Steel rail state evaluation method and device based on neural network

Abstract

The invention relates to the technical field of rail transit, in particular to a steel rail state evaluation method and device based on a neural network. A steel rail image is collected, a steel rail surface damage state is analyzed, and damage state grades are divided; classifying and expanding the data levels to obtain a training set; the ResNet50 pre-training model and the Yov5 detection head model are spliced and embedded by introducing a pyramid segmentation attention module to construct a new training model, and the new training model is defined as a YPResNet50 model; training a YPResNet50 model through the training set and optimizing the YPResNet50 model to obtain a state evaluation model; and importing a new steel rail surface image into the state evaluation model, and outputting a damage state grade. According to the method, the steel rail surface damage state is automatically classified and recognized through the constructed YPResNet50 model, and the fatigue state evolution process of the whole life cycle of the steel rail surface can be reflected through the five stages of recognition and classification. Therefore, the maintenance scheme is reasonably selected, and the phenomena that the maintenance cost is increased and unnecessary expenditure is wasted due to excessive detection and excessive maintenance are avoided.

Description

Technical Field

[0001] The present invention relates to the field of rail transportation technology, and in particular to a rail state assessment method and device based on a neural network. Background Art

[0002] With the in-depth research of artificial intelligence in the field of machine vision, many network models have been widely used due to their excellent performance. Among them, ResNet50, Parallel Spatial Attention Module and Yolo v5 neural network are important components of deep learning, and they play a key role in computer vision tasks.

[0003] ResNet50 is a deep convolutional neural network with a 50-layer deep network structure. Its core idea is to solve the gradient vanishing and representation bottleneck problems in the deep neural network training process by introducing residual blocks. Parallel Spatial Attention Module is a spatial attention module that can capture spatial context information in parallel at different levels of the network to enhance the representation ability of features. By introducing parallel spatial attention modules, the network can pay more attention to important spatial areas and ignore irrelevant noise information, thereby improving the discriminability of features and the performance of tasks. Yolo v5 is a real-time target detection algorithm that uses multi-scale feature fusion, anchor box prediction, loss function optimization and other technologies to improve the accuracy and speed of target detection.

[0004] With the continuous development of artificial intelligence technology, the theoretical research and practical application of track surface disease detection based on machine vision have also made certain progress. A variety of technical means have been developed for the detection of rail damage. In order to further study the laws of rail damage, scholars around the world have explored the automatic classification of rail damage. Through long-term research on cracks and vibrations of the track, the fatigue stage is judged, so as to achieve the purpose of making simple predictions of rail failures in advance. However, general research is limited to the final stage of failure, and cannot reflect the evolution law of the entire life cycle.

[0005] Based on the problems in the prior art, the present invention provides a rail condition assessment method and device based on a neural network. Summary of the invention

[0006] The purpose of the present invention is to provide a method and device for evaluating rail condition based on a neural network, so as to solve the technical problem that rail damage detection in the prior art cannot quickly identify the current rail condition at the positioning stage in the life cycle.

[0007] The technical solution of the present invention is: a rail state assessment method based on a neural network, comprising: collecting rail surface image data, analyzing the damage state of the rail surface, and dividing the damage state level; classifying the data according to the damage state level, and expanding the data in the classification database to obtain a training set; constructing a ResNet50 pre-training model and a Yo l ov5 detection head model, and introducing a pyramid segmentation attention module, and splicing and chimerizing the three to construct a new training model, which is defined as a YPResNet50 model; importing and training the YPResNet50 model through the training set, verifying the assessment accuracy of the YPResNet50 model, and optimizing and adjusting the training options and weights of the model to obtain a state assessment model; importing a new rail surface image into the state assessment model, and outputting the damage state level.

[0008] Preferably, the YOLOv5 detection head model includes a 1x1 convolution layer, a batch normalization layer, a ReLU layer, a fully connected layer, a Softmax layer and a classification layer.

[0009] Preferably, the PSA-Module model includes an input layer, a PSA-Module layer, a 1x1 convolution layer, a batch normalization layer, and a ReLU layer.

[0010] Preferably, the construction process of the YPResNet50 model is:

[0011] Get all the layers before the last convolutional layer of the ResNet50 pre-trained model and use them as input, concatenate and embed them into the PSA-Module layer in the pyramid segmentation attention module, and concatenate and embed the ReLU layer of the pyramid segmentation attention module into the 1x1 convolutional layer of the Yolo v5 detection head model;

[0012] When building the YPResNet50 model, the weights of the PSA-Module layer in the YOLOv5 detection head model and the pyramid segmentation attention module are randomly initialized.

[0013] Preferably, the damage status level includes five levels, namely: Note 1 indicates a brand new rail, Note 3 indicates a rail that needs to be polished and maintained, Note 4 indicates a rail that has a major safety hazard, and Note 5 indicates a rail that needs to be replaced immediately.

[0014] Preferably, the data of the training set is divided into two parts, namely a training data set and a validation data set, corresponding to the training phase and validation phase of the YPResNet50 model.

[0015] Preferably, a new data set is obtained by rotating the collected images by 90 degrees, 180 degrees and 270 degrees to expand the data set capacity and reduce the difference in the number of images in the classification database under different damage states.

[0016] Preferably, when collecting data, the most severely damaged area on the surface of each rail is captured as the object of collection and preservation.

[0017] Preferably, the size of the captured images is standardized, and the size of the images is uniformly changed to 224x224.

[0018] A rail state assessment device, used to implement the rail state assessment method based on neural network, comprising a camera, a control processor, a storage medium, an intelligent control screen, and a rechargeable energy storage power supply;

[0019] The camera, the alarm unit, the storage medium and the intelligent control screen are electrically connected to the control processor respectively, and the rechargeable energy storage power supply is used to power the camera, the control processor, the storage medium and the intelligent control screen; the camera is used to collect rail images, the logic program corresponding to the YPResNet50 model is stored in the storage medium or configured in the control processor, and the images collected by the camera are stored in the storage medium;

[0020] The control processor calls the logic program corresponding to the YPResNet50 model, analyzes and evaluates the damage status of the rail surface image collected by the camera, and displays the rail image and the evaluated damage status level on the intelligent control screen.

[0021] Compared with the prior art, the advantages of the present invention are:

[0022] Based on the ResNet50 model, the Yolo v5 detection head model and the PSA-Module model, the present invention constructs the YPResNet50 model to automatically classify and identify the damage status of the rail surface. The five stages of identification and classification can reflect the fatigue state evolution process of the rail surface throughout its life cycle, and make timely, rapid and effective judgments and evaluations, so as to reasonably select the maintenance plan, avoiding the increase of maintenance costs due to excessive detection and excessive maintenance, and wasting unnecessary expenses. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] The present invention will be further described below in conjunction with the accompanying drawings and embodiments:

[0024] Figure 1 It is a flowchart of the rail condition assessment method based on neural network according to the present invention;

[0025] Figure 2 A picture of the data collection environment and a schematic diagram of the collection result image of the present invention;

[0026] Figure 3 It is a schematic diagram of a data set of five classified images after multi-angle rotation according to the present invention;

[0027] Figure 4 A schematic diagram of the network structure of the YPResNet50 model of the present invention;

[0028] Figure 5 This is a schematic diagram of the training results of the ResNet50 model described in the present invention;

[0029] Figure 6 This is a schematic diagram of the training results of the YPResNet50 model described in the present invention;

[0030] Figure 7 A schematic diagram showing a comparison of training results of multiple existing models described in the present invention;

[0031] Figure 8 This is a schematic diagram of the automatic rail damage assessment results of YPResNet50 described in the present invention. DETAILED DESCRIPTION

[0032] The present invention is further described in detail below in conjunction with specific embodiments:

[0033] For ease of understanding, the application scenario of this application is first explained. The data collected by this application comes from the quality inspection area of ​​the Zhuozishan Line Workshop of the Jining Maintenance Section of the Hohhot Bureau of China Railway. The daily inspection of the operating lines within the jurisdiction of the workshop, and the collected rail status reflect the actual data set on site. As a heavy-duty line and heavy-duty 10,000-ton channel under the jurisdiction of the Zhuozishan Line Workshop, the Tangbao Line runs an average of more than 190 pairs of freight trains per day. The daily transportation is mainly coal transportation, with a large carrying capacity, and the daily impact and wear on the rails is relatively large. In addition, the terrain of the Zhuozishan Line is complex, including mountains, tunnels, heavy-duty railways, turnout joints, long and long downhill slopes, and many small-radius curves. The linear conditions of the Tangbao Line with many small-radius curves cause the train to bump and impact the rails due to turning inertia, crossing the turnout area, etc., which increases the diversity of rail diseases. It is impossible for humans to make a quick and comprehensive judgment, which greatly reduces the service life of the rails.

[0034] In this application, rail wear data is mainly measured and analyzed periodically using rail wear meters, ruler measurements, and visual discrimination methods. The system data set is formed by collecting photos of rail damage of different degrees over a long period of time, with no less than 1,000 photos of the full state of the rails; then, through the constructed YPResNet50 neural network model, according to the artificial intelligence algorithm, the photos of the fatigue state of the rail surface taken are automatically classified and compared with the images of the collected rail surface defects to conduct damage assessment, and the rail state is quickly, accurately, and efficiently assessed.

[0035] like Figure 1 As shown, a rail condition assessment method based on a neural network comprises the following steps:

[0036] Step 1: Collect rail surface image data, analyze and classify damage status levels.

[0037] The surface condition of the rails is photographed using high-definition cameras, mainly capturing the most severely damaged areas on the surface of each rail as the research object.

[0038] The captured images are then saved in the corresponding independent data sets. More than 1,000 rail damage data images are collected cumulatively, and the images that can reflect the rolling fatigue state of the rail surface are finally selected as the initial data set. Some images of the data collection environment and the collection result images are shown in the figure. Figure 2 shown.

[0039] Step 2: Data classification and expansion to obtain data sets.

[0040] Due to the influence of the horizontal rotation of the lens during the shooting process, the collected images are rotated 90 degrees, 180 degrees or even 270 degrees to form a new data set. The data set examples refer to the attached Figure 3 .

[0041] Long-term rail damage detection experts in the field conducted subjective evaluations on the collected data according to the Note1-5 classifications, and used the evaluation results as the reference standard for the constructed neural network evaluation model.

[0042] Specifically, Note 1 indicates a brand new rail, Note 2 indicates a rail that is not brand new but has no damage that could pose a safety hazard, Note 3 indicates a rail that has damage and needs grinding and maintenance, Note 4 indicates a rail that has a major safety hazard, and Note 5 indicates a rail that needs to be replaced immediately.

[0043] According to the scoring format of 1 to 5 points, the images of each category are placed into the corresponding category. Finally, the five data sets need to be labeled separately and created in the same data set.

[0044] In addition, it is worth mentioning that after the image rotation operation, the data set is expanded so that the database of each fatigue damage stage can be expanded with a large number of pictures to avoid too large differences in the number of images of each type of data.

[0045] Since the input data of the convolutional neural network must be in the form of a square whose side length is divisible by 2, and the size of most neural networks is 512x512, 256x256, or 224x224, the size of the acquired images is standardized and the image size is uniformly changed to 224x224 to accommodate most convolutional neural networks.

[0046] Step 3: Build the YPResNet50 model

[0047] A ResNet50 pre-trained model and a YOLOv5 detection head model are constructed, and a pyramid segmentation attention module is introduced. The YOLOv5 detection head model includes a 1x1 convolution layer, a batch normalization layer, a ReLU layer, a fully connected layer, a Softmax layer, and a classification layer; the PSA-Module model includes an input layer, a PSA-Module layer, a 1x1 convolution layer, a batch normalization layer, and a ReLU layer.

[0048] Get all the layers before the last convolutional layer of the ResNet50 pre-trained model as input, splice and embed them into the PSA-Module layer in the pyramid segmentation attention module, and splice the ReLU layer of the pyramid segