Steel rail damage identification method and device based on source denoising, equipment and medium

Abstract

The invention relates to a steel rail damage identification method and device based on source denoising, equipment and a medium, and the method comprises the steps: obtaining original ultrasonic signals collected by a plurality of channels arranged in a to-be-identified steel rail region, and dividing the original ultrasonic signals into a plurality of groups of ultrasonic data segments according to channel types; detecting a noise point region in the ultrasonic data segment by adopting multi-threshold wavelet transform processing and envelope averaging processing to obtain a target noise point region; removing noise in the target noise point region to obtain target B-display image data; processing the target B-display image data based on a target deep learning network to obtain a steel rail damage position of the to-be-identified steel rail region; and outputting a steel rail damage identification result according to the steel rail damage position. According to the steel rail damage identification method based on source denoising provided by the invention, ultrasonic noise interference is removed from the source by adopting a wavelet transform and envelope averaging method, so that the damage detection difficulty is reduced, and the accuracy of steel rail damage detection is effectively improved.

Description

Technical Field

[0001] The present application relates to the fields of signal processing and computer vision technology, and in particular to a method, device, equipment and medium for identifying rail damage based on source denoising. Background Art

[0002] Automatically locating and identifying rail flaws using ultrasonic flaw detection signals is difficult to achieve with single-modal data analysis and processing techniques. Noise interference in ultrasonic signals not only increases the difficulty of flaw detection, but also causes more targets to be detected as flaws. This not only renders ultrasonic flaw detection ineffective, but also increases the workload. Summary of the Invention

[0003] Based on this, it is necessary to provide a rail damage identification method, device, equipment and medium based on source denoising, which can locate and classify internal rail damage through signal processing and deep learning technology, avoid noise interference to a large extent, and improve detection accuracy, in order to address the above technical problems.

[0004] In a first aspect, the present application provides a rail damage identification method based on source denoising, the method comprising:

[0005] Acquire original ultrasonic signals collected by multiple channels set in the rail area to be identified, and divide the original ultrasonic signals into multiple groups of ultrasonic data segments according to channel type;

[0006] Using multi-threshold wavelet transform processing and envelope averaging processing to detect noise point areas in the ultrasonic data segment to obtain target noise point areas;

[0007] removing noise in the target noise point area to obtain target B display image data;

[0008] Processing the target B-display image data based on a target deep learning network to obtain a rail damage location in the rail area to be identified, wherein the target deep learning network adopts a U2Net network structure;

[0009] Outputting a rail damage identification result according to the rail damage position.

[0010] In one embodiment, after the target deep learning network processes the target B-display image data to obtain the rail damage location of the rail area to be identified, the method further includes:

[0011] Acquire a damage image of the damaged position of the rail according to the output data of the target deep learning network;

[0012] Identifying the damage type of the damaged image based on a target classification network, wherein the target classification network adopts a ResNet network structure;

[0013] Output a rail damage identification result according to the rail damage location and the damage type.

[0014] In one embodiment, the detecting of the noise point region in the ultrasonic data segment by using multi-threshold wavelet transform processing and envelope averaging processing respectively to obtain the target noise point region includes:

[0015] For each set of ultrasonic echo data in the ultrasonic data segment, the multi-threshold wavelet transform is used to detect a noise point region of the ultrasonic echo data to obtain a first noise point region;

[0016] For ultrasonic echo data corresponding to the same detection target in multiple groups of ultrasonic data segments, detecting a noise point area of ​​the ultrasonic echo data using the envelope averaging process to obtain a second noise point area;

[0017] The target noise point area is obtained by combining the first noise point area and the second noise point area.

[0018] In one embodiment, dividing the original ultrasonic signal into multiple groups of ultrasonic data segments according to channel type includes:

[0019] The ultrasonic signal is divided into a plurality of groups of ultrasonic data segments, the number of which is the same as the number of channels, according to the channel type.

[0020] In one embodiment, the U2Net network structure includes at least 6 residual U-shaped network structures, and the target-based deep learning network processes the target B-display image data to obtain the rail damage location of the rail area to be identified, including:

[0021] Inputting the target B-display image data into the target deep learning network, and obtaining the output feature map of each residual U-shaped network structure;

[0022] A target grayscale image is obtained based on the output feature image, and the rail damage position is determined according to the target grayscale image.

[0023] In one embodiment, obtaining a damage image of the damaged position of the rail according to the output data of the target deep learning network includes:

[0024] Obtaining an output feature map of each residual U-shaped network structure, and obtaining a target grayscale image based on the output feature map;

[0025] Binarize the target grayscale image according to a preset grayscale image threshold to obtain a target binary image, wherein the circumscribed rectangular area with a pixel value of 1 in the target binary image is the area corresponding to the damaged position of the rail;

[0026] The image corresponding to the circumscribed rectangular area with pixels of 1 in the target binary image is intercepted as a mask, and is subjected to dot multiplication processing with the target B display image data to obtain the damaged image.

