Rail screw hole crack identification method, system and readable storage medium

Abstract

The present invention discloses a method, system and readable storage medium for identifying cracks in rail screw holes. The method comprises: inputting at least one ultrasonic detection B-scan image sample into an improved EfficientNet network, proposing a TokenDrop regularization, and the EfficientNet network classifies at least one ultrasonic detection B-scan image sample to obtain at least one classification result; inputting at least one classification result and an actual rail screw hole crack recognition result corresponding to at least one classification result into an improved GNN, training and generating a rail screw hole crack recognition model; obtaining a real-time ultrasonic detection B-scan image obtained by ultrasonically detecting the rail according to a preset Star strategy; inputting the real-time ultrasonic detection B-scan image into the rail screw hole crack recognition model, and outputting the screw hole crack recognition result. The complex process of manually extracting features is avoided, the recognition rate is improved, and the model volume is reduced.

Description

Technical Field

[0001] The present invention belongs to the technical field of railway rail detection, and in particular relates to a rail screw hole crack identification method, system and readable storage medium. Background Art

[0002] Railways are the main transportation force, which leads to long-term load on rails, resulting in a large number of hidden dangers. The quality of rails directly affects the safety and stability of railway lines. As the use time increases, the rails will age and wear. This situation may cause wear, deformation or breakage on the surface of the rails. Unreasonable or untimely maintenance and repair may cause the problem to deteriorate further or even cause accidents. Environmental factors will also have an impact on the rails. Changes in climatic conditions, erosion by chemical substances, and impacts from external objects may all cause damage or corrosion to the rails. Therefore, it is urgent to strengthen maintenance management.

[0003] At present, the main method for internal flaw detection of rails is ultrasonic testing, but this requires workers to rely on solid professional knowledge and test data to judge the damage of the railway, which greatly affects the development of transportation work. Therefore, attempts are made to apply neural networks to defect detection to relieve the pressure on workers and improve efficiency.

[0004] According to current research results, the large-scale use of neural networks provides a new idea for solving the problem. Images of rail screw holes are collected and fault diagnosis is performed through neural networks. However, the recognition success rate has been unsatisfactory and cannot be promoted and used on a large scale, and the efficiency has not been greatly improved. Summary of the invention

[0005] The present invention provides a rail screw hole crack recognition method, system and readable storage medium, which are used to solve the technical problem of low recognition success rate of fault diagnosis of rail screw hole images by neural networks.

[0006] In a first aspect, the present invention provides a method for identifying a rail screw hole crack, comprising:

[0007] At least one ultrasonic detection B-scan image sample is obtained, and the at least one ultrasonic detection B-scan image sample is input into an improved EfficientNet network, wherein the improved EfficientNet network classifies the at least one ultrasonic detection B-scan image sample according to image contour features in the at least one ultrasonic detection B-scan image sample to obtain at least one classification result, wherein the improved EfficientNet network includes a regularization strategy, and the expression of the regularization strategy is:

[0008] ,

[0009] ,

[0010] ,

[0011] In the formula, is the input feature of the jump connection between the input node and the output node, is the output feature of each intermediate node in the top-down channel of the fusion layer, is the final output of the adaptive weighted fusion algorithm, To adjust the convolution of the input channel in the jump layer, is the input feature weight, It corresponds to the two groups of E-ELAN and MPConv modules from bottom to top in the fusion layer. is the parameter weight, is the mean of the eigenvectors, is the characteristic value, is the real boundary, is the feature vector after dimensionality reduction, is the feature vector, It corresponds to a group of Conv and BatchNorm modules from bottom to top in the fusion layer. is the number of superpixel blocks, is the total spectral distance, is the pixel space distance, is the spectral vector, A small amount to ensure the stability of weight values;

[0012] Inputting the at least one classification result and the actual rail screw hole crack recognition result corresponding to the at least one classification result into the improved GNN to train and generate a rail screw hole crack recognition model;

[0013] Acquire a real-time ultrasonic testing B-scan image obtained by ultrasonic testing the rail according to a preset STAR strategy;

[0014] The real-time ultrasonic detection B-scan image is input into a rail screw hole crack recognition model, and the rail screw hole crack recognition model outputs a screw hole crack recognition result.

[0015] In a second aspect, the present invention provides a rail screw hole crack identification system, comprising:

[0016] The classification module is configured to obtain at least one ultrasonic detection B-scan image sample, input the at least one ultrasonic detection B-scan image sample into the improved EfficientNet network, and the improved EfficientNet network classifies the at least one ultrasonic detection B-scan image sample according to the image contour features in the at least one ultrasonic detection B-scan image sample to obtain at least one classification result, wherein the improved EfficientNet network includes a regularization strategy, and the expression of the regularization strategy is:

[0017] ,

[0018] ,

[0019] ,

[0020] In the formula, is the input feature of the jump connection between the input node and the output node, is the output feature of each intermediate node in the top-down channel of the fusion layer, is the final output of the adaptive weighted fusion algorithm, To adjust the convolution of the input channel in the jump layer, is the input feature weight, It corresponds to the two groups of E-ELAN and MPConv modules from bottom to top in the fusion layer. is the parameter weight, is the mean of the eigenvectors, is the characteristic value, is the real boundary, is the feature vector after dimensionality reduction, is the feature vector, It corresponds to a group of Conv and BatchNorm modules from bottom to top in the fusion layer. is the number of superpixel blocks, is the total spectral distance, is the pixel space distance, is the spectral vector, A small amount to ensure the stability of weight values;

[0021] A training module, configured to input the at least one classification result and an actual rail screw hole crack recognition result corresponding to the at least one classification result into the improved GNN, to train and generate a rail screw hole crack recognition model;

[0022] An acquisition module configured to acquire a real-time ultrasonic detection B-scan image obtained by ultrasonically detecting the rail according to a preset STA strategy;

[0023] The output module is configured to input the real-time ultrasonic detection B-scan image into a rail screw hole crack recognition model, and the rail screw hole crack recognition model outputs a screw hole crack recognition result.

[0024] According to a third aspect, an electronic device is provided, comprising: at least one processor, and a memory communicatively connected to the at least one processor, wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute the steps of the rail screw hole crack identification method of any embodiment of the present invention.

[0025] In a fourth aspect, the present invention further provides a computer-readable storage medium having a computer program stored thereon, wherein when the program instructions are executed by a processor, the processor executes the steps of the rail screw hole crack identification method of any embodiment of the present invention.

[0026] The rail screw hole crack recognition method, system and readable storage medium of the present application obtain a real-time ultrasonic detection B-scan image obtained by ultrasonically detecting the rail according to a preset STA strategy; the real-time ultrasonic detection B-scan image is input into a rail screw hole crack recognition model, and the rail screw hole crack recognition model outputs a screw hole crack recognition result, so that it has an autonomous recognition function, avoids the complex process of manually extracting features, improves the recognition rate, reduces the model volume, and is conducive to the realization of engineering applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0028] Figure 1 A flow chart of a rail screw hole crack identification method provided by one embodiment of the present invention;

[0029] Figure 2 A structural block diagram of a rail screw hole crack identification system provided by an embodiment of the present invention;

[0030] Figure 3 It is a schematic diagram of the structure of an electronic device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0031] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0032] See also Figure 1 , which shows a flow chart of a rail screw hole crack identification method of the present application.

[0033] like Figure 1 As shown, the rail screw hole crack identification method specifically includes the following steps:

[0034] Step S101, obtaining at least one ultrasonic detection B-scan image sample, inputting the at least one ultrasonic detection B-scan image sample into an improved EfficientNet network, and classifying the at least one ultrasonic detection B-scan image sample according to image contour features in the at least one ultrasonic detection B-scan image sample to obtain at least one classification result.

[0035] In this step, the existing Vision Transformer model usually directly uses the dropout algorithm commonly used in CNN as a regularization method. This algorithm randomly drops the feature units of the tokens generated after the patch embedding operation and sets a uniform drop probability for different depths of the attention layer. In the time-frequency diagram, the application of the dropout layer that randomly deactivates the feature tokens is not ideal. In order to solve the above problems, a new regularization strategy is designed, which randomly deactivates some related feature units. These deactivated feature units are mapped to the time-frequency diagram equivalent to the deactivated part of the spatial continuous area. Screw hole cracks usually appear as anomalies in local areas, so regularization methods that help the model better capture these local features are very important for improving the accuracy of screw hole crack recognition. The regularization strategy allows the network to focus on learning the features of key components, reduce overfitting, and improve the generalization ability of the model.

