Rail damage identification model generation device, damage detection device and system

Abstract

The present invention discloses a rail damage identification model generation device, a damage detection device and a system. The damage identification model generation device includes: a first data parsing unit, which parses a single echo point group data file and extracts information including position, depth, channel and speed of the single echo point group; a reconstructed image generation unit, which generates a single reconstructed image according to the information including position, depth and channel of the single echo point group; a rail damage classification network, which trains and tests the single reconstructed image using a convolutional neural network and outputs a damage identification model. The image input to the convolutional neural network is an RGB reconstructed image with a direction formed by superimposing N channels. The present invention can solve the technical problem that the network input of the existing rail damage identification method directly adopts physical images or custom images, resulting in loss or error of channel information, and increases the difficulty and inaccuracy of machine learning.

Description

Technical Field

[0001] The present invention relates to the technical field of rail engineering machinery, and in particular to a rail damage identification model generating device, a damage detection device and a system for use in railway track ultrasonic detection. Background Art

[0002] At present, with the vigorous development of railway transportation, the mileage of railway construction in my country is increasing day by day. High-density, high-speed and heavy-load railway traffic have gradually become the main characteristics of modern railways. The structural performance and health status of railway rails are crucial to ensure the safety of railway transportation. Railway rails are prone to various fatigue cracks and wear during the long-term high-speed and high-density travel of trains. These early damages will continue to develop over time. After a certain period of time, they can develop into serious damage, resulting in rail breakage and causing destructive safety accidents. Therefore, it is of great significance to quickly and accurately detect and identify damage to rails.

[0003] At present, a damage identification mode of large-scale high-speed rail flaw detection vehicles is: software based on fixed rules and manual playback. Since the detection software's damage judgment principle is based on fixed decision tree rules, it cannot adapt well to all changes in damage during actual use, so there are technical problems such as high false alarm rate and low reference degree. After the software detection is completed, the software detection results need to be manually replayed, which not only reduces efficiency, but also easily affects the final judgment results due to factors such as manual fatigue and differences in personnel experience.

[0004] On the basis of software damage judgment, domestic and foreign experts and scholars have also proposed some improvement methods, such as rail damage recognition methods based on traditional machine learning algorithms, seeking to use the machine's self-learning ability to overcome the drawbacks of using fixed rules to identify damage. For example, China's invention application CN110246134A discloses a rail damage classification device based on a machine learning algorithm. However, traditional machine learning algorithms also have certain limitations, and the quality of their recognition effect depends on whether the feature extraction algorithm and the classifier classification model are appropriately selected. Usually, the selection of feature extraction algorithms and classification models mainly depends on human judgment, so it is one-sided, especially the feature extraction algorithm. If the extracted features are not enough to represent the sample, or too many features that the machine does not need to learn are extracted indiscriminately, the final recognition effect will be unsatisfactory.

[0005] Some scholars have improved traditional machine learning algorithms and applied deep learning algorithms to this field, using B-type image screenshots (i.e. physical images) as input to intelligent learning networks. However, if the positions and depths of some echo points in the image are the same, they will overlap, and it will be difficult for the machine to learn the channel information of the overlapping echo points, which will cause channel information loss. In addition, directly using physical images is easily affected by image resolution, which reduces the accuracy of machine learning.

[0006] In addition, some scholars have improved the above intelligent algorithm and proposed to change the input of the intelligent learning network from a physical image to a custom image, that is, using different pixel values ​​to represent different channels, such as: "0" represents a black background (no echo), "-2" represents R45 (backward 45° channel), and "2" represents F45 (forward 45° channel). However, it will appear that: if the R45 echo point and the F45 echo point appear at the same position and depth, the pixel value will be 0 after superposition, that is, the custom image processes the overlapping position of R45 and F45 into the background color (that is, the superimposed pixel value is the same as the defined pixel value), and the channel information of this position is lost. In addition, the machine only learns from the pixel superposition value, and cannot accurately learn which channels of echo point pixel values ​​are superimposed on the pixel superposition value. Summary of the invention

[0007] In view of this, the purpose of the present invention is to provide a rail damage identification model generation device, a damage detection device and a system to solve the technical problems existing in the existing rail damage identification method that the network input directly adopts physical images or custom images, resulting in channel information loss or errors, and increasing the difficulty and inaccuracy of machine learning.

[0008] In order to achieve the above-mentioned invention object, the present invention specifically provides a technical implementation scheme of a rail damage identification model generation device, the rail damage identification model generation device comprising:

[0009] A first data analysis unit analyzes a single echo point group data file to extract information of the single echo point group including position, depth, channel and speed;

[0010] A reconstructed image generation unit generates a single reconstructed image according to information including position, depth and channel of a single echo point group;

[0011] The rail damage classification network uses a convolutional neural network to train and test a single reconstructed image and output a damage recognition model.

[0012] The image input to the convolutional neural network is a directional RGB reconstructed image formed by superimposing N channels.

[0013] Furthermore, the rail damage identification model generating device also includes a data enhancement unit, which performs data enhancement processing on the single reconstructed image output by the reconstructed image generating unit, generates an enhanced single reconstructed image and outputs it to the rail damage classification network.

[0014] Furthermore, the rail damage classification network includes an input layer, a first convolution layer, a first pooling layer, a second convolution layer, a second pooling layer, a fully connected layer, a classifier and an output layer connected in sequence. The input layer is a single reconstructed image. The first convolution layer and the second convolution layer both contain a number of learnable convolution kernels, which use the convolution kernels to scan the input image or feature map of the previous layer from left to right and from top to bottom, perform convolution operations on them, and then obtain the feature map of the current layer after the activation function. The first pooling layer and the second pooling layer first divide the feature map after the convolution operation into non-overlapping areas, and then calculate the maximum value or average value in these areas respectively, and use this value to replace the value of the area. The fully connected layer combines multiple groups of feature maps formed after multiple convolution and pooling operations into a group of feature maps, and transmits them to the classifier. The output layer outputs the damage classification result.

[0015] Furthermore, a set proportion of each type of damage samples is randomly selected and input into the rail damage classification network for training the damage recognition model, and the remaining proportion of the damage samples of this type is used for testing the damage recognition model. The training process includes:

[0016] Forward propagation stage: a damage sample and its corresponding category label value are taken from the training sample set of a single reconstructed image, and the damage sample is input into the rail damage classification network. The convolution kernel weights in the rail damage classification network are randomly initialized, and then the corresponding actual output label value is calculated.

[0017] Backward propagation stage: Calculate the difference between the actual output label value and the ideal output label value, and then adjust the weight matrix by backpropagation according to the minimization error method.

[0018] After repeating the forward propagation stage and the backward propagation stage for all damaged samples in the training sample set, it is determined whether the preset index requirements are met. If not, it returns to the forward propagation stage and continues to iterate. If it is satisfied, the training ends.

[0019] After the training is completed, the test sample set is input into the rail damage classification network, and the output result of the rail damage classification network is obtained. The result is compared with the corresponding category label to test the classification accuracy of the rail damage classification network.

[0020] Furthermore, N channels correspond to N arrows with different directions and different RGB values, and the angle between the directions of two adjacent channels is 360° / N. When there are multiple echo points appearing at the same position and the same depth, the RGB values ​​are superimposed to form a new synthetic color that is different from the RGB values ​​of the N channels, and the direction arrows still retain the directions of their original channels.

[0021] The present invention also specifically provides a technical implementation scheme of a rail damage detection device, the rail damage detection device comprising:

[0022] The second data analysis unit analyzes the overall data file of a flaw detection operation and extracts information including position, depth, channel and speed of all echo point groups recorded in the file;

[0023] a reconstructed image segment generating unit, which forms a reconstructed image segment according to the information extracted by the second data parsing unit, and the whole data file generated by one flaw detection operation only forms one reconstructed image segment;

[0024] The reconstructed image segment cutting unit automatically cuts the reconstructed image segment by setting a sliding window with a certain step length, and outputs the cut reconstructed image segment set to a damage recognition model generated based on a convolutional neural network to obtain a preliminary damage recognition result; the image input to the convolutional neural network is an RGB reconstructed image with a direction formed by superimposing N channels;

[0025] The damage identification unit performs redundancy and overlap removal processing on the preliminary damage identification results to obtain the damage identification results for the flaw detection operation.

[0026] Furthermore, N channels correspond to N arrows with different directions and different RGB values, and the angle between the directions of two adjacent channels is 360° / N. When there are multiple echo points appearing at the same position and the same depth, the RGB values ​​are superimposed to form a new synthetic color that is different from the RGB values ​​of the N channels, and the direction arrows still retain the directions of their original channels.

[0027] The present invention also specifically provides a technical implementation scheme of a rail damage detection system, the rail damage detection system comprising: a vehicle-mounted terminal and a ground terminal. The vehicle-mounted terminal comprises a detection computer, a vehicle-mounted computer, a first display module and a first storage module. The ground terminal comprises a ground computer, a sample library server, a deep learning workstation, a second display module and a second storage module.

[0028] The detection computer shares the data file with the vehicle-mounted computer. After parsing the data file and identifying damage, the vehicle-mounted computer displays the identification result in the form of a B-type graph in the first display module and stores it in the first storage module. The vehicle-mounted computer sends the damage sample and the damage identification result to the sample library server.

[0029] The detection computer shares the data file with the ground computer. After parsing the data file and identifying damage, the ground computer displays the identification result in the form of a B-type diagram on the second display module and stores it in the second storage module. The ground computer transmits the damage sample and the damage identification result to the sample library server, and the sample library server implements the sample library management function.

[0030] The deep learning workstation is connected to the sample library server and optimizes the damage recognition model according to the selected samples in the sample library server. When the damage recognition model passes the test, the damage recognition model in the on-board computer and the ground computer is upgraded.

[0031] The damage recognition model is generated based on a convolutional neural network, and the image input into the convolutional neural network is a directional RGB reconstructed image formed by superimposing N channels.

[0032] The present invention also specifically provides another technical implementation scheme of a rail damage detection system, the rail damage detection system comprising: a vehicle-mounted terminal and a ground terminal. The vehicle-mounted terminal comprises a detection computer, a first display module and a first storage module. The ground terminal comprises a sample library server and a deep learning workstation.

[0033] After the detection computer analyzes the data file and identifies the damage, it displays the identification result in the form of a B-type image in the first display module and stores it in the first storage module. The detection computer sends the damage sample and the damage identification result to the sample library server, and the sample library server implements the sample library management function.

[0034] The deep learning workstation is connected to the sample library server and optimizes the damage recognition model according to the selected samples in the sample library server. When the damage recognition model passes the test, the damage recognition model in the detection computer is upgraded.

[0035] The damage recognition model is generated based on a convolutional neural network, and the image input into the convolutional neural network is a directional RGB reconstructed image formed by superimposing N channels.

[0036] The present invention also specifically provides a third technical implementation scheme of a rail damage detection system, the rail damage detection system comprising: a vehicle-mounted terminal and a ground terminal. The vehicle-mounted terminal comprises a detection computer, a vehicle-mounted computer, a first display module and a first storage module. The ground terminal comprises a sample library server and a deep learning workstation.

[0037] The detection computer shares data files with the vehicle-mounted computer. After parsing and identifying damages in the data files, the vehicle-mounted computer displays the identification results in the form of a B-type graph in the first display module and stores them in the first storage module. The vehicle-mounted computer sends the damage samples and damage identification results to the sample library server, which implements the sample library management function.

[0038] The deep learning workstation is connected to the sample library server and optimizes the damage recognition model according to the selected samples in the sample library server. When the damage recognition model passes the test, the damage recognition model in the on-board computer is upgraded.

[0039] The damage recognition model is generated based on a convolutional neural network, and the image input into the convolutional neural network is a directional RGB reconstructed image formed by superimposing N channels.

[0040] The present invention also specifically provides a fourth technical implementation scheme of a rail damage detection system, the rail damage detection system comprising: a vehicle-mounted terminal and a ground terminal. The vehicle-mounted terminal comprises a detection computer, and the ground terminal comprises a ground computer, a sample library server, a deep learning workstation, a second display module, and a second storage module.

[0041] The detection computer shares data files with the ground computer. After the ground computer parses the data files and identifies damage, it displays the identification results in the form of a B-type diagram in the second display module and stores them in the second storage module. The ground computer transmits the damage samples and damage identification results to the sample library server, and the sample library server implements the sample library management function. The deep learning workstation is connected to the sample library server, and optimizes the damage identification model according to the selected samples in the sample library server. When the damage identification model passes the test, the damage identification model in the ground computer is upgraded;

[0042] The damage recognition model is generated based on a convolutional neural network, and the reconstructed image input into the convolutional neural network is an RGB reconstructed image with a direction formed by superimposing N channels.

[0043] Furthermore, N channels correspond to N arrows with different directions and different RGB values, and the angle between the directions of two adjacent channels is 360° / N. When there are multiple echo points appearing at the same position and the same depth, the RGB values ​​are superimposed to form a new synthetic color that is different from the RGB values ​​of the N channels, and the direction arrows still retain the directions of their original channels.

