some embodiments of the present invention, the above-mentioned final feature also includes the wheel flat depth.

[0084] In some embodiments of the present invention, the above loss value is a cross entropy loss value.

[0085] The following is an explanation of the training process of the wheel flat fault recognition model, which specifically includes the following steps:

[0086] Step I, initializing the model parameters of the wheel flat fault recognition model;

[0087] Step II, inputting the wheel flat fault samples in the wheel flat fault sample data set into the wheel flat fault recognition model one by one to obtain an initial recognition result; the recognition result is a predicted value of the wheel flat length corresponding to each vertical vibration acceleration signal;

[0088] Step III, calculating the cross entropy loss value of the wheel flat fault recognition model according to the predicted value of the wheel flat length and the actual wheel flat length in the wheel flat fault sample data set;

[0089] Step IV, if the cross entropy loss value is less than the preset loss threshold (for example, 0.01), it means that the wheel flat fault recognition model has been fitted and can be used as the trained wheel flat fault recognition model; otherwise, the model parameters of the wheel flat fault recognition model are updated to obtain an updated wheel flat fault recognition model, and the updated wheel flat fault recognition model is used as the wheel flat fault recognition model in step II, and the process returns to step II.

[0090] In one embodiment of the present invention, the model parameters of the trained wheel flat fault recognition model are shown in Table 1.

[0091] Table 1

[0092]

[0093] Step 14, inputting the collected vertical vibration acceleration signal to be measured into the trained wheel flat fault recognition model to obtain the wheel flat fault recognition result.

[0094] In an embodiment of the present invention, the wheel flat scar fault identification result is a mild flat scar, a moderate flat scar or a severe flat scar; wherein, the flat scar length corresponding to a mild flat scar is less than or equal to 25 mm, the flat scar length corresponding to a moderate flat scar is greater than 25 mm and less than or equal to 40 mm, and the flat scar length corresponding to a severe flat scar is greater than 40 mm and less than or equal to 60 mm.

[0095] For example, in some embodiments of the present invention, when the wheel flat fault identification result is a mild flat, it means that the vehicle's wheel flat fault is mild and can be temporarily left untreated. When the wheel flat fault identification result is a moderate flat, it means that the vehicle's wheel flat fault is severe and needs to be treated and maintained within a certain period of time. When the wheel flat fault identification result is a severe flat, it means that the vehicle's wheel flat fault is extremely critical and needs to be stopped immediately for treatment and maintenance.

[0096] The following is an explanation of the processing process of the channel attention module in the wheel flat fault recognition model.

[0097] Specifically, the first maximum pooling layer and the first global pooling layer are used to perform maximum pooling operations and global pooling operations on the input two-dimensional time-frequency image to obtain local feature maps and global feature vectors; then a multi-layer perceptron network (MLP) is used to process the local feature maps and global feature vectors after the pooling operation, and an element-by-element addition operation is performed on them, and then the sigmoid activation function in the first activation layer is normalized to generate the weight of each channel of the feature map, that is, the channel attention feature map.

[0098] The expression of the obtained channel attention feature map is as follows:

[0099]

[0100] Among them, M c (F) represents the channel attention feature map, σ(·) represents the sigmoid activation function, W 0 ,W 1 represents the shared weight parameter in the multilayer perceptron network, MLP(·) represents the multilayer perceptron network, AvgPool(·) represents global pooling, MaxPool(·) represents maximum pooling, F represents a two-dimensional time-frequency image, represents the local feature map, represents the global eigenvector.

[0101] The following is an explanation of the processing process of the spatial feature module in the wheel flat fault recognition model.

[0102] Specifically, the merging layer is first used to perform element-by-element multiplication of the channel attention feature map and the two-dimensional time-frequency image to obtain a new feature map; then the second maximum pooling layer and the second global pooling layer are used for pooling respectively, and then the splicing layer is used to splice the pooling results, and then the convolution layer is used to perform channel dimensionality reduction on the splicing results, and finally the sigmoid activation function in the second activation layer is used to normalize the weight of each channel of the feature map to generate a spatial attention feature map.

[0103] The expression of the generated spatial attention feature map is as follows:

[0104]

[0105] Among them, M s (F ′ ) represents the spatial attention feature map, f 7×7 represents the convolution operation with a kernel size of 7×7, (·;·) represents concatenation, and F ′ Represents the new feature map obtained by element-by-element multiplication of the channel attention feature map and the two-dimensional time-frequency image. Indicates F ′ The global pooling result is Indicates F ′ The maximum pooling result.

[0106] In order to further verify the accuracy of the wheel flat fault identification method provided by the present invention, in another embodiment of the present invention, the wheel flat fault identification method provided by the present invention, LeNet-5

[103] , VGG-16

[104] , and ResNet are compared, wherein the model parameter settings of LeNet-5 and VGG-16 are kept consistent; at the same time, in order to reflect fairness, the initial hyperparameter settings of all the comparison methods are kept consistent with the proposed method. Due to the existence of the primary suspension, the vertical vibration acceleration response of the frame end is somewhat weakened by the noise signal interference. Therefore, only the recognition accuracy of the wheel flat fault by different methods under SNR=2~10dB (increment 2dB) and noise-free conditions is considered, and the results are shown in Table 2.

[0107] Table 2

[0108]

[0109] It can be seen from Table 2 that under the 6 noise conditions considered, the recognition accuracy of the wheel flat fault recognition method provided by the present invention is better than other comparison methods. Specifically, the average recognition accuracy of the wheel flat fault recognition method provided by the present invention under different noise conditions is improved by 4.15%, 3.27% and 1.56% respectively compared with the LeNet-5 model, VGG-16 model and ResNet model, which shows that the present method has strong recognition ability. Secondly, as the proportion of added noise continues to increase, the test recognition accuracy of each model gradually decreases, and the higher the noise proportion, the higher the noise resistance requirement for the model; and under different noise conditions, the recognition accuracy of the wheel flat fault recognition method provided by the present invention is improved by up to 6.07%, 4.12% and 3.04% respectively compared with other comparison models, and its recognition accuracy is maintained above 93%, which shows that the wheel flat fault recognition method provided by the present invention has better noise resistance under the noise background, can effectively reduce the interference of redundant information in the noisy signal, accurately extract the detailed features of the wheel flat fault, thereby improving the accuracy of wheel flat fault recognition.

[0110] In order to more intuitively highlight the test recognition accuracy of different methods under different noise conditions, the following plots are made based on Table 2: Figure 5The bar graph shown in the figure shows that CBAM-ResdNet is a wheel flat fault identification method provided by the present invention. By comparing the identification bar graphs under different noise conditions, it can be clearly seen that as the proportion of added noise increases, the recognition accuracy of this method is improved more significantly compared with other comparison methods, which also shows that the present invention has good noise resistance. In addition, from the perspective of the standard deviation changes of various methods under different noise conditions, the standard deviation of the present invention as a whole has a smaller range of change, which can reflect that the present invention has good stability.

[0111] In order to further explore the recognition effects of different methods, taking the recognition situation under the noise condition of SNR=2dB as an example, the last hidden layer features of the models provided by different methods are visualized by t-SNE two-dimensional display, as shown in the figure below: Figure 6a , 6b , 6c, 6d. Figure 6a , 6b It is not difficult to see from 6c and 6d that the features extracted by LeNet-5, VGG-16 and ResNet have serious confusion between the normal and mild categories. This is because when there is a small flat scar in the mild fault category, the time-frequency graph obtained after preprocessing has a small energy within the flat scar fault frequency range, which is easily confused with the time-frequency graph of the normal category, resulting in the model's poor recognition ability for the mild fault category under noise conditions. It can be seen from the visualization results of the present invention that the confusion between the normal and mild fault categories has been significantly improved, and the degree of aggregation of other fault categories is also better than the comparison method, which shows that the present invention can effectively extract the detailed features of the fault under noise interference conditions, reduce the interference of redundant information in the noisy signal, and thus improve the recognition accuracy of difficult-to-identify categories.

[0112] In order to analyze the specific recognition accuracy of different methods for each fault category, Figure 7a , Figure 7b , Figure 7c , Figure 7d The confusion matrix results of the LeNet-5 model, VGG-6 model, ResNet model and the method provided by the present invention under the condition of SNR=2dB noise are given respectively. Figure 7a , Figure 7b , Figure 7c , Figure 7d It can be seen that all methods have good recognition accuracy for normal, moderate and severe fault categories, all maintained above 90%. For the recognition ability of mild fault categories, the comparison method is obviously insufficient. However, the recognition accuracy of the mild fault category of the present invention is greatly improved compared with the comparison method, which is improved by 16.5%, 21% and 17% respectively. This fully demonstrates that the present invention can effectively improve the recognition ability of difficult-to-identify categories under noise interference conditions, and has better robustness than the comparison method.

[0113] The wheel flat fault identification system provided by the present invention is described below.

[0114] like Figure 8 As shown, the wheel flat fault identification system 800 includes:

[0115] The data acquisition module 801 is used to acquire a wheel flat fault sample data set; the wheel flat fault sample data set includes a plurality of wheel flat lengths and a vertical vibration acceleration signal generated by the vibration of the vehicle frame at each wheel flat length, and the plurality of wheel flat lengths and the vertical vibration acceleration signal correspond one to one in the time domain;

[0116] The preprocessing module 802 is used to convert the vertical vibration acceleration signal based on the time domain into the space domain for each wheel flat fault sample, and perform time-frequency processing on the converted space domain samples to obtain a two-dimensional time-frequency image of the wheel flat fault; the space domain samples are used to describe the spatial relationship between the wheel flat and the vertical vibration acceleration signal;

[0117] The model training module 803 is used to construct a wheel flat fault recognition model for wheel flat fault recognition, and train the wheel flat fault recognition model using a two-dimensional time-frequency image until the loss value of the wheel flat fault recognition model is less than a preset loss threshold, thereby obtaining a trained wheel flat fault recognition model; the wheel flat fault recognition model includes a channel attention module for pooling the two-dimensional time-frequency image to generate a channel attention feature map, a spatial feature module for multiplying the channel attention feature map and the two-dimensional time-frequency image element by element to generate a spatial attention feature map, and a recognition module for obtaining a wheel flat fault recognition result according to the spatial attention feature map;

[0118] The wheel flat identification module 804 is used to input the collected vertical vibration acceleration signal to be tested into the trained wheel flat fault identification model to obtain the wheel flat fault identification result.

[0119] It should be noted that the information interaction, execution process, etc. between the above-mentioned devices / units are not related to the implementation method of the present application.

[0120] The examples are based on the same concept, and their specific functions and technical effects can be specifically referred to in the method embodiment section, which will not be repeated here. Those skilled in the art can clearly understand that for the convenience and simplicity of description, only the division of the above-mentioned functional units and modules is used as an example. In practical applications, the above-mentioned function allocation can be completed by different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiments can be integrated in a processing unit, or each unit can exist physically separately, or two or more units can be integrated in one unit, and the above-mentioned integrated units can be implemented in the form of hardware or in the form of software functional units. In addition, the specific names of the functional units and modules are only for the convenience of distinguishing each other, and are not used to limit the scope of protection of this application. The specific working process of the units and modules in the above-mentioned system can refer to the corresponding process in the aforementioned method embodiment, which will not be repeated here.

[0121] like Fig. 9 As shown, an embodiment of the present invention provides a terminal device, such as Fig. 9 As shown, the terminal device D10 of this embodiment includes: at least one processor D100 ( Fig. 9 Only one processor is shown in the figure), a memory D101, and a computer program D102 stored in the memory D101 and executable on the at least one processor D100, wherein the processor D100 implements the steps of any of the above-mentioned method embodiments when executing the computer program D102.

[0122] Specifically, when the processor D100 executes the computer program D102, a wheel flat fault sample data set is obtained; for each wheel flat fault sample, the vertical vibration acceleration signal based on the time domain is converted to the spatial domain, and the converted spatial domain samples are processed in a time-frequency manner to obtain a two-dimensional time-frequency image of the wheel flat fault; a wheel flat fault recognition model for identifying wheel flat faults is constructed, and the wheel flat fault recognition model is trained using the two-dimensional time-frequency image until the loss value of the wheel flat fault recognition model is less than a preset loss threshold, thereby obtaining a trained wheel flat fault recognition model; the collected vertical vibration acceleration signal to be tested is input into the trained wheel flat fault recognition model to obtain a wheel flat fault recognition result. Among them, the wheel flat fault recognition model constructed by this method includes a channel attention module and a spatial feature module, which can extract and combine the feature information of the vertical vibration acceleration signal and the wheel flat in the two dimensions of channel and space, accurately mine the correlation characteristics between the vertical vibration acceleration signal and the wheel flat, which is beneficial to improving the accuracy of wheel flat fault recognition; in the spatial feature module, the channel attention feature map and the two-dimensional time-frequency image are multiplied element by element, which can realize adaptive feature optimization and dynamically adjust the feature weights, so that the wheel flat fault recognition model can mine the important features of the wheel flat affecting the vertical vibration acceleration signal, reduce noise interference, and improve the accuracy of wheel flat fault recognition; in addition, this method does not rely on people's subjective experience and judgment, can reduce the influence of human factors, and further improve the accuracy of wheel flat fault recognition.

[0123] The processor D100 may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor, etc.

[0124] In some embodiments, the memory D101 may be an internal storage unit of the terminal device D10, such as a hard disk or memory of the terminal device D10. In other embodiments, the memory D101 may also be an external storage device of the terminal device D10, such as a plug-in hard disk, a smart memory card (SMC, SmartMedia Card), a secure digital (SD, Secure Digital) card, a flash card (Flash Card), etc. equipped on the terminal device D10. Further, the memory D101 may also include both an internal storage unit of the terminal device D10 and an external storage device. The memory D101 is used to store an operating system, an application program, a boot loader (BootLoader), data and other programs, such as the program code of the computer program, etc. The memory D101 may also be used to temporarily store data that has been output or is to be output.

[0125] An embodiment of the present application further provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps in the above-mentioned method embodiments can be implemented.

[0126] An embodiment of the present application provides a computer program product. When the computer program product is run on a terminal device, the terminal device can implement the steps in the above-mentioned method embodiments when executing the computer program product.

[0127] A person skilled in the art should understand that the discussion of any of the above embodiments is merely illustrative and is not intended to imply that the scope of protection of the present application is limited to these examples. In line with the concept of the present application, the technical features in the above embodiments or different embodiments may be combined, the steps may be implemented in any order, and there are many other variations of different aspects of one or more embodiments of the present application as described above, which are not provided in detail for the sake of simplicity.

[0128] One or more embodiments of the present application are intended to cover all such substitutions, modifications and variations that fall within the broad scope of the present application. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of one or more embodiments of the present application should be included in the protection scope of the present application.

Claims

1. A wheel flat fault identification method, characterized in that: include: Acquire a wheel flat fault sample data set; the wheel flat fault sample data set includes a plurality of wheel flat lengths and a vertical vibration acceleration signal generated by the vibration of the vehicle frame at each of the wheel flat lengths, wherein the plurality of wheel flat lengths correspond to the vertical vibration acceleration signal in a one-to-one manner in the time domain; For each wheel flat fault sample, the vertical vibration acceleration signal based on the time domain is converted to the space domain, and the converted space domain samples are subjected to time-frequency processing to obtain a two-dimensional time-frequency image of the wheel flat fault; the space domain samples are used to describe the spatial relationship between the wheel flat and the vertical vibration acceleration signal; Constructing a wheel flat fault recognition model for wheel flat fault recognition, and training the wheel flat fault recognition model using the two-dimensional time-frequency image until a loss value of the wheel flat fault recognition model is less than a preset loss threshold, thereby obtaining a trained wheel flat fault recognition model; The wheel flat fault recognition model includes a channel attention module for pooling the two-dimensional time-frequency image to generate a channel attention feature map, a spatial feature module for multiplying the channel attention feature map and the two-dimensional time-frequency image element by element to generate a spatial attention feature map, and a recognition module for obtaining a wheel flat fault recognition result according to the spatial attention feature map; The collected vertical vibration acceleration signal to be tested is input into the trained wheel flat fault recognition model to obtain a wheel flat fault recognition result.

2. The wheel flat fault identification method according to claim 1, characterized in that: The channel attention module, the spatial feature module and the recognition module are connected in sequence; The channel attention module includes a first maximum pooling layer, a first global pooling layer, a multi-layer perceptron network and a first activation layer; wherein the input ends of the first maximum pooling layer and the first global pooling layer receive the two-dimensional time-frequency image, the input end of the multi-layer perceptron network is respectively connected to the output end of the first maximum pooling layer and the output end of the first global pooling layer, the output end of the multi-layer perceptron network is connected to the input end of the first activation layer, and the output end of the first activation layer outputs the channel attention feature map; The spatial feature module includes a merging layer, a splicing layer, a second maximum pooling layer, a second global pooling layer, a convolution layer and a second activation layer; wherein the merging layer is used to perform element-by-element multiplication of the channel attention feature map and the two-dimensional time-frequency image to obtain a new feature map, the input ends of the second maximum pooling layer and the second global pooling layer receive the new feature map, the input end of the splicing layer is respectively connected to the output end of the second maximum pooling layer and the output end of the second global pooling layer, the output end of the splicing layer is connected to the input end of the convolution layer, the output end of the convolution layer is connected to the input end of the second activation layer, and the output end of the second activation layer outputs the spatial attention feature map.