[0036] The expression of the regularization strategy is:

[0037] ,

[0038] ,

[0039] ,

[0040] In the formula, is the input feature of the jump connection between the input node and the output node, is the output feature of each intermediate node in the top-down channel of the fusion layer, is the final output of the adaptive weighted fusion algorithm, To adjust the convolution of the input channel in the jump layer, is the input feature weight, It corresponds to the two groups of E-ELAN and MPConv (Multi-scale Pyramid Convolution) modules in the fusion layer from bottom to top. is the parameter weight, is the mean of the eigenvectors, is the characteristic value, is the real boundary, is the feature vector after dimensionality reduction, is the feature vector, It corresponds to a group of Conv and BatchNorm (batch normalization) modules in the fusion layer from bottom to top. is the number of superpixel blocks, is the total spectral distance, is the pixel space distance, is the spectral vector, A small amount to ensure the stability of weight values;

[0041] Furthermore, a PFE module is proposed, which combines the characteristics of DCN and ViT architecture to extract local and global features in the early stage and provide additional local information for the attention mechanism to enhance the ability to extract richer features. The PFE module consists of a 1-D convolution block, a 1-DDCN block and a linear layer. The proposed PFE module solves the limitations of the traditional transformer-based model by integrating DCN. The DCN layer can sample adaptive shapes in local areas, enhancing the model's ability to capture shape and structured information from the input. By dynamically adjusting the shape and position of the convolution kernel, the linear layer plays an important role in the PFE module. It converts the extracted features into higher-dimensional feature representations through linear mapping. This mapping helps the model better capture the complex patterns and structures in the data and understand the relationship between the data, thereby improving the performance of the model in crack recognition and feature extraction. After feature extraction, the PFE module connects the feature sequence to the CLS token, which represents the classification head. The CLS token is a learnable embedding that is only used for the final crack classification. Finally, the absolute position encoding is added to generate the encoder input sequence. The computational description of this process is:

[0042] ,

[0043] In the formula, is the input sequence of the encoder, is the patch sequence after operation, is the classification head vector, is the linear mapping matrix, is the position encoding matrix, is the target image feature, is the degraded image feature;

[0044] ,

[0045] ,

[0046] In the formula, is the regression function, is the number of samples used for classification training, is a 0-1 label, To improve the output of the EfficientNet network, is the weight coefficient, To improve the weight matrix of the EfficientNet network, is the number of test samples, is the mean absolute percentage error, is the weighted mean absolute percentage error, is the weight parameter to balance the spatial distance and spectral distance, is the new cluster center, is the mean value of the pixels, is the number of pixels in the neighborhood;

[0047] Specifically, the convolution kernel is used as a filter to find the contour features in the image and transform the image area matrix With the filter matrix The corresponding elements are multiplied and then added, the convolution step size is used to slide the area, and the next image area matrix is ​​convolved to obtain the feature matrix , where the image area matrix The expression is:

[0048]

[0049] In the formula, is the first channel of the corresponding feature map, is the second channel of the corresponding feature map, is the third channel of the corresponding feature map, is the 4th channel of the corresponding feature map, is the 5th channel of the corresponding feature map, is the sixth channel of the corresponding feature map, is the 7th channel of the corresponding feature map, is the 8th channel of the corresponding feature map, is the 9th channel of the corresponding feature map;

[0050] Filter Matrix The expression is:

[0051] ,

[0052] In the formula, is the first redundant pixel value, is the second redundant pixel value, is the third redundant pixel value, is the 4th redundant pixel value;

[0053] Feature Matrix The expression is:

[0054] ,

[0055] In the formula, is the first feature pixel, is the second feature pixel, is the third feature pixel, is the fourth feature pixel;

[0056] The pixel value of each point in the image is between 0 and 255, and the image size is expressed as , after the size of , the sliding step length is of After convolution kernels, the image size becomes:

[0057] ,

[0058] According to the feature matrix At least one ultrasonic detection B-scan image sample is classified.

[0059] It should be noted that the pooling layer of the improved EfficientNet network is set as follows: In order to remove unimportant or repeated information in the features, a pooling layer for downsampling is added after the convolution layer:

[0060] ,

[0061] In the formula, is the input feature map of the Cth channel, is the pre-activation function, is the length of the pooling kernel, is the width of the pooling kernel, is the input characteristic length, is the number of input feature channels, is the input feature width, is the average value of the input features, is the weight of the Cth channel, is the input of the Cth channel;

[0062] Assume that is the parameter pooling layer error, then the error of this layer transferred to the previous layer for:

[0063] ,

[0064] in, , we can get the back propagation error formula of the parameter pooling layer:

[0065] ,

[0066] In the formula, is the error formula, is the pre-activation function;

[0067] Use the optimization algorithm to update the weight parameters of the parameter pooling layer. Suppose the lth layer is the parameter pooling layer, and the gradient of the layer parameters with respect to C is:

[0068] ,

[0069] ,

[0070] In the formula, is the deviation, is the step length, For the filter, is a hyperparameter, is the maximum value window, Fill the edges.

[0071] Step S102: input the at least one classification result and the actual rail screw hole crack recognition result corresponding to the at least one classification result into the improved GNN to train and generate a rail screw hole crack recognition model.

[0072] In this step, the expression of the alpha loss function of the improved GNN is:

[0073] ,

[0074] ,

[0075] In the formula, To improve the alpha loss function of GNN, is the weight coefficient, is the set of feature target values, is the target length of the feature sequence, is the feature bias, is the sample size;

[0076] ,

[0077] In the formula, , , , , , , are respectively the first alpha moment invariant, the second alpha moment invariant, the third alpha moment invariant, the fourth alpha moment invariant, the fifth alpha moment invariant, the sixth alpha moment invariant and the seventh alpha moment invariant, , , , , , , The 20th sample is the normalized center distance, the 2nd sample is the normalized center distance, the 11th sample is the normalized center distance, the 30th sample is the normalized center distance, the 12th sample is the normalized center distance, the 21st sample is the normalized center distance, and the 3rd sample is the normalized center distance.

[0078] In a specific embodiment, before inputting at least one classification result and the actual rail screw hole crack recognition result corresponding to the at least one classification result into the improved GNN, the hyperparameters of the improved GNN are reconstructed according to the hyperparameter reconstruction strategy, wherein the expression of the hyperparameter reconstruction strategy is:

[0079] ,

[0080] In the formula, To reconstruct the function, is the upper limit of the parameter, is the lower limit of the parameter, To reconstruct the upper limit, is the optimal reduction factor, is the number of parameters, is the scale factor;

[0081] ,

[0082] In the formula, is the reduction factor, To reduce the weight, is a time series, is the reduction function, To memorize information, To hide information, is the outlier coefficient;

[0083] ,

[0084] In the formula, is the actual value of the ratio, is the predicted value of the ratio, is the prediction error, is the number of predictions.

[0085] Step S103, obtaining a real-time ultrasonic detection B-scan image obtained by ultrasonically detecting the rail according to a preset Star strategy.

[0086] In this step, a double-track rail flaw detector is used to scan the problematic rails. According to the pulse signal received by the probe each time, the depth of the defect and the amplitude of the reflected signal are displayed. The horizontal axis represents the propagation time of the ultrasonic signal in the object. The propagation distance of the signal can be calculated to determine the location of the defect. Multiple scanning waveforms are combined to display the sound wave reflection of the longitudinal section of the rail. Therefore, the depth of the echo signal and its distribution on the section can be displayed. After processing, it is drawn into an image according to a certain ratio. However, its detection rate is too slow. Therefore, a Star strategy is proposed to improve the detection rate of the double-track rail ultrasonic flaw detector. The expression of the Star strategy is:

[0087] ,

[0088] ,

[0089] In the formula, is the detection rate, To detect the line speed, To detect the angular velocity, is the input speed, is the speed error, is the perturbation boundary, is the rate estimate, is the sliding mode boundary, is the target linear velocity, is the target angular velocity, To detect the first derivative of linear velocity, To detect the first derivative of angular velocity, To detect the second-order derivative of linear velocity, To detect the second derivative of angular velocity.

[0090] Step S104, inputting the real-time ultrasonic detection B-scan image into a rail screw hole crack recognition model, and the rail screw hole crack recognition model outputs a screw hole crack recognition result.

[0091] In summary, the method of the present application obtains a real-time ultrasonic detection B-scan image obtained by ultrasonically detecting the rail according to a preset STA strategy; the real-time ultrasonic detection B-scan image is input into a rail screw hole crack recognition model, and the rail screw hole crack recognition model outputs a screw hole crack recognition result, so that it has an autonomous recognition function, avoids the complex process of artificial feature extraction, improves the recognition rate, reduces the model size, and is conducive to engineering application.

[0092] See also Figure 2 , which shows a structural block diagram of a rail screw hole crack identification system of the present application.

[0093] like Figure 2 As shown, the rail screw hole crack identification system 200 includes a classification module 210, a training module 220, an acquisition module 230 and an output module 240.

[0094] The classification module 210 is configured to obtain at least one ultrasonic detection B-scan image sample, input the at least one ultrasonic detection B-scan image sample into the improved EfficientNet network, and the improved EfficientNet network classifies the at least one ultrasonic detection B-scan image sample according to the image contour features in the at least one ultrasonic detection B-scan image sample to obtain at least one classification result, wherein the improved EfficientNet network includes a regularization strategy, and the expression of the regularization strategy is:

[0095] ,

[0096] ,

[0097] ,

[0098] In the formula, is the input feature of the jump connection between the input node and the output node, is the output feature of each intermediate node in the top-down channel of the fusion layer, is the final output of the adaptive weighted fusion algorithm, To adjust the convolution of the input channel in the jump layer, is the input feature weight, It corresponds to the two groups of E-ELAN and MPConv modules from bottom to top in the fusion layer. is the parameter weight, is the mean of the eigenvectors, is the characteristic value, is the real boundary, is the feature vector after dimensionality reduction, is the feature vector, It corresponds to a group of Conv and BatchNorm modules from bottom to top in the fusion layer. is the number of superpixel blocks, is the total spectral distance, is the pixel space distance, is the spectral vector, A small amount to ensure the stability of weight values;

[0099] The training module 220 is configured to input the at least one classification result and the actual rail screw hole crack recognition result corresponding to the at least one classification result into the improved GNN to train and generate a rail screw hole crack recognition model;

[0100] An acquisition module 230 is configured to acquire a real-time ultrasonic detection B-scan image obtained by ultrasonically detecting the rail according to a preset Star strategy;

[0101] The output module 240 is configured to input the real-time ultrasonic detection B-scan image into a rail screw hole crack recognition model, and the rail screw hole crack recognition model outputs a screw hole crack recognition result.

[0102] It should be understood that Figure 2 Modules and references documented in Figure 1 Therefore, the operations and features described above for the method and the corresponding technical effects are also applicable to Figure 2 The modules in it will not be described in detail here.

[0103] In some other embodiments, the embodiments of the present invention further provide a computer-readable storage medium having a computer program stored thereon, wherein when the program instructions are executed by a processor, the processor is caused to execute the rail screw hole crack identification method in any of the above method embodiments;

[0104] As an implementation mode, the computer-readable storage medium of the present invention stores computer-executable instructions, and the computer-executable instructions are configured as follows:

[0105] At least one ultrasonic detection B-scan image sample is obtained, and the at least one ultrasonic detection B-scan image sample is input into an improved EfficientNet network, wherein the improved EfficientNet network classifies the at least one ultrasonic detection B-scan image sample according to image contour features in the at least one ultrasonic detection B-scan image sample to obtain at least one classification result, wherein the improved EfficientNet network includes a regularization strategy, and the expression of the regularization strategy is:

[0106] ,

[0107] ,

[0108] ,

[0109] In the formula, is the input feature of the jump connection between the input node and the output node, is the output feature of each intermediate node in the top-down channel of the fusion layer, is the final output of the adaptive weighted fusion algorithm, To adjust the convolution of the input channel in the jump layer, is the input feature weight, It corresponds to the two groups of E-ELAN and MPConv modules from bottom to top in the fusion layer. is the parameter weight, is the mean of the eigenvectors, is the characteristic value, is the real boundary, is the feature vector after dimensionality reduction, is the feature vector, It corresponds to a group of Conv and BatchNorm modules from bottom to top in the fusion layer. is the number of superpixel blocks, is the total spectral distance, is the pixel space distance, is the spectral vector, A small amount to ensure the stability of weight values;

[0110] Inputting the at least one classification result and the actual rail screw hole crack recognition result corresponding to the at least one classification result into the improved GNN to train and generate a rail screw hole crack recognition model;

[0111] Acquire a real-time ultrasonic testing B-scan image obtained by ultrasonic testing the rail according to a preset STAR strategy;

[0112] The real-time ultrasonic detection B-scan image is input into a rail screw hole crack recognition model, and the rail screw hole crack recognition model outputs a screw hole crack recognition result.

[0113] The computer-readable storage medium may include a program storage area and a data storage area, wherein the program storage area may store an operating system and an application required by at least one function; the data storage area may store data created according to the use of the rail screw hole crack identification system, etc. In addition, the computer-readable storage medium may include a high-speed random access memory, and may also include a memory, such as at least one disk storage device, a flash memory device, or other non-volatile solid-state storage device. In some embodiments, the computer-readable storage medium may optionally include a memory remotely arranged relative to the processor, and these remote memories may be connected to the rail screw hole crack identification system via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0114] Figure 3 is a schematic diagram of the structure of an electronic device provided by an embodiment of the present invention, such as Figure 3 As shown, the device includes: a processor 310 and a memory 320. The electronic device may also include: an input device 330 and an output device 340. The processor 310, the memory 320, the input device 330 and the output device 340 may be connected via a bus or other means. Figure 3 The example of the connection via bus is taken. The memory 320 is the computer-readable storage medium mentioned above. The processor 310 executes various functional applications and data processing of the server by running the non-volatile software programs, instructions and modules stored in the memory 320, that is, the rail screw hole crack identification method of the above method embodiment is implemented. The input device 330 can receive input digital or character information, and generate key signal input related to user settings and function control of the rail screw hole crack identification system. The output device 340 may include a display device such as a display screen.

[0115] The electronic device can execute the method provided by the embodiment of the present invention, and has the functional modules and beneficial effects corresponding to the execution method. For technical details not described in detail in this embodiment, please refer to the method provided by the embodiment of the present invention.

[0116] As an implementation mode, the electronic device is applied to a rail screw hole crack identification system and is used for a client, and includes: at least one processor; and a memory connected to the at least one processor in communication; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can:

[0117] At least one ultrasonic detection B-scan image sample is obtained, and the at least one ultrasonic detection B-scan image sample is input into an improved EfficientNet network, wherein the improved EfficientNet network classifies the at least one ultrasonic detection B-scan image sample according to image contour features in the at least one ultrasonic detection B-scan image sample to obtain at least one classification result, wherein the improved EfficientNet network includes a regularization strategy, and the expression of the regularization strategy is:

[0118] ,

[0119] ,

[0120] ,

[0121] In the formula, is the input feature of the jump connection between the input node and the output node, is the output feature of each intermediate node in the top-down channel of the fusion layer, is the final output of the adaptive weighted fusion algorithm, To adjust the convolution of the input channel in the jump layer, is the input feature weight, It corresponds to the two groups of E-ELAN and MPConv modules from bottom to top in the fusion layer. is the parameter weight, is the mean of the eigenvectors, is the characteristic value, is the real boundary, is the feature vector after dimensionality reduction, is the feature vector, It corresponds to a group of Conv and BatchNorm modules from bottom to top in the fusion layer. is the number of superpixel blocks, is the total spectral distance, is the pixel space distance, is the spectral vector, A small amount to ensure the stability of weight values;

[0122] Inputting the at least one classification result and the actual rail screw hole crack recognition result corresponding to the at least one classification result into the improved GNN to train and generate a rail screw hole crack recognition model;

[0123] Acquire a real-time ultrasonic testing B-scan image obtained by ultrasonic testing the rail according to a preset STAR strategy;

[0124] The real-time ultrasonic detection B-scan image is input into a rail screw hole crack recognition model, and the rail screw hole crack recognition model outputs a screw hole crack recognition result.

[0125] Through the description of the above implementation modes, those skilled in the art can clearly understand that each implementation mode can be implemented by means of software plus a necessary general hardware platform, or of course by hardware. Based on such an understanding, the above technical solution can essentially or in other words be embodied in the form of a software product that contributes to the prior art. The computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a magnetic disk, an optical disk, etc., and includes a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods of each embodiment or some parts of the embodiment.

[0126] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A rail screw hole crack identification method, characterized in that: include: At least one ultrasonic detection B-scan image sample is obtained, and the at least one ultrasonic detection B-scan image sample is input into an improved EfficientNet network, wherein the improved EfficientNet network classifies the at least one ultrasonic detection B-scan image sample according to image contour features in the at least one ultrasonic detection B-scan image sample to obtain at least one classification result, wherein the improved EfficientNet network includes a regularization strategy, and the expression of the regularization strategy is: ， ， ， In the formula, is the input feature of the jump connection between the input node and the output node, is the output feature of each intermediate node in the top-down channel of the fusion layer, is the final output of the adaptive weighted fusion algorithm, To adjust the convolution of the input channel in the jump layer, is the input feature weight, It corresponds to the two groups of E-ELAN and MPConv modules from bottom to top in the fusion layer. is the parameter weight, is the mean of the eigenvectors, is the characteristic value, is the real boundary, is the feature vector after dimensionality reduction, is the feature vector, It corresponds to a group of Conv and BatchNorm modules from bottom to top in the fusion layer. is the number of superpixel blocks, is the total spectral distance, is the pixel space distance, is the spectral vector, A small amount to ensure the stability of weight values; The improved EfficientNet network includes a PFE module, and the PFE module includes a 1-D convolution block, a 1-DDCN block, and a linear layer; The regression function expression of the PFE module is: ， ， In the formula, is the regression function, is the number of samples used for classification training, is a 0-1 label, To improve the output of the EfficientNet network, is the weight coefficient, To improve the weight matrix of the EfficientNet network, is the number of test samples, is the mean absolute percentage error, is the weighted mean absolute percentage error, is the weight parameter to balance the spatial distance and spectral distance, is the new cluster center, is the mean value of the pixels, is the number of pixels in the neighborhood; Inputting the at least one classification result and the actual rail screw hole crack recognition result corresponding to the at least one classification result into the improved GNN to train and generate a rail screw hole crack recognition model; A real-time ultrasonic detection B-scan image is obtained by ultrasonically detecting the rail according to a preset Star strategy, wherein the expression of the Star strategy is: ， ， In the formula, is the detection rate, To detect the line speed, To detect the angular velocity, is the input speed, is the speed error, is the perturbation boundary, is the rate estimate, is the sliding mode boundary, is the target linear velocity, is the target angular velocity, To detect the first derivative of linear velocity, To detect the first derivative of angular velocity, To detect the second-order derivative of linear velocity, To detect the second derivative of angular velocity; The real-time ultrasonic detection B-scan image is input into a rail screw hole crack recognition model, and the rail screw hole crack recognition model outputs a screw hole crack recognition result.

2. A rail screw hole crack identification method according to claim 1, characterized in that: The expression of the alpha loss function of the improved GNN is: ， ， In the formula, To improve the alpha loss function of GNN, is the weight coefficient, is the set of feature target values, is the target length of the feature sequence, is the feature bias, is the sample size; ， In the formula, , , , , , , are respectively the first alpha moment invariant, the second alpha moment invariant, the third alpha moment invariant, the fourth alpha moment invariant, the fifth alpha moment invariant, the sixth alpha moment invariant and the seventh alpha moment invariant, , , , , , , The 20th sample is the normalized center distance, the 2nd sample is the normalized center distance, the 11th sample is the normalized center distance, the 30th sample is the normalized center distance, the 12th sample is the normalized center distance, the 21st sample is the normalized center distance, and the 3rd sample is the normalized center distance.

3. A rail screw hole crack identification method according to claim 1, characterized in that: Before inputting the at least one classification result and the actual rail screw hole crack identification result corresponding to the at least one classification result into the improved GNN, the method further includes: The hyperparameters of the improved GNN are reconstructed according to the hyperparameter reconstruction strategy, wherein the expression of the hyperparameter reconstruction strategy is: ， In the formula, To reconstruct the function, is the upper limit of the parameter, is the lower limit of the parameter, To reconstruct the upper limit, is the optimal reduction factor, is the number of parameters, is the scale factor; ， In the formula, is the reduction factor, To reduce the weight, is a time series, is the reduction function, To memorize information, To hide information, is the outlier coefficient; ， In the formula, is the actual value of the ratio, is the predicted value of the ratio, is the prediction error, is the number of predictions.

4. A rail screw hole crack identification method according to claim 1, characterized in that: The classifying the at least one ultrasonic detection B-scan image sample according to the image contour feature in the at least one ultrasonic detection B-scan image sample to obtain at least one classification result comprises: Use the convolution kernel as a filter to find the contour features in the image and transform the image area matrix With the filter matrix The corresponding elements are multiplied and then added, the convolution step size is used to slide the area, and the next image area matrix is ​​convolved to obtain the feature matrix , where the image area matrix The expression is: In the formula, is the first channel of the corresponding feature map, is the second channel of the corresponding feature map, is the third channel of the corresponding feature map, is the 4th channel of the corresponding feature map, is the 5th channel of the corresponding feature map, is the sixth channel of the corresponding feature map, is the 7th channel of the corresponding feature map, is the 8th channel of the corresponding feature map, is the 9th channel of the corresponding feature map; Filter Matrix The expression is: ， In the formula, is the first redundant pixel value, is the second redundant pixel value, is the third redundant pixel value, is the 4th redundant pixel value; Feature Matrix The expression is: ， In the formula, is the first feature pixel, is the second feature pixel, is the third feature pixel, is the 4th feature pixel; The pixel value of each point in the image is between 0 and 255, and the image size is expressed as , after the size of , the sliding step length is of After convolution kernels, the image size becomes: ， According to the feature matrix The at least one ultrasonic detection B-scan image sample is classified.

5. A rail screw hole crack identification system, characterized in that: include: The classification module is configured to obtain at least one ultrasonic detection B-scan image sample, input the at least one ultrasonic detection B-scan image sample into the improved EfficientNet network, and the improved EfficientNet network classifies the at least one ultrasonic detection B-scan image sample according to the image contour features in the at least one ultrasonic detection B-scan image sample to obtain at least one classification result, wherein the improved EfficientNet network includes a regularization strategy, and the expression of the regularization strategy is: ， ， ， In the formula, is the input feature of the jump connection between the input node and the output node, is the output feature of each intermediate node in the top-down channel of the fusion layer, is the final output of the adaptive weighted fusion algorithm, To adjust the convolution of the input channel in the jump layer, is the input feature weight, It corresponds to the two groups of E-ELAN and MPConv modules from bottom to top in the fusion layer. is the parameter weight, is the mean of the eigenvectors, is the characteristic value, is the real boundary, is the feature vector after dimensionality reduction, is the feature vector, It corresponds to a group of Conv and BatchNorm modules from bottom to top in the fusion layer. is the number of superpixel blocks, is the total spectral distance, is the pixel space distance, is the spectral vector, A small amount to ensure the stability of weight values; The improved EfficientNet network includes a PFE module, and the PFE module includes a 1-D convolution block, a 1-DDCN block, and a linear layer; The regression function expression of the PFE module is: ， ， In the formula, is the regression function, is the number of samples used for classification training, is a 0-1 label, To improve the output of the EfficientNet network, is the weight coefficient, To improve the weight matrix of the EfficientNet network, is the number of test samples, is the mean absolute percentage error, is the weighted mean absolute percentage error, is the weight parameter to balance the spatial distance and spectral distance, is the new cluster center, is the mean value of the pixels, is the number of pixels in the neighborhood; A training module, configured to input the at least one classification result and an actual rail screw hole crack recognition result corresponding to the at least one classification result into the improved GNN, to train and generate a rail screw hole crack recognition model; The acquisition module is configured to acquire a real-time ultrasonic detection B-scan image obtained by ultrasonically detecting the rail according to a preset Star strategy, wherein the expression of the Star strategy is: ， ， In the formula, is the detection rate, To detect the line speed, To detect the angular velocity, is the input speed, is the speed error, is the perturbation boundary, is the rate estimate, is the sliding mode boundary, is the target linear velocity, is the target angular velocity, To detect the first derivative of linear velocity, To detect the first derivative of angular velocity, To detect the second-order derivative of linear velocity, To detect the second derivative of angular velocity; The output module is configured to input the real-time ultrasonic detection B-scan image into a rail screw hole crack recognition model, and the rail screw hole crack recognition model outputs a screw hole crack recognition result.

6. An electronic device, characterized in that: include: At least one processor, and a memory communicatively connected to the at least one processor, wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the method described in any one of claims 1 to 4.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the method according to any one of claims 1 to 4 is implemented.

Images