[0044] By implementing the technical solutions of the rail damage identification model generation device, damage detection device and system provided by the present invention, the following beneficial effects are achieved:

[0045] (1) In the rail damage identification model generation device, damage detection device and system of the present invention, the image input to the convolutional neural network is an RGB reconstructed image with direction formed by superposition of N channels. When there are multiple echo points appearing at the same position and the same depth, the RGB values ​​are added respectively to form a new synthetic color different from the aforementioned N channel colors, and the direction arrows still retain the direction arrows of their respective original channels. It is not only not affected by the image resolution, but also can solve the technical problem that the network input of the existing rail damage identification method directly uses physical images or custom images, resulting in channel information loss or errors, increasing the difficulty and inaccuracy of machine learning, and greatly improving the accuracy and operability of machine learning;

[0046] (2) The rail damage identification model generation device, damage detection device and system of the present invention can solve the technical problem of the overall low feasibility of the existing system, realize offline and real-time detection of rail damage, and integrate intelligent recognition function, sample library management function and model optimization function. The deep learning workstation can use the selected samples in the sample library server to optimize the model at any time. When the test passes, the model can be stored, and the intelligent recognition module in the vehicle-mounted terminal or the ground terminal can be upgraded through the network or mobile storage device. At the same time, it can be further trained and optimized based on any model, and the detection effect can be continuously optimized with the advantage of low cost, freeing up manpower. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the prior art descriptions. Obviously, the drawings described below are only some embodiments of the present invention, and for ordinary technicians in this field, other embodiments can be obtained based on these drawings without creative work.

[0048] Figure 1 It is a schematic diagram of the principle of generating the B-type diagram of rail flaw detection;

[0049] Figure 2 yes Figure 1Schematic diagram of the B-type display interface of the middle channel F0 / R0 and FLER / RLER;

[0050] Figure 3 It is a schematic block diagram of the structure of a specific embodiment of the rail damage identification model generation device of the present invention;

[0051] Figure 4 It is a schematic block diagram of the principle of a rail damage classification network architecture in a specific embodiment of the rail damage identification model generation device of the present invention;

[0052] Figure 5 is a schematic diagram of a single-hole reconstructed image based on the device and system of the present invention;

[0053] Figure 6 It is a schematic block diagram of the structure of a specific embodiment of the rail damage detection device of the present invention;

[0054] Figure 7 It is a system structure block diagram of a specific embodiment of the rail damage detection system of the present invention;

[0055] Figure 8 It is a system structure block diagram of another specific embodiment of the rail damage detection system of the present invention;

[0056] Fig. 9 It is a system structure block diagram of a third specific embodiment of the rail damage detection system of the present invention;

[0057] Fig.10 It is a system structure block diagram of a fourth specific embodiment of the rail damage detection system of the present invention;

[0058] In the figure: 1-rail, 2-screw hole, 3-probe wheel, 4-ultrasonic chip, 10-vehicle terminal, 11-detection computer, 12-vehicle computer, 13-router, 14-first display module, 15-first storage module, 20-ground terminal, 21-ground computer, 22-sample library server, 23-deep learning workstation, 24-second display module, 25-second storage module, 100-rail damage classification network, 101-input layer, 102-first convolution layer, 103-first pooling layer, 104-second convolution layer Accumulation layer, 105-second pooling layer, 106-fully connected layer, 107-classifier, 108-output layer, 200-rail damage recognition model generation device, 201-first data parsing unit, 202-reconstructed image generation unit, 203-data enhancement unit, 204-damage recognition model, 300-rail damage detection device, 301-second data parsing unit, 302-reconstructed image segment generation unit, 303-reconstructed image segment cutting unit, 304-damage recognition unit, 400-rail damage detection system. DETAILED DESCRIPTION

[0059] 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 only 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.

[0060] It should be noted that when an element is referred to as being "fixed on" or "set on" another element, it can be directly on the other element or indirectly set on the other element; when an element is referred to as being "connected to" another element, it can be directly connected to the other element or indirectly connected to the other element.

[0061] It should be understood that the terms "length", "width", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside", etc., indicating the orientation or position relationship, are based on the orientation or position relationship shown in the drawings, and are only for the convenience of describing the present application and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be understood as a limitation on the present application.

[0062] In addition, the terms "first" and "second" are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the features. In the description of this application, "multiple" and "several" mean two or more, unless otherwise clearly and specifically defined.

[0063] It should be noted that the structures, proportions, sizes, etc. illustrated in the drawings of this specification are only used to match the contents disclosed in the specification for people familiar with this technology to understand and read, and are not used to limit the conditions under which this application can be implemented. Therefore, they have no substantive technical significance. Any structural modification, change in proportional relationship or adjustment of size should still fall within the scope of the technical content disclosed in this application without affecting the effects and purposes that can be achieved by this application.

[0064] As attached Figure 1 To Attachment Fig.10 As shown, specific embodiments of the rail damage identification model generation device, damage detection device and system of the present invention are given. The present invention is further described below in conjunction with the drawings and specific embodiments.

[0065] Example 1

[0066] Taking the detection of the screw hole 2 at the waist of the rail by the ultrasonic chip 4 of the 0° channel as an example, the generation principle of its B-type diagram (a display method of ultrasonic flaw detection cross-sectional diagram, which can display the longitudinal cross-sectional image of the inspected rail and indicate the approximate size and relative position of the ultrasonic reflector in the rail; wherein the horizontal axis represents the mileage position of the damage, and the vertical axis represents the depth of the damage under the rail surface) is as shown in the attached figure. Figure 1 As shown. Among them, as shown in A in the figure, it is the rail head position of the rail 1, as shown in B, it is the rail waist position, as shown in C, it is the rail bottom position, as shown in D1, D2, D3, D4, E1 and E2, they are the initial pulse, interface wave, damage echo, rail bottom reflection echo, 0° monitoring gate (MON) threshold and 0° bottom wave monitoring gate (LER) threshold respectively. The 0° ultrasonic chip in the probe wheel 3 emits an initial pulse after excitation, and enters the rail 1 through the surface of the rail. If there is no obstacle in the rail 1, the ultrasonic wave of the ultrasonic chip 4 will pass through the rail head and rail waist, reach the bottom of the rail, and be reflected by it. When there is an obstacle in the rail 1, such as a screw hole 2, the ultrasonic wave is reflected by it and received by the ultrasonic chip 4. After data processing, if the voltage amplitude of the damage echo exceeds the 0° monitoring gate threshold set by the system, the echo information will be stored and displayed in the form of echo points on the display interface of the B-type diagram, as shown in the attached figure. Figure 2 At the same time, due to the obstruction of screw hole 2, the voltage amplitude of the reflected echo from the rail bottom is less than the 0° bottom wave monitoring gate threshold, thus forming a bottom wave disappearance at the bottom of the rail, which will also be recorded and displayed on the B-type display interface, as shown in the attached figure. Figure 2 The middle channel FLER / RLER is shown.

[0067] In this embodiment, the background file corresponding to the B-type image is a data file in binary format, which records the position, depth, channel, speed and other information of the echo point on the B-type image. Among them, the channel information is a 24-bit binary number, each bit represents a channel, when the value of the bit is 1, it means that the corresponding channel has a damage signal, if multiple channels appear at the same position and depth, the 24-bit binary number corresponding to the position and depth is set to 1, and the specific channel definition is shown in Table 1 below. Since the B-type image only records 20 channels, the last 4 bits of the 24-bit binary number are not defined.

[0068] Table 1 Channel Definition

[0069]

[0070] As attached Figure 3 As shown, an embodiment of a rail damage identification model generation device 200 of the present invention specifically includes:

[0071] The first data parsing unit 201 parses the single echo (point) group data file in binary format to extract the information of the single echo (point) group including position, depth, channel and speed;

[0072] The reconstructed image generating unit 202 generates a single reconstructed image according to the information including the position, depth and channel of a single echo point group;

[0073] The rail damage classification network 100 uses a convolutional neural network to train and test a single reconstructed image and output a damage recognition model 204.

[0074] The image input to the convolutional neural network is a directional RGB reconstructed image composed of N channels. Each pixel of a color image is usually represented by three components: red (R), green (G), and blue (B), and each component is between (0, 255).

[0075] As a further improvement of this embodiment, the rail damage identification model generation device 200 further includes a data enhancement unit 203, which performs data enhancement processing (using data enhancement means such as position transformation and noise addition) on the single reconstructed image output by the reconstructed image generation unit 202, generates an enhanced single reconstructed image and outputs it to the rail damage classification network 100 for training and testing. Finally, a stable damage identification model 204 is obtained. The robustness of the damage identification model 204 can be further enhanced by data enhancement processing.

[0076] The purpose of designing the damage recognition model generation device 200 is that the traditional machine learning recognition algorithm uses artificially designed features with the following disadvantages: 1) damage recognition and judgment are time-consuming and labor-intensive, and a large number of features need to be used to achieve better results; 2) it is necessary to rely on professional knowledge, experience and intuition, which is difficult to achieve in some professional fields; 3) it is difficult to design multi-layer features. Therefore, the number and quality of features have become the bottleneck of traditional feature recognition algorithms. With the rapid development of deep learning algorithms, its outstanding advantage is that it can mine shallow to deep features of the target, which well meets the needs of machine self-learning features rather than artificially designed features. Among many deep neural network models, since convolutional neural networks (CNN) avoid complex feature extraction and show obvious advantages when the network input is a multi-dimensional image, the rail damage classification network 100 in this embodiment is designed based on CNN to obtain a damage recognition model 204.

[0077] The CNN architecture designed in this embodiment mainly includes the following structures: input layer 101, convolution layer, pooling layer, fully connected layer 106 and output layer 108. The input layer is a 224×224×3 RGB reconstructed image. Each convolution layer contains multiple learnable convolution kernels, which are used to scan the input image / feature map of the previous layer from left to right and from top to bottom, perform convolution operations on them, and then obtain the feature map of the current layer after the activation function. The pooling layer first divides the feature map after the convolution operation into non-overlapping areas, and then calculates the maximum value or average value in these areas respectively, and uses this value to replace the value of the area, so as to achieve the effect of aggregating features and reducing the amount of calculation. The fully connected layer combines multiple groups of feature maps formed after multiple convolution pooling into a group of feature maps, and transmits them to the softmax classifier. The output layer outputs the damage classification result. As shown in the attached Figure 4 As shown, as a typical specific embodiment of the present invention, the rail damage classification network 100 further includes an input layer 101, a first convolutional layer 102, a first pooling layer 103, a second convolutional layer 104, a second pooling layer 105, a fully connected layer 106, a classifier 107 and an output layer 108 connected in sequence. Among them, the input layer 101 is a single reconstructed image. The first convolutional layer 102 and the second convolutional layer 104 both contain a number of learnable convolution kernels, and the specific number of convolution kernels can be determined according to the test results of the damage recognition model 204. These convolution kernels are used to scan the input image or feature map of the previous layer from left to right and from top to bottom, perform convolution operations on them, and then obtain the feature map of the current layer after the activation function. The first pooling layer 103 and the second pooling layer 105 first divide the feature map after the convolution operation into non-overlapping areas, and then calculate the maximum value or average value in these areas respectively, and use this value to replace the value of the area. The fully connected layer 106 combines multiple sets of feature maps formed after multiple convolution and pooling operations into one set of feature maps, and transmits it to the classifier 107. The output layer 108 outputs the damage classification result. The number of convolutional layers and pooling layers is not limited to two layers, and can be three or more layers.

[0078] In this embodiment, N channels correspond to N arrows with different directions and different RGB values, and the angle between the directions of two adjacent channels is 360° / N. When there are multiple echo points appearing at the same position and the same depth, the RGB values ​​are superimposed to form a new synthetic color that is different from the RGB values ​​of the aforementioned N channels, and the direction arrows still retain the directions of their original channels. As a typical specific embodiment of the present invention, according to the data format of the channel and the input conditions of CNN, this embodiment defines the input image of the rail damage classification network 100 as an RGB reconstructed image formed by the superposition of 20 channels. Among them, 20 channels correspond to 20 arrows with different directions and different RGB values, and the angle between the directions of two adjacent channels is 18°. Taking the RGB reconstructed image of a single hole as an example, the relevant channels that constitute the single hole are F0, F45, R45 and R0. Therefore, its RGB reconstructed image can only be displayed at the position and depth corresponding to the echo points of the F0, F45, R45 and R0 channels. As shown in the attached figure Figure 5 As shown in the figure, a 20×20×3 color RGB reconstructed image is shown, in which the F0, F45, R45 and R0 channels correspond to 4 RGB color points with arrows of different specified directions (different line types are used to represent different colors for distinction). Figure 5 As shown in point A of , if two echo points appear at the same position and depth, their RGB values ​​are added to form a new composite color that is different from the 20 channel colors mentioned above, and the direction arrows still retain the direction arrows of their original channels. If multiple echo points appear at the same position and depth, the above processing method is the same.

[0079] Each RGB reconstructed image represents a specific type of sample. The sample types involved in this embodiment are shown in Table 2 below. Each sample type is composed of some or all channels in Table 1 above.

[0080] Table 2 Types of rail damage samples

[0081]

[0082]

[0083] The rail damage identification model generation device 200 described in this embodiment adopts an architecture (CNN architecture) based on the rail damage classification network 100, which includes an input layer 101, an output layer 108, and a certain number of convolutional layers, pooling layers and fully connected layers 106. According to the data format of the channel and the input conditions of CNN, this embodiment specifically defines the input image of the convolutional neural network as an RGB reconstructed image formed by the superposition of 20 channels. Among them, the 20 channels correspond to 20 arrows with different directions and different RGB values, and the angle between the directions of two adjacent channels is 18°. When there are multiple echo points appearing at the same position and the same depth, the RGB values ​​are added separately to form a new synthetic color that is different from the above 20 channel colors, and the direction arrows still retain the direction arrows of their respective original channels, which will not cause the loss of channel information.

[0084] A set proportion of each type of damage samples is randomly selected and input into the rail damage classification network 100 for training the damage identification model 204, and the remaining proportion of the damage samples of this type is used for testing the damage identification model 204. The sample set of this embodiment is specifically composed of the 13 types of samples in Table 2 above, and the number of samples in each type is the same, ensuring the balance of the data. As a typical specific embodiment of the present invention, 70% of the number of samples in each type is randomly selected for model training, and 30% is used for model testing. The training process further includes:

[0085] The first stage, the forward propagation stage: extract a damaged sample X and its corresponding class label value Y from the training sample set of a single reconstructed image P , the damage sample X is input into the rail damage classification network 100, the convolution kernel weights in the rail damage classification network 100 are randomly initialized, and then the corresponding actual output label value O is calculated. P .

[0086] The second stage, the back propagation stage: calculate the actual output label value O P and the ideal output label value Y P The difference is then calculated by back propagation according to the minimization error method to adjust the weight matrix.

[0087] In the third stage, after repeating the forward propagation stage and the backward propagation stage for all the damaged samples in the training sample set, it is determined whether the preset index requirements are met. If not, it returns to the first stage (the weights at this time are no longer randomly initialized, but use the values ​​of the previous stage) and continues to iterate. If it is satisfied, the training ends.

[0088] After the training is completed, the test sample set is input into the rail damage classification network 100, and the output result of the rail damage classification network 100 is obtained, and the result is compared with the corresponding category label to test the classification accuracy of the rail damage classification network 100.

[0089] The rail damage identification model generation device described in this embodiment adopts deep learning technology to realize rail damage identification, automatically performs spatial grouping and damage classification on the echo points in the data file, and can output information such as damage type, location, and size.

[0090] Example 2

[0091] In this embodiment, the rail damage detection device 300 actually performs damage identification on the binary overall data file generated by each operation of the flaw detection vehicle, rather than on a single echo (point) group data file. Figure 6 As shown, an embodiment of a rail damage detection device 300 of the present invention specifically includes:

[0092] The second data analysis unit 301 analyzes the overall binary data file of a flaw detection operation and extracts information including position, depth, channel and speed of all echo (point) groups recorded in the file;

[0093] The reconstructed image segment generating unit 302 forms an RGB reconstructed image segment according to the information extracted by the second data parsing unit 301. The entire binary data file generated by one flaw detection operation only forms one RGB reconstructed image segment.

[0094] The reconstructed image segment cutting unit 303 automatically cuts the RGB reconstructed image segments by setting a sliding window with a certain step length, and outputs the cut reconstructed image segment set to the damage recognition model 204 generated based on the convolutional neural network to obtain a preliminary damage recognition result; the image input to the convolutional neural network is an RGB reconstructed image with a direction formed by superimposing N channels;

[0095] The damage identification unit 304 removes redundancy and overlap on the preliminary damage identification results to obtain the damage identification results for the flaw detection operation. Because the sliding window is cut continuously, there are some echo (point) groups that do not display the complete RGB reconstructed images, and their identification results are meaningless and need to be removed. After the algorithm removes the redundant and overlapping initial identification results, the damage identification results for the operation can be obtained.

[0096] As a typical specific embodiment of the present invention, N channels correspond to N arrows with different directions and different RGB values, and the angle between the directions of two adjacent channels is 360° / N. When there are multiple echo points appearing at the same position and the same depth, the RGB values ​​are superimposed to form a new synthetic color that is different from the RGB values ​​of the aforementioned N channels, and the direction arrows still retain the directions of their original channels.

[0097] Example 3

[0098] As attached Figure 7 As shown, an embodiment of a rail damage detection system 400 based on deep learning of the present invention is used to realize offline detection of rail damage, and specifically includes: a vehicle-mounted terminal 10 and a ground terminal 20. The vehicle-mounted terminal 10 further includes a detection computer 11, a vehicle-mounted computer 12, a router 13, a first display module 14 and a first storage module 15. The ground terminal 20 further includes a ground computer 21, a sample library server 22, a deep learning workstation 23, a second display module 24 and a second storage module 25.

[0099] The detection computer 11 shares the binary data file with the vehicle-mounted computer 12 through the router 13. After parsing the data file and automatically identifying damage, the vehicle-mounted computer 12 displays the identification result in the form of a B-type diagram in the first display module 14 and stores it in the first storage module 15. The vehicle operator can mark and modify the erroneous or omitted damage identification results during playback. The vehicle-mounted computer 12 sends the damage sample and damage identification result to the sample library server 22 through the network or a mobile storage device.

[0100] The detection computer 11 shares binary format data files with the ground computer 21 through the network or mobile storage device. After the ground computer 21 parses the data file and automatically identifies damage, it displays the identification results in the form of a B-type diagram in the second display module 24 and stores them in the second storage module 25. The ground operator can mark and modify the erroneous or omitted damage identification results during playback. The ground computer 21 transmits the damage samples and damage identification results to the sample library server 22 through the network. The sample library server 22 is used to implement the management functions of the sample library including browsing, searching, filtering, classification, deletion, editing, and batch generation of data. The deep learning workstation 23 is connected to the sample library server 22 and can optimize the damage identification model 204 at any time according to the selected samples in the sample library server 22. When the damage identification model 204 passes the test, it can be stored, and the damage identification model 204 in the on-board computer 12 and the ground computer 21 can be upgraded through the network or mobile storage device. The deep learning workstation 23 is a high-performance computer with a rail damage identification model. During actual verification, if problems are found in the new model, the model optimization module can roll back to the specified historical model, and the model optimization module can further train and optimize the model based on any model.

[0101] As a typical specific embodiment of the present invention, the damage identification model 204 is generated based on a convolutional neural network, and the image input to the convolutional neural network is an RGB reconstructed image with directions formed by superimposing N channels. N channels correspond to N arrows with different directions and different RGB values, and the angle between the directions of two adjacent channels is 360° / N. When there are multiple echo points appearing at the same position and the same depth, the RGB values ​​are superimposed respectively to form a new synthetic color that is different from the RGB values ​​of the aforementioned N channels, and the direction arrows still retain the directions of their respective original channels.

[0102] The rail damage detection system 400 based on offline detection described in this embodiment realizes the model optimization function through the high-configuration deep learning workstation 23 and the model optimization module inside it. The model optimization module can use the selected samples in the management module of the sample library server 22 to optimize the model at any time. When the test passes, the model can be stored, and the damage identification model 204 in the on-board computer 12 and the ground computer 21 can be upgraded through the network or mobile storage device. The rail damage detection system 400 described in this embodiment can realize the three major functions of intelligent identification, sample library management and model optimization.

[0103] Example 4

[0104] When the detection speed of the damage recognition model reaches the maximum detection speed of the flaw detection vehicle, the damage recognition model 204 can be directly integrated into the detection computer 11 to achieve real-time detection of rail damage. Figure 8 As shown, another embodiment of a rail damage detection system 400 based on deep learning of the present invention is used to realize online detection of rail damage, and specifically includes: a vehicle-mounted terminal 10 and a ground terminal 20. The vehicle-mounted terminal 10 further includes a detection computer 11, a first display module 14 and a first storage module 15. The ground terminal 20 further includes a sample library server 22 and a deep learning workstation 23.

[0105] After the detection computer 11 analyzes the data file and automatically identifies the damage, it displays the identification result in the form of a B-type diagram in the first display module 14 and stores it in the first storage module 15. The on-board operator can mark and modify the erroneous or omitted damage identification results in real time / offline. The detection computer 11 sends the damage sample and the damage identification result to the sample library server 22 through the network or a mobile storage device.

[0106] The sample library server 22 is used to implement management functions of the sample library including browsing, searching, filtering, classifying, deleting, editing, and batch generation of data. The deep learning workstation 23 is connected to the sample library server 22 and can optimize the damage recognition model 204 at any time according to the selected samples in the sample library server 22. When the damage recognition model 204 passes the test, it can be stored and the damage recognition model 204 in the detection computer 11 can be upgraded through the network or a mobile storage device.

[0107] As a typical specific embodiment of the present invention, the damage identification model 204 is generated based on a convolutional neural network, and the image input to the convolutional neural network is an RGB reconstructed image with directions formed by superimposing N channels. N channels correspond to N arrows with different directions and different RGB values, and the angle between the directions of two adjacent channels is 360° / N. When there are multiple echo points appearing at the same position and the same depth, the RGB values ​​are superimposed respectively to form a new synthetic color that is different from the RGB values ​​of the aforementioned N channels, and the direction arrows still retain the directions of their respective original channels.

[0108] The rail damage detection system 400 based on offline detection described in this embodiment realizes the model optimization function through the high-configuration deep learning workstation 23 and the model optimization module inside it. The model optimization module can use the selected samples in the management module of the sample library server 22 to optimize the model at any time. When the test passes, the model can be stored, and the damage identification model 204 in the detection computer 11 can be upgraded through the network or mobile storage device. The rail damage detection system 400 described in this embodiment can realize the three functions of intelligent identification, sample library management and model optimization.

[0109] Example 5

[0110] As attached Fig. 9 As shown, the third embodiment of the rail damage detection system 400 based on deep learning of the present invention is used to realize offline detection of rail damage, and specifically includes: a vehicle-mounted terminal 10 and a ground terminal 20. The vehicle-mounted terminal 10 further includes a detection computer 11, a vehicle-mounted computer 12, a router 13, a first display module 14 and a first storage module 15. The ground terminal 20 further includes a sample library server 22 and a deep learning workstation 23.

[0111] The detection computer 11 shares the data file with the vehicle-mounted computer 12 through the router 13. After parsing the data file and automatically identifying damage, the vehicle-mounted computer 12 displays the identification result in the form of a B-type diagram in the first display module 14 and stores it in the first storage module 15. The vehicle-mounted computer 12 sends the damage sample and the damage identification result to the sample library server 22 through the network or a mobile storage device.

[0112] The sample library server 22 is used to implement management functions of the sample library including browsing, searching, filtering, classifying, deleting, editing, and batch generation of data. The deep learning workstation 23 is connected to the sample library server 22 and can optimize the damage recognition model 204 at any time according to the selected samples in the sample library server 22. When the damage recognition model 204 passes the test, it can be stored and the damage recognition model 204 in the vehicle-mounted computer 12 can be upgraded through the network or a mobile storage device.

[0113] As a typical specific embodiment of the present invention, the damage identification model 204 is generated based on a convolutional neural network, and the image input to the convolutional neural network is an RGB reconstructed image with directions formed by superimposing N channels. N channels correspond to N arrows with different directions and different RGB values, and the angle between the directions of two adjacent channels is 360° / N. When there are multiple echo points appearing at the same position and the same depth, the RGB values ​​are superimposed respectively to form a new synthetic color that is different from the RGB values ​​of the aforementioned N channels, and the direction arrows still retain the directions of their respective original channels.

[0114] The rail damage detection system 400 based on offline detection described in this embodiment realizes the model optimization function through the high-configuration deep learning workstation 23 and the model optimization module inside it. The model optimization module can use the selected samples in the management module of the sample library server 22 to optimize the model at any time. When the test passes, the model can be stored, and the damage identification model 204 in the on-board computer 12 can be upgraded through the network or mobile storage device. The rail damage detection system 400 described in this embodiment can realize the three functions of intelligent identification, sample library management and model optimization.

[0115] Example 6

[0116] As attached Fig.10 As shown, a fourth embodiment of a rail damage detection system 400 based on deep learning of the present invention is used to realize offline detection of rail damage, and specifically includes: a vehicle-mounted terminal 10 and a ground terminal 20. The vehicle-mounted terminal 10 further includes a detection computer 11, and the ground terminal 20 further includes a ground computer 21, a sample library server 22, a deep learning workstation 23, a second display module 24 and a second storage module 25.

[0117] The detection computer 11 shares data files with the ground computer 21 through the network or mobile storage device. After the ground computer 21 analyzes the data files and automatically identifies damage, it displays the identification results in the form of a B-type diagram in the second display module 24 and stores them in the second storage module 25. The ground computer 21 transmits the damage samples and damage identification results to the sample library server 22 through the network. The sample library server 22 is used to implement the management functions of the sample library including browsing, searching, filtering, classification, deletion, editing, and batch generation of data. The deep learning workstation 23 is connected to the sample library server 22 and can optimize the damage identification model 204 at any time according to the selected samples in the sample library server 22. When the damage identification model 204 passes the test, it can be stored, and the damage identification model 204 in the ground computer 21 can be upgraded through the network.

[0118] As a typical specific embodiment of the present invention, the damage identification model 204 is generated based on a convolutional neural network, and the image input to the convolutional neural network is an RGB reconstructed image with directions formed by superimposing N channels. N channels correspond to N arrows with different directions and different RGB values, and the angle between the directions of two adjacent channels is 360° / N. When there are multiple echo points appearing at the same position and the same depth, the RGB values ​​are superimposed respectively to form a new synthetic color that is different from the RGB values ​​of the aforementioned N channels, and the direction arrows still retain the directions of their respective original channels.