[0027] In a second aspect, the present application further provides a rail damage identification device based on source denoising, the device comprising:

[0028] a data acquisition module, configured to acquire original ultrasonic signals collected by multiple channels set in the rail area to be identified, and to divide the original ultrasonic signals into multiple groups of ultrasonic data segments according to channel types;

[0029] A noise detection module is used to detect the noise point area in the ultrasonic data segment by using multi-threshold wavelet transform processing and envelope averaging processing respectively, to obtain a target noise point area;

[0030] A noise removal module is used to remove noise in the target noise point area to obtain target B display image data;

[0031] a damage identification module, configured to process the target B-display image data based on a target deep learning network to obtain a rail damage location in the rail area to be identified, wherein the target deep learning network adopts a U2Net network structure;

[0032] The result output module is used to output the rail damage identification result according to the rail damage position.

[0033] In a third aspect, the present application further provides a computer device comprising a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the steps of the rail damage identification method based on source denoising described in the first aspect are implemented.

[0034] In a fourth aspect, the present application further provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the rail damage identification method based on source denoising described in the first aspect.

[0035] In a fifth aspect, the present application also provides a computer program product, which includes a computer program, and when the computer program is executed by a processor, it implements the steps of the rail damage identification method based on source denoising described in the first aspect.

[0036] In summary, the present application proposes a rail damage identification method, device, equipment and medium based on source denoising, including: obtaining original ultrasonic signals collected by multiple channels set in the rail area to be identified, and dividing the original ultrasonic signals into multiple groups of ultrasonic data segments according to the channel type; using multi-threshold wavelet transform processing and envelope averaging processing to detect the noise point area in the ultrasonic data segment to obtain the target noise point area; removing the noise in the target noise point area to obtain the target B-display image data; processing the target B-display image data based on the target deep learning network to obtain the rail damage position in the rail area to be identified; and outputting the rail damage identification result according to the rail damage position. The rail damage identification method based on source denoising provided by the present application adopts wavelet transform and envelope averaging methods to remove ultrasonic noise interference from the source, thereby reducing the difficulty of damage detection and effectively improving the accuracy of rail damage detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 FIG2 is an application environment diagram of a rail damage identification method based on source denoising in one embodiment;

[0038] Figure 2 1 is a flow chart of a rail damage identification method based on source denoising in one embodiment;

[0039] Figure 3 A schematic flow chart of a rail damage identification method based on source denoising in another embodiment;

[0040] Figure 4 A schematic diagram of a process for obtaining a target noise point area in one embodiment;

[0041] Figure 5 A schematic diagram of the U2Net network structure in one embodiment;

[0042] Figure 6 1 is a structural block diagram of a rail damage identification device based on source denoising in one embodiment;

[0043] Figure 7 FIG. 1 is a diagram showing the internal structure of a computer device in one embodiment. DETAILED DESCRIPTION

[0044] In order to make the purpose, technical solutions and advantages of this application more clear, the following further describes this application in detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain this application and are not intended to limit this application.

[0045] The rail damage identification method based on source denoising provided in the embodiment of the present application can be applied to Figure 1In the application environment shown, the terminal 102 communicates with the server 104 via a network. The data storage system can store data that the server 104 needs to process. The data storage system can be integrated with the server 104 or placed on the cloud or other network servers. The terminal 102 can be a rail ultrasonic flaw detection device with multi-channel detection capabilities. The rail ultrasonic flaw detection device in this embodiment can be used to detect damage defects in areas covering the rail head, rail waist, and rail bottom. The server 104 can be implemented as a standalone server or a server cluster consisting of multiple servers.

[0046] In related technologies, rail ultrasonic flaw detection signals are often accompanied by significant noise due to various factors during the data acquisition process, including mechanical vibration, gain adjustment differences, and environmental influences. The presence of noise—echo signals without actual physical meaning—constantly interferes with rail flaw detection operations, making it difficult to identify damage. In intelligent rail flaw detection systems, this chaotic and disordered noise not only impairs flaw detection but also prevents the system from accurately capturing damage information, as the interfering noise signals themselves are often identified as damage.

[0047] To solve the above problems, this application provides a rail damage identification method based on source denoising, which reduces the noise interference in the rail ultrasonic echo signal from the source, enhances the expression of the damaged target, and effectively reduces the difficulty of subsequent intelligent model building.

[0048] In one embodiment, Figure 2 As shown in the figure, a rail damage identification method based on source denoising is provided. Figure 1 The following steps are used as an example to illustrate the terminal in the figure:

[0049] S201 , obtaining original ultrasonic signals collected by multiple channels set in the rail area to be identified, and dividing the original ultrasonic signals into multiple groups of ultrasonic data segments according to channel types.

[0050] In this embodiment, the rail area to be identified refers to the rail area where the rail ultrasonic flaw detection equipment is installed. The rail ultrasonic flaw detection equipment provided in this embodiment includes an ultrasonic emission source and multiple data acquisition channels. The data acquisition channels can be ultrasonic echo signal acquisition devices such as ultrasonic probes. In actual application, the ultrasonic probe can be set to the corresponding detection position and detection depth based on the actual identification requirements of the rail area to be identified, and is used to detect the detection area, including the rail head, rail web, and related areas of the rail bottom.

[0051] In actual applications, after acquiring the raw ultrasonic signal, the ultrasonic detection signal can be split according to actual business needs. For example, the ultrasonic echo signal can be divided into rail head, rail web, and rail bottom signals based on the echo detection location. Alternatively, the ultrasonic echo signal can be divided into 20 different groups based on the number of channels of each ultrasonic probe. Furthermore, the ultrasonic echo signal can be split into forward and backward signals based on the detection direction.

[0052] In this embodiment, the original ultrasonic signal is divided into multiple groups of ultrasonic data segments according to the channel type, that is, the original ultrasonic signal is divided according to the number of channels, channel functions, or channel detection directions to obtain multiple groups of ultrasonic data segments. For example, assuming that the number of channels for collecting ultrasonic signals is 20, that is, the number of ultrasonic probes is 20, each ultrasonic probe collects corresponding ultrasonic data and summarizes them into the original ultrasonic signal. In this case, by dividing the original ultrasonic signal according to the number of channels, 20 groups of ultrasonic data segments can be obtained. If the original ultrasonic signal is divided according to the channel detection target, less than or equal to 20 groups of ultrasonic data segments may be obtained. It should be noted that the actual number of ultrasonic data segments in this embodiment can be determined based on the requirements of the actual application scenario and the specific definition of the channel type.

[0053] In one embodiment, as a preferred solution, S201 divides the original ultrasonic signal into multiple groups of ultrasonic data segments according to channel type, including:

[0054] The ultrasonic signal is divided into a plurality of ultrasonic data segments according to the channel type, the number of which is the same as the number of channels.

[0055] In this embodiment, in order to avoid the problem of missing or over-reporting of damage caused by noise signals, the original ultrasonic signal is directly converted into It is divided into 20 groups of different ultrasonic data segments, each group of ultrasonic data segments is recorded as Among them, each set of ultrasonic data segments Each of these signals represents the echo signal collected by an ultrasonic probe. Under normal conditions of ultrasonic flaw detection equipment (e.g., vehicle flaw detection), there will be no abnormalities such as missing channel data. Based on the above steps, effective noise signal removal is guaranteed.

[0056] S202 , using multi-threshold wavelet transform processing and envelope averaging processing to detect noise point regions in the ultrasonic data segment to obtain target noise point regions.

[0057] In this embodiment, the specific steps of multi-threshold wavelet transform processing include: first, determining the number of decomposition layers of the wavelet signal, decomposing the ultrasonic data segment to obtain multiple decomposition layers, each decomposition layer including multiple detail signals; second, determining the threshold of the detail signal under each decomposition layer, and performing threshold quantization processing on the detail signal; third, reconstructing the detail signal and the approximation signal after threshold processing to obtain a noise-reduced signal.

[0058] After determining the number of decomposition layers of the wavelet signal, the ultrasonic data segment can be decomposed using appropriate wavelet basis functions. Wavelet decomposition is a multi-scale analysis method that decomposes the signal into sub-bands of different frequencies. Each decomposition layer contains an approximation signal (low-frequency component) and multiple detail signals (high-frequency components). The approximation signal reflects the overall trend of the signal, while the detail signal contains high-frequency detail information of the signal at different scales. For example, for a 5-layer wavelet decomposition, the detail signal of the first layer mainly contains the highest-frequency component of the signal. As the number of decomposition layers increases, the frequency components contained in the detail signal gradually decrease.

[0059] After threshold quantization, the processed detail signals are obtained. These processed detail signals are combined with the approximation signal of the final layer and reconstructed using a wavelet reconstruction algorithm. Wavelet reconstruction is the inverse of wavelet decomposition, combining the approximation signal and detail signal at different scales into an approximate representation of the original signal. This reconstruction yields a de-noised signal. For example, for a five-layer wavelet decomposition, reconstructing the processed detail signals from layers 1 through 5 and the approximation signal of layer 5 yields a de-noised ultrasonic signal.

[0060] Wavelet transforms can be divided into soft threshold and hard threshold denoising methods based on the threshold selection. Hard threshold denoising retains coefficients in the wavelet decomposition detail signal whose absolute values ​​are greater than the threshold, while setting coefficients whose absolute values ​​are less than the threshold to 0. Soft threshold denoising is a smoothing method based on hard threshold denoising.

[0061] In this embodiment, the soft threshold denoising method can be mainly used to remove noise from each segment of ultrasonic echo signal. The wavelet soft threshold denoising expression is as follows:

[0062]

[0063] in, Expressed as a symbolic function, Indicates the current segment echo signal, represents the wavelet threshold of the echo signal, represents the signal after wavelet transformation. The calculation formula is as follows:

[0064]

[0065] in, For the The standard deviation of the layer detail signal, It is an empirical constant, usually between 2.5 and 3.5.

[0066] In this embodiment, wavelet transform is used to denoise the signal scales of different sliding window sizes. The noise filtering of the echo signal of each ultrasonic data segment can remove the scattered noise of the current signal, but does not affect the fluctuation of the echo signal. Therefore, the place where the echo signal is prominent is the noise point area in the ultrasonic data segment.

[0067] In this embodiment, the envelope averaging process includes the following steps: first, preprocessing the ultrasonic data segment to remove significant outliers and noise, and identifying the extreme values ​​within the ultrasonic data segment; second, interpolating the extreme values ​​using the cubic spline interpolation method, wherein cubic spline interpolation is performed on local maxima and local minima. Taking local maxima as an example, these maxima are used as interpolation nodes, and a smooth curve passing through these maxima is obtained using the cubic spline interpolation algorithm. This curve serves as a preliminary estimate of the upper envelope; third, smoothing algorithms such as moving average and Gaussian filtering are used to process the interpolated curve to obtain a smoothed curve; and fourth, determining the upper and lower envelopes based on the smoothed curve. The smoothed curve passing through the local maxima is the upper envelope. This curve, located above the signal, can reflect the overall upper bound of the signal. For example, in an ultrasonic echo signal, the upper envelope can represent the maximum amplitude variation trend of the echo signal. The smoothed curve passing through the local minima is the lower envelope. The lower envelope is located below the signal and reflects the overall lower bound of the signal. For example, in an echo signal, the lower envelope can indicate the minimum amplitude variation trend of the signal. Fifth, the values ​​of the corresponding points of the upper and lower envelopes are averaged to obtain the average envelope.

[0068] According to the average value of the upper and lower envelopes, the influence of some scattered noise can be removed. The cubic spline difference method is defined as: the echo signal In each small area The above is a cubic polynomial, denoted as ,in Indicates a time node. If at a time node Given ,satisfy ,but is a cubic spline interpolation function. After averaging the upper and lower envelopes of the ultrasonic echo signal, some prominent points with even upper and lower peaks can be weakened, effectively removing noise interference.

[0069] Based on the above steps, the noise point region in each ultrasonic data segment can be obtained by combining the denoising results of the wavelet transform and the upper and lower envelope averaging processing. In some embodiments, the denoised ultrasonic echo signal can also be directly obtained by combining the denoising results of the wavelet transform and the upper and lower envelope averaging processing.

[0070] S203, removing noise in the target noise point area to obtain target B display image data;

[0071] In this embodiment, when performing wavelet transform processing and envelope averaging processing on the ultrasonic data segment, the denoised ultrasonic echo signal can be directly obtained, and the corresponding target noise point area can be screened out, and then most of the scattered noise in the ultrasonic echo signal can be removed according to the target noise point area.

[0072] On this basis, this embodiment uses the unified diagram of a large rail flaw detection vehicle to plot the ultrasonic echo signal in each ultrasonic data segment as B-display image data.

[0073] S204, processing the target B display image data based on the target deep learning network to obtain the rail damage position of the rail area to be identified, wherein the target deep learning network adopts a U2Net network structure.

[0074] In this embodiment, the U2Net network is used to automatically detect and identify the significant areas of the B-display image data, and then the rail damage position is determined based on the significant areas in the B-display image data.

[0075] In this embodiment, the U2Net network is composed of two closely connected U structures, such as Figure 5 As shown in Figure 1, the residual U structure used here mixes receptive fields of different sizes, which can capture more contextual information from different scales. In addition, these residual U structure modules use pooling operations, so the depth of the network can be increased without significantly increasing the computational cost. In this embodiment, the U2Net network structure is composed of nested U-shaped residual modules, where En_k represents the Kth layer in the encoding stage and De_l represents the lth layer in the decoding stage.

[0076] In one embodiment, the U2Net model of the target deep learning network includes a total of 6 encoding layers, that is, continuous downsampling to obtain high-level information of the main target, and then the output feature map of the downsampling is spliced ​​through dimensionality, and the feature map size is normalized after 5 layers of upsampling decoding to obtain the pixel-level position information of the significant area, that is, the position information of the rail damage area of ​​the rail area to be identified, referred to as the rail damage location.

[0077] S205: Outputting rail damage identification results according to the rail damage location.

[0078] Noise is inevitable in the data collection process of rail ultrasonic damage automatic identification technology. The earlier the noise is removed, the easier it will be to reduce the difficulty of the analysis model and also significantly improve the data expression ability.

[0079] Based on the above steps, we use wavelet transform and envelope characteristics to analyze noise areas at the source of the data and reduce noise interference. We then use the U2Net model to detect significant areas in the B-display image data and determine the location of rail damage, achieving end-to-end localization of rail damage. By reducing noise interference, we significantly improve the accuracy of localization of rail damage.

[0080] In one embodiment, S204 processes the target B display image data based on the target deep learning network to obtain the rail damage position of the rail area to be identified, such as Figure 3 As shown in Figure 2, the rail damage identification method based on source denoising also includes:

[0081] S301, obtaining a damage image of the damaged position of the rail according to the output data of the target deep learning network.

[0082] S302: Identify the damage type of the damaged image based on a target classification network, wherein the target classification network adopts a ResNet network structure.

[0083] S303: Outputting rail damage identification results based on the rail damage location and damage type.

[0084] In this embodiment, the output data of the target deep learning network, i.e., the output data of the U2Net network, is a grayscale image, and this grayscale image includes a significant region, i.e., the region where the rail damage is located. In this embodiment, the grayscale image can be binarized by setting a grayscale threshold of 125 to determine the coordinates of the significant region in the grayscale image. Specifically, the grayscale image is binarized based on the grayscale threshold of 125. For pixels in the grayscale image with a grayscale value exceeding 125, the grayscale value of the pixel is directly set to 1. For pixels with a grayscale value less than 125, the grayscale value of the pixel is directly set to 0.

[0085] Through the above steps, the grayscale image is converted into a binary image. The binary image is used to cut out the bounding rectangle with pixels set to 1. The image corresponding to the cutout outer rectangle is then used as a mask and dot-multiplied with the target B-display image data to obtain a damage image of the rail damage location in the B-display image data.

[0086] In this embodiment, for each damage image of a rail damage location, this embodiment uses a ResNet classification model to classify the damage image to determine the specific type of rail damage in the damage image. The specific steps of using the ResNet classification model for classification include:

[0087] First, obtain the damage images corresponding to all rail damage locations. Second, crop and scale each damage image to unify the image scale, for example, unify the image scale size to 224. 224. Third, each scaled damage image is input into the ResNet-50 network to output each image label, which is the damage type corresponding to the damage image.

[0088] In one embodiment, a damage image of the damaged position of the rail is obtained based on the output data of the target deep learning network, including: obtaining the output feature map of each residual U-shaped network structure, and obtaining a target grayscale image based on the output feature map; binarizing the target grayscale image according to a preset grayscale image threshold to obtain a target binary image, wherein the circumscribed rectangular area with pixels of 1 in the target binary image is the area corresponding to the damaged position of the rail; intercepting the image corresponding to the circumscribed rectangular area with pixels of 1 in the target binary image as a mask, and performing dot multiplication processing with the target B-display image data to obtain a damage image.

[0089] In summary, this embodiment provides a rail damage identification method based on source denoising. This method not only denoises the raw ultrasonic data signal at the source, effectively reducing noise interference, but also enhances the expressiveness of useful signals and reduces the difficulty of model building. It also employs multiple convolutional neural networks (U2Net models) for feature extraction and fusion. By combining different features, the accuracy of the fine-grained classification model is improved, thereby reducing false positives and false negatives caused by human intervention. Furthermore, by using the ResNet-50 network to determine the final damage type, the influence of artificially designed features can be effectively avoided.

[0090] In one embodiment, Figure 4 As shown, S202 uses multi-threshold wavelet transform processing and envelope averaging processing to detect noise point areas in the ultrasonic data segment to obtain target noise point areas, including:

[0091] S401, for each set of ultrasonic echo data in the ultrasonic data segment, a multi-threshold wavelet transform is used to detect a noise point region of the ultrasonic echo data to obtain a first noise point region;

[0092] S402, for ultrasonic echo data corresponding to the same detection target in multiple groups of ultrasonic data segments, detecting a noise point region of the ultrasonic echo data using envelope averaging processing to obtain a second noise point region;

[0093] S403 , combining the first noise point area and the second noise point area to obtain a target noise point area.

[0094] In this embodiment, wavelet transform is used to denoise each segment of ultrasonic data after segmentation. Since wavelet transform uses different wavelet coefficients to continuously segment data, determining the signal frequency domain information under different wavelet coefficients is equivalent to continuously intercepting the ultrasonic signal with sliding windows of different sizes.

[0095] In actual application, since the ultrasonic data segments in this embodiment are divided according to channel types (number of channels), more noise will be left if only wavelet transform denoising is used.

[0096] Therefore, this embodiment further removes scattered noise based on the average value of the upper and lower envelopes for ultrasonic echo data corresponding to the same detection target in multiple groups of ultrasonic data segments, that is, targets of some channel combination echoes, such as joints, welds, etc.

[0097] Based on the above steps, this embodiment complements the denoising information provided by wavelet transformation and the denoising information provided by envelope averaging processing, which can effectively improve the denoising level of the ultrasonic echo signal, and thus play a constructive role in the construction and function of subsequent detection models and classification models, greatly improving the detection precision and accuracy of rail damage.

[0098] In one embodiment, Figure 5 As shown in the figure, the U2Net network structure includes at least 6 residual U-shaped network structures. Based on the target deep learning network, the target B-display image data is processed to obtain the rail damage location of the rail area to be identified, including:

[0099] The target B-display image data is input into the target deep learning network, and the output feature map of each residual U-shaped network structure is obtained; the target grayscale image is obtained based on the output feature map, and the rail damage location is determined according to the target grayscale image.

[0100] In this embodiment, reference Figure 5Each residual block of the U2Net network structure provided in this embodiment is a U-shaped network structure, and its specific structure is as follows: En_1 and De_1 both pass through two Conv+BN+Relu convolutional layers, then five 1 / 2 downsampling+Conv+BN+Relu convolutional layers, and then pass through one Conv+BN+Relu convolutional layer with an expansion factor of 2 to complete the downsampling step. After the upsampling passes through the Conv+BN+Relu convolutional layer, it is completed by five upsampling operations with an amplification factor of 2+Conv+BN+Relu. Among them, the downsampled feature map is added to the upsampled feature map step by step for transmission.

[0101] Both En_2 and De_2 complete the downsampling step after passing through two Conv+BN+Relu convolutional layers, four 1 / 2 downsampling+Conv+BN+Relu convolutional layers, and one Conv+BN+Relu convolutional layer with a dilation factor of 2. Upsampling is completed after passing through a Conv+BN+Relu convolutional layer and four more upsampling layers with a dilation factor of 2+Conv+BN+Relu. The downsampled feature maps are added to the upsampled feature maps step by step.

[0102] Both En_3 and De_3 complete the downsampling step after passing through two Conv+BN+Relu convolutional layers, three 1 / 2 downsampling+Conv+BN+Relu convolutional layers, and one Conv+BN+Relu convolutional layer with a dilation factor of 2. Upsampling is completed after passing through a Conv+BN+Relu convolutional layer and then three upsampling layers with a dilation factor of 2+Conv+BN+Relu. The downsampled feature maps are added to the upsampled feature maps step by step.

[0103] Both En_4 and De_4 complete the downsampling step after passing through two Conv+BN+Relu convolutional layers, two 1 / 2 downsampling+Conv+BN+Relu convolutional layers, and one Conv+BN+Relu convolutional layer with a dilation factor of 2. Upsampling is completed after passing through a Conv+BN+Relu convolutional layer and then two upsampling layers with a dilation factor of 2+Conv+BN+Relu. The downsampled feature maps are added to the upsampled feature maps step by step.

[0104] Both En_5 and De_5 complete the downsampling step after passing through two Conv+BN+Relu convolutional layers, one Conv+BN+Relu convolutional layer with a dilation factor of 2, one Conv+BN+Relu convolutional layer with a dilation factor of 4, and one Conv+BN+Relu convolutional layer with a dilation factor of 8. The upsampling operation is completed after passing through one Conv+BN+Relu convolutional layer with a dilation factor of 4, one Conv+BN+Relu convolutional layer with a dilation factor of 2, and one Conv+BN+Relu convolutional layer. The downsampled feature maps are added to the upsampled feature maps step by step.

[0105] En_6 is the bottom layer of the U-shaped structure. It first passes through two Conv+BN+Relu convolutional layers, then one Conv+BN+Relu convolutional layer with a dilation factor of 2, then one Conv+BN+Relu convolutional layer with a dilation factor of 4, and finally one Conv+BN+Relu convolutional layer with a dilation factor of 8 to complete the downsampling step. The upsampling operation is completed by passing through one Conv+BN+Relu convolutional layer with a dilation factor of 4, then one Conv+BN+Relu convolutional layer with a dilation factor of 2, and finally one Conv+BN+Relu convolutional layer. The downsampled feature map is added to the upsampled feature map at each level for transmission.

[0106] In this embodiment, the output feature map of each residual U-shaped structure block will output a target grayscale image after operations such as superposition, convolution, and activation (sigmoid), and the rail damage location can be determined based on the target grayscale image.

[0107] In summary, this embodiment provides a rail damage identification method based on source denoising. It employs multiple convolutional neural networks (U2Net models) for feature extraction and fusion. By combining different features, the accuracy of fine-grained classification models can be improved, thereby reducing manual false positives and false negatives. Furthermore, by using a ResNet-50 network to determine the final damage type, the influence of manually designed features can be effectively avoided.

[0108] It should be understood that, although the various steps in the flowcharts involved in the various embodiments described above are displayed in sequence according to the instructions of the arrows, these steps are not necessarily executed in sequence in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order restriction on the execution of these steps, and these steps can be executed in other orders. Moreover, at least a portion of the steps in the flowcharts involved in the various embodiments described above can include multiple steps or multiple stages, and these steps or stages are not necessarily executed and completed at the same time, but can be executed at different times, and the execution order of these steps or stages is not necessarily to be carried out in sequence, but can be executed in turn or alternately with other steps or at least a portion of steps or stages in other steps.

[0109] Based on the same inventive concept, embodiments of the present application also provide a source denoising-based rail damage identification device for implementing the aforementioned source denoising-based rail damage identification method. The solution provided by this device is similar to the solution described in the aforementioned method. Therefore, the specific limitations of one or more of the following embodiments of the source denoising-based rail damage identification device can be found in the above-mentioned limitations of the source denoising-based rail damage identification method and will not be further elaborated here.

[0110] In one embodiment, Figure 6 As shown, a rail damage identification device 600 based on source denoising is provided, comprising: a data acquisition module 610, a noise detection module 620, a noise removal module 630, a damage identification module 640 and a result output module 650, wherein:

[0111] The data acquisition module 610 is used to acquire original ultrasonic signals collected by multiple channels set in the rail area to be identified, and divide the original ultrasonic signals into multiple groups of ultrasonic data segments according to channel type;

[0112] The noise detection module 620 is configured to detect the noise point region in the ultrasonic data segment by using multi-threshold wavelet transform processing and envelope averaging processing respectively, to obtain a target noise point region;

[0113] A noise removal module 630 is configured to remove noise in the target noise point area to obtain target B display image data;

[0114] a damage identification module 640 for processing the target B-display image data based on a target deep learning network to obtain a rail damage location in the rail area to be identified, wherein the target deep learning network adopts a U2Net network structure;

[0115] The result output module 650 is used to output the rail damage identification result according to the rail damage position.

[0116] In one embodiment, the damage identification module 640 is further configured to obtain a damage image of the rail damage location based on the output data of the target deep learning network; identify the damage type of the damage image based on a target classification network, wherein the target classification network adopts a ResNet network structure;

[0117] The result output module 650 is further configured to output a rail damage identification result according to the rail damage location and the damage type.

[0118] In one embodiment, the noise detection module 620 is specifically used to detect the noise point area of ​​the ultrasonic echo data in each group of ultrasonic data segments using the multi-threshold wavelet transform processing to obtain a first noise point area; for the ultrasonic echo data corresponding to the same detection target in multiple groups of ultrasonic data segments, the envelope averaging processing is used to detect the noise point area of ​​the ultrasonic echo data to obtain a second noise point area; and the first noise point area and the second noise point area are combined to obtain the target noise point area.

[0119] In one embodiment, the data acquisition module 610 is configured to divide the ultrasonic signal into a plurality of ultrasonic data segments, the number of which is equal to the number of channels, according to the channel type.

[0120] In one embodiment, the damage identification module 640 is specifically used to input the target B-display image data into the target deep learning network and obtain the output feature map of each residual U-shaped network structure; obtain the target grayscale map based on the output feature map, and determine the rail damage position according to the target grayscale map.

[0121] In one embodiment, the damage identification module 640 is specifically used to obtain the output feature map of each residual U-shaped network structure, and obtain a target grayscale image based on the output feature map; binarize the target grayscale image according to a preset grayscale image threshold to obtain a target binary image, wherein the circumscribed rectangular area with pixels of 1 in the target binary image is the area corresponding to the damaged position of the rail; intercept the image corresponding to the circumscribed rectangular area with pixels of 1 in the target binary image as a mask, and perform dot multiplication processing with the target B-display image data to obtain the damage image.

[0122] In summary, this embodiment provides a rail damage identification device based on source denoising. It also employs multiple convolutional neural networks (U2Net models) for feature extraction and fusion. By combining different features, the accuracy of fine-grained classification models can be improved, thereby reducing manual false positives and false negatives. Furthermore, by using a ResNet-50 network to determine the final damage type, the influence of manually designed features can be effectively avoided.

[0123] Each module in the aforementioned source denoising-based rail damage identification device can be implemented in whole or in part through software, hardware, or a combination thereof. Each module can be embedded in or independent of a processor in a computer device in hardware form, or stored in a computer device memory in software form, so that the processor can call and execute the corresponding operations of each module.

[0124] In one embodiment, a computer device is provided. The computer device may be a terminal, and its internal structure diagram may be as follows: Figure 7 As shown. The computer device includes a processor, memory, an input / output interface, a communication interface, a display unit, and an input device. The processor, memory, and input / output interface are connected via a system bus, while the communication interface, display unit, and input device are connected to the system bus via the input / output interface. The processor of the computer device provides computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system and computer programs. The internal memory provides an environment for the operating system and computer programs stored in the non-volatile storage medium. The input / output interface of the computer device is used to exchange information between the processor and external devices. The communication interface of the computer device is used to communicate with external terminals via wired or wireless means, and the wireless means can be achieved via Wi-Fi, mobile cellular networks, NFC (near-field communication), or other technologies. When executed by the processor, the computer program implements a rail damage identification method based on source denoising. The display unit of the computer device is used to produce a visually visible image and can be a display screen, a projection device, or a virtual reality imaging device. The display screen can be a liquid crystal display screen or an electronic ink display screen, and the input device of the computer device can be a touch layer covering the display screen, or a button, trackball or touchpad set on the computer device casing, or an external keyboard, touchpad or mouse.

[0125] Those skilled in the art will understand that Figure 7The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than shown in the figure, or combine certain components, or have a different component arrangement.

[0126] In one embodiment, a computer device is provided, including a memory and a processor, wherein a computer program is stored in the memory, and when the processor executes the computer program, the following steps are implemented:

[0127] Acquire original ultrasonic signals collected by multiple channels set in the rail area to be identified, and divide the original ultrasonic signals into multiple groups of ultrasonic data segments according to channel types;

[0128] The noise point area in the ultrasonic data segment is detected by using multi-threshold wavelet transform processing and envelope averaging processing respectively, and the target noise point area is obtained;

[0129] Remove the noise in the target noise point area to obtain the target B display image data;

[0130] The target B-display image data is processed based on the target deep learning network to obtain the rail damage location of the rail area to be identified. The target deep learning network adopts the U2Net network structure;

[0131] Output rail damage identification results based on rail damage location.

[0132] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the following steps are implemented:

[0133] Acquire original ultrasonic signals collected by multiple channels set in the rail area to be identified, and divide the original ultrasonic signals into multiple groups of ultrasonic data segments according to channel types;

[0134] The noise point area in the ultrasonic data segment is detected by using multi-threshold wavelet transform processing and envelope averaging processing respectively, and the target noise point area is obtained;

[0135] Remove the noise in the target noise point area to obtain the target B display image data;

[0136] The target B-display image data is processed based on the target deep learning network to obtain the rail damage location of the rail area to be identified. The target deep learning network adopts the U2Net network structure;

[0137] Output rail damage identification results based on rail damage location.

[0138] In one embodiment, a computer program product is provided, comprising a computer program, which, when executed by a processor, implements the following steps:

[0139] Acquire original ultrasonic signals collected by multiple channels set in the rail area to be identified, and divide the original ultrasonic signals into multiple groups of ultrasonic data segments according to channel types;

[0140] The noise point area in the ultrasonic data segment is detected by using multi-threshold wavelet transform processing and envelope averaging processing respectively, and the target noise point area is obtained;

[0141] Remove the noise in the target noise point area to obtain the target B display image data;

[0142] The target B-display image data is processed based on the target deep learning network to obtain the rail damage location of the rail area to be identified. The target deep learning network adopts the U2Net network structure;

[0143] Output rail damage identification results based on rail damage location.

[0144] Those skilled in the art will appreciate that all or part of the processes in the above-mentioned embodiments can be implemented by instructing the relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the above-mentioned embodiments. In particular, any reference to memory, database, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM). The databases involved in the various embodiments provided herein may include at least one of a relational database and a non-relational database. Non-relational databases may include, but are not limited to, distributed databases based on blockchains. The processors involved in the various embodiments provided herein may be, but are not limited to, general-purpose processors, central processing units (CPUs), graphics processing units (GPUs), digital signal processors (DSPs), programmable logic devices (PLDs), data processing logic devices based on quantum computing, and the like.

[0145] The technical features of the above embodiments can be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0146] The above-described embodiments merely represent several implementation methods of the present application. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present application. It should be noted that a person of ordinary skill in the art may make various modifications and improvements without departing from the spirit of the present application, and these modifications and improvements fall within the scope of protection of the present application. Therefore, the scope of protection of the present application shall be determined by the appended claims.

Claims

1. A rail damage identification method based on source denoising, characterized in that: The method comprises: Acquire original ultrasonic signals collected by multiple channels set in the rail area to be identified, and divide the original ultrasonic signals into multiple groups of ultrasonic data segments according to channel type; Using multi-threshold wavelet transform processing and envelope averaging processing to detect noise point areas in the ultrasonic data segment to obtain target noise point areas; removing noise in the target noise point area to obtain target B display image data; Processing the target B-display image data based on a target deep learning network to obtain a rail damage location in the rail area to be identified, wherein the target deep learning network adopts a U2Net network structure; Outputting a rail damage identification result according to the rail damage position.

2. The method according to claim 1, characterized in that After the target deep learning network is used to process the target B-display image data to obtain the rail damage position of the rail area to be identified, the method further includes: Acquire a damage image of the damaged position of the rail according to the output data of the target deep learning network; Identifying the damage type of the damaged image based on a target classification network, wherein the target classification network adopts a ResNet network structure; Output a rail damage identification result according to the rail damage location and the damage type.

3. The method according to claim 1, characterized in that The method of detecting the noise point area in the ultrasonic data segment by using multi-threshold wavelet transform processing and envelope averaging processing to obtain the target noise point area includes: For each set of ultrasonic echo data in the ultrasonic data segment, the multi-threshold wavelet transform is used to detect a noise point region of the ultrasonic echo data to obtain a first noise point region; For ultrasonic echo data corresponding to the same detection target in multiple groups of ultrasonic data segments, detecting a noise point area of ​​the ultrasonic echo data using the envelope averaging process to obtain a second noise point area; The target noise point area is obtained by combining the first noise point area and the second noise point area.

4. The method according to claim 3, characterized in that The dividing the original ultrasonic signal into multiple groups of ultrasonic data segments according to channel type includes: The ultrasonic signal is divided into a plurality of groups of ultrasonic data segments, the number of which is the same as the number of channels, according to the channel type.

5. The method according to claim 2, characterized in that The U2Net network structure includes at least 6 residual U-shaped network structures. The target-based deep learning network processes the target B-display image data to obtain the rail damage position of the rail area to be identified, including: Inputting the target B-display image data into the target deep learning network, and obtaining the output feature map of each residual U-shaped network structure; A target grayscale image is obtained based on the output feature image, and the rail damage position is determined according to the target grayscale image.

6. The method according to claim 5, characterized in that The step of obtaining a damage image of the damaged position of the rail according to the output data of the target deep learning network includes: Obtaining an output feature map of each residual U-shaped network structure, and obtaining a target grayscale image based on the output feature map; Binarize the target grayscale image according to a preset grayscale image threshold to obtain a target binary image, wherein the circumscribed rectangular area with a pixel value of 1 in the target binary image is the area corresponding to the damaged position of the rail; The image corresponding to the circumscribed rectangular area with pixels of 1 in the target binary image is intercepted as a mask, and is subjected to dot multiplication processing with the target B display image data to obtain the damaged image.

7. A rail damage identification device based on source denoising, characterized in that: The device comprises: a data acquisition module, configured to acquire original ultrasonic signals collected by multiple channels set in the rail area to be identified, and to divide the original ultrasonic signals into multiple groups of ultrasonic data segments according to channel types; A noise detection module is used to detect the noise point area in the ultrasonic data segment by using multi-threshold wavelet transform processing and envelope averaging processing respectively, to obtain a target noise point area; A noise removal module is used to remove noise in the target noise point area to obtain target B display image data; a damage identification module, configured to process the target B-display image data based on a target deep learning network to obtain a rail damage location in the rail area to be identified, wherein the target deep learning network adopts a U2Net network structure; The result output module is used to output the rail damage identification result according to the rail damage position.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the rail damage identification method based on source denoising according to any one of claims 1 to 6 are implemented.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the rail damage identification method based on source denoising according to any one of claims 1 to 6 are implemented.

10. A computer program product comprising a computer program, characterized in that When the computer program is executed by a processor, the steps of the rail damage identification method based on source denoising according to any one of claims 1 to 6 are implemented.

Images