3. The wheel flat fault identification method according to claim 2 is characterized in that: The expression of the channel attention feature map is as follows: Among them, M c (F) represents the channel attention feature map, σ(·) represents the sigmoid activation function, W00, W1 represent the shared weight parameters in the multi-layer perceptron network, MLP(·) represents the multi-layer perceptron network, AvgPool(·) represents global pooling, MaxPool(·) represents maximum pooling, F represents the two-dimensional time-frequency image, represents the local feature map, represents the global eigenvector.

4. The wheel flat fault identification method according to claim 3 is characterized in that: The expression of the spatial attention feature map is as follows: Among them, M s (F′) represents the spatial attention feature map, f 7×7 represents a convolution operation with a convolution kernel size of 7×7, (·;·) represents concatenation, F′ represents a new feature map obtained by element-by-element multiplication of the channel attention feature map and the two-dimensional time-frequency image, represents the global pooling result of F′, Represents the maximum pooling result of F′.

5. The wheel flat fault identification method according to claim 4, characterized in that: The wheel flat scar fault identification result is a mild flat scar, a moderate flat scar or a severe flat scar; wherein, the flat scar length corresponding to a mild flat scar is less than or equal to 25 mm, the flat scar length corresponding to a moderate flat scar is greater than 25 mm and less than or equal to 40 mm, and the flat scar length corresponding to a severe flat scar is greater than 40 mm and less than or equal to 60 mm.

6. The wheel flat fault identification method according to claim 2, characterized in that: It also includes a fully connected layer for enhancing the nonlinear feature extraction capability and a dropout layer for preventing the overfitting of the wheel flat fault recognition model; The input end of the fully connected layer is connected to the output end of the second global pooling layer, and the output end of the fully connected layer is connected to the input end of the convolutional layer; The input end of the dropout layer is connected to the output end of the second activation layer, and the output end of the dropout layer is connected to the input end of the recognition module.

7. The wheel flat fault identification method according to claim 1, characterized in that: The loss value is a cross entropy loss value.

8. A wheel flat fault identification system, characterized in that: include: A data acquisition module is used to acquire a wheel flat fault sample data set; the wheel flat fault sample data set includes a plurality of wheel flat lengths and a vertical vibration acceleration signal generated by the vibration of the vehicle frame at each of the wheel flat lengths, wherein the plurality of wheel flat lengths correspond to the vertical vibration acceleration signal in a one-to-one manner in the time domain; A preprocessing module is used to convert the vertical vibration acceleration signal based on the time domain into the space domain for each wheel flat fault sample, and perform time-frequency processing on the converted space domain samples to obtain a two-dimensional time-frequency image of the wheel flat fault; the space domain samples are used to describe the spatial relationship between the wheel flat and the vertical vibration acceleration signal; A model training module, used for constructing a wheel flat fault recognition model for wheel flat fault recognition, and training the wheel flat fault recognition model using the two-dimensional time-frequency image until a loss value of the wheel flat fault recognition model is less than a preset loss threshold, thereby obtaining a trained wheel flat fault recognition model; The wheel flat fault recognition model includes a channel attention module for pooling the two-dimensional time-frequency image to generate a channel attention feature map, a spatial feature module for multiplying the channel attention feature map and the two-dimensional time-frequency image element by element to generate a spatial attention feature map, and a recognition module for obtaining a wheel flat fault recognition result according to the spatial attention feature map; The wheel flat identification module is used to input the collected vertical vibration acceleration signal to be tested into the trained wheel flat fault identification model to obtain a wheel flat fault identification result.

9. A terminal device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the wheel flat fault identification method as described in any one of claims 1 to 7 is implemented.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the wheel flat fault identification method as described in any one of claims 1 to 7 is implemented.

Images