[0119] The rail damage detection system 400 based on offline detection described in this embodiment realizes the model optimization function through the high-configuration deep learning workstation 23 and the model optimization module inside it. The model optimization module can use the selected samples in the management module of the sample library server 22 to optimize the model at any time. When the test passes, the model can be stored, and the damage identification model 204 in the ground computer 21 can be upgraded through the network or mobile storage device. The rail damage detection system 400 described in this embodiment can realize the three functions of intelligent identification, sample library management and model optimization.

[0120] The rail damage identification model generation device, damage detection device and system described in the specific embodiment of the present invention change the damage identification mode based on traditional machine learning recognition to an identification mode based on a deep learning algorithm, and at the same time design a set of intelligent damage detection systems based on the algorithm, and achieve the purpose of enhancing the detection effect through continuous iterative updates. The specific embodiment of the present invention avoids the manual selection of features and achieves the purpose of machine self-learning by customizing the convolutional neural network architecture in the deep learning algorithm, freeing up manpower while ensuring recognition accuracy. The specific embodiment of the present invention also provides a technical implementation scheme for an offline rail damage detection system and a real-time rail damage detection system based on a deep learning algorithm, integrating intelligent recognition functions, sample library management functions and model optimization functions, while achieving a streamlined structure and strong operability.

[0121] By implementing the technical solutions of the rail damage identification model generation device, damage detection device and system described in the specific embodiments of the present invention, the following technical effects can be achieved:

[0122] (1) The rail damage identification model generation device, damage detection device and system described in the specific embodiment of the present invention, the image input to the convolutional neural network is an RGB reconstructed image with direction formed by superposition of N channels. When there are multiple echo points appearing at the same position and the same depth, the RGB values ​​are added respectively to form a new synthetic color different from the aforementioned N channel colors, and the direction arrows still retain the direction arrows of their respective original channels. It is not only not affected by the image resolution, but also can solve the technical problem that the network input of the existing rail damage identification method directly uses physical images or custom images, resulting in channel information loss or errors, increasing the difficulty and inaccuracy of machine learning, and greatly improving the accuracy and operability of machine learning;

[0123] (2) The rail damage identification model generation device, damage detection device and system described in the specific embodiments of the present invention can solve the technical problem that the overall feasibility of the existing system is not strong, realize offline and real-time detection of rail damage, and integrate intelligent recognition function, sample library management function and model optimization function. The deep learning workstation can use the selected samples in the sample library server to optimize the model at any time. When the test passes, the model can be stored, and the intelligent recognition module in the vehicle-mounted terminal or the ground terminal can be upgraded through the network or mobile storage device. At the same time, it can be further trained and optimized based on any model, and the detection effect can be continuously optimized with the advantage of low cost, freeing up manpower.

[0124] The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referenced to each other.

[0125] The above is only a preferred embodiment of the present invention and does not limit the present invention in any form. Although the present invention has been disclosed as above in the preferred embodiment, it is not used to limit the present invention. Any technician familiar with the art can make many possible changes and modifications to the technical solution of the present invention by using the methods and technical contents disclosed above without departing from the spirit and technical solution of the present invention, or modify it into an equivalent embodiment of equivalent changes. Therefore, any simple modification, equivalent replacement, equivalent change and modification made to the above embodiments according to the technical essence of the present invention without departing from the content of the technical solution of the present invention still falls within the scope of protection of the technical solution of the present invention.

Claims

1. A rail damage identification model generation device, characterized in that: include: A first data analysis unit (201) analyzes a single echo point group data file to extract information of the single echo point group including position, depth, channel and speed; A reconstructed image generating unit (202) generates a single reconstructed image according to information including position, depth and channel of a single echo point group; A rail damage classification network (100) uses a convolutional neural network to train and test a single reconstructed image and output a damage recognition model (204); the image input to the convolutional neural network is an RGB reconstructed image with a direction formed by superimposing N channels; N channels correspond to N arrows with different directions and different RGB values, and the angle between the directions of two adjacent channels is 360° / N; when there are multiple echo points appearing at the same position and the same depth, the RGB values ​​are superimposed to form a new synthetic color that is different from the RGB values ​​of the aforementioned N channels, and the direction arrows still retain the directions of their original channels.

2. The rail damage identification model generation device according to claim 1, characterized in that: The rail damage identification model generation device (200) further comprises a data enhancement unit (203), wherein the data enhancement unit (203) performs data enhancement processing on the single reconstructed image output by the reconstructed image generation unit (202), generates an enhanced single reconstructed image, and outputs the enhanced single reconstructed image to the rail damage classification network (100).

3. The rail damage identification model generation device according to claim 1 or 2, characterized in that: The rail damage classification network (100) comprises an input layer (101), a first convolution layer (102), a first pooling layer (103), a second convolution layer (104), a second pooling layer (105), a fully connected layer (106), a classifier (107) and an output layer (108) which are connected in sequence; the input layer (101) is a single reconstructed image; the first convolution layer (102) and the second convolution layer (104) both comprise a plurality of learnable convolution kernels, and the convolution kernels are used to sequentially scan the input image of the previous layer from left to right and from top to bottom. The first pooling layer (103) and the second pooling layer (105) first divide the feature map after the convolution operation into non-overlapping areas, and then calculate the maximum value or average value in these areas respectively, and use the value to replace the value of the area; the fully connected layer (106) combines multiple groups of feature maps formed after multiple convolution and pooling operations into one group of feature maps, and transmits it to the classifier (107); the output layer (108) outputs the damage classification result.

4. The rail damage identification model generation device according to claim 3, characterized in that: Randomly selecting a set proportion of each type of damage samples and inputting them into a rail damage classification network (100) for training the damage recognition model (204), and the remaining proportion of the damage samples of this type are used for testing the damage recognition model (204); The training process includes: Forward propagation stage: a damage sample and its corresponding category label value are taken from the training sample set of the single reconstructed image, and the damage sample is input into the rail damage classification network (100), the convolution kernel weights in the rail damage classification network (100) are randomly initialized, and then the corresponding actual output label value is calculated; Backward propagation stage: calculate the difference between the actual output label value and the ideal output label value, and then adjust the weight matrix by backpropagation according to the minimization error method; After repeating the forward propagation stage and the backward propagation stage for all damaged samples in the training sample set, determine whether the preset index requirements are met. If not, return to the forward propagation stage and continue iterating; if satisfied, the training ends; After the training is completed, the test sample set is input into the rail damage classification network (100), and the output result of the rail damage classification network (100) is obtained, and the result is compared with the corresponding category label to test the classification accuracy of the rail damage classification network (100).

5. A rail damage detection device, characterized in that: include: A second data analysis unit (301) analyzes the overall data file of a flaw detection operation and extracts information including position, depth, channel and speed of all echo point groups recorded in the file; A reconstructed image segment generating unit (302) is configured to form a reconstructed image segment based on the information extracted by the second data parsing unit (301), wherein the entire data file generated by one flaw detection operation only forms one reconstructed image segment; The reconstructed image segment cutting unit (303) automatically cuts the reconstructed image segment by setting a sliding window with a certain step length, and outputs the cut reconstructed image segment set to the damage recognition model (204) generated based on the convolutional neural network to obtain a preliminary damage recognition result; the image input to the convolutional neural network is an RGB reconstructed image with a direction formed by superimposing N channels; A damage identification unit (304) performs redundancy and overlap removal processing on the preliminary damage identification result to obtain a damage identification result for the flaw detection operation; N channels correspond to N arrows with different directions and different RGB values, and the angle between the directions of two adjacent channels is 360° / N; when there are multiple echo points appearing at the same position and the same depth, the RGB values ​​are superimposed to form a new synthetic color that is different from the RGB values ​​of the aforementioned N channels, and the direction arrows still retain the directions of their original channels.

6. A rail damage detection system, characterized in that: include: A vehicle-mounted terminal (10) and a ground terminal (20); the vehicle-mounted terminal (10) comprises a detection computer (11), a vehicle-mounted computer (12), a first display module (14) and a first storage module (15); the ground terminal (20) comprises a ground computer (21), a sample library server (22), a deep learning workstation (23), a second display module (24) and a second storage module (25); The detection computer (11) shares a data file with the on-board computer (12); after parsing the data file and identifying damage, the on-board computer (12) displays the identification result in the form of a B-type diagram on a first display module (14) and stores it in a first storage module (15); the on-board computer (12) sends the damage sample and the damage identification result to a sample library server (22); The detection computer (11) shares a data file with a ground computer (21); after parsing the data file and identifying damage, the ground computer (21) displays the identification result in the form of a B-type diagram on a second display module (24) and stores it in a second storage module (25); the ground computer (21) transmits the damage sample and the damage identification result to a sample library server (22); and the sample library server (22) implements a sample library management function; The deep learning workstation (23) is connected to the sample library server (22), and optimizes the damage recognition model (204) according to the selected samples in the sample library server (22); when the damage recognition model (204) passes the test, the damage recognition model (204) in the vehicle-mounted computer (12) and the ground computer (21) is upgraded; The damage recognition model (204) is generated based on a convolutional neural network, and the image input into the convolutional neural network is an RGB reconstructed image with directions formed by superimposing N channels; the N channels correspond to N arrows with different directions and different RGB values, and the angle between the directions of two adjacent channels is 360° / N; when there are multiple echo points appearing at the same position and the same depth, the RGB values ​​are superimposed respectively to form a new synthetic color that is different from the RGB values ​​of the aforementioned N channels, and the direction arrows still retain the directions of their respective original channels.

7. A rail damage detection system, characterized in that: include: A vehicle-mounted terminal (10) and a ground terminal (20); the vehicle-mounted terminal (10) comprises a detection computer (11), a first display module (14) and a first storage module (15); the ground terminal (20) comprises a sample library server (22) and a deep learning workstation (23); After the detection computer (11) parses the data file and identifies the damage, it displays the identification result in the form of a B-type diagram in a first display module (14) and stores it in a first storage module (15); the detection computer (11) sends the damage sample and the damage identification result to a sample library server (22), and the sample library server (22) implements a sample library management function; The deep learning workstation (23) is connected to the sample library server (22), and optimizes the damage recognition model (204) according to the selected samples in the sample library server (22); when the damage recognition model (204) passes the test, the damage recognition model (204) in the detection computer (11) is upgraded; The damage recognition model (204) is generated based on a convolutional neural network, and the image input into the convolutional neural network is an RGB reconstructed image with directions formed by superimposing N channels; the N channels correspond to N arrows with different directions and different RGB values, and the angle between the directions of two adjacent channels is 360° / N; when there are multiple echo points appearing at the same position and the same depth, the RGB values ​​are superimposed respectively to form a new synthetic color that is different from the RGB values ​​of the aforementioned N channels, and the direction arrows still retain the directions of their respective original channels.

8. A rail damage detection system, characterized in that: include: A vehicle-mounted terminal (10) and a ground terminal (20); the vehicle-mounted terminal (10) comprises a detection computer (11), a vehicle-mounted computer (12), a first display module (14) and a first storage module (15); the ground terminal (20) comprises a sample library server (22) and a deep learning workstation (23); The detection computer (11) shares a data file with the on-board computer (12); after parsing the data file and identifying damage, the on-board computer (12) displays the identification result in the form of a B-type diagram on a first display module (14) and stores it in a first storage module (15); the on-board computer (12) sends the damage sample and the damage identification result to a sample library server (22); the sample library server (22) implements a sample library management function; The deep learning workstation (23) is connected to the sample library server (22), and optimizes the damage recognition model (204) according to the selected samples in the sample library server (22); when the damage recognition model (204) passes the test, the damage recognition model (204) in the vehicle-mounted computer (12) is upgraded; The damage recognition model (204) is generated based on a convolutional neural network, and the image input into the convolutional neural network is an RGB reconstructed image with directions formed by superimposing N channels; the N channels correspond to N arrows with different directions and different RGB values, and the angle between the directions of two adjacent channels is 360° / N; when there are multiple echo points appearing at the same position and the same depth, the RGB values ​​are superimposed respectively to form a new synthetic color that is different from the RGB values ​​of the aforementioned N channels, and the direction arrows still retain the directions of their respective original channels.

9. A rail damage detection system, characterized in that: include: A vehicle-mounted terminal (10) and a ground terminal (20); the vehicle-mounted terminal (10) comprises a detection computer (11), and the ground terminal (20) comprises a ground computer (21), a sample library server (22), a deep learning workstation (23), a second display module (24) and a second storage module (25); The detection computer (11) shares a data file with a ground computer (21); after the ground computer (21) parses the data file and identifies damage, it displays the identification result in the form of a B-type diagram in a second display module (24) and stores it in a second storage module (25); the ground computer (21) transmits the damage sample and the damage identification result to a sample library server (22); the sample library server (22) implements a sample library management function; the deep learning workstation (23) is connected to the sample library server (22), and optimizes the damage identification model (204) according to the selected samples in the sample library server (22); when the damage identification model (204) passes the test, the damage identification model (204) in the ground computer (21) is upgraded; The damage recognition model (204) is generated based on a convolutional neural network, and the image input into the convolutional neural network is an RGB reconstructed image with directions formed by superimposing N channels; the N channels correspond to N arrows with different directions and different RGB values, and the angle between the directions of two adjacent channels is 360° / N; when there are multiple echo points appearing at the same position and the same depth, the RGB values ​​are superimposed respectively to form a new synthetic color that is different from the RGB values ​​of the aforementioned N channels, and the direction arrows still retain the directions of their respective original channels.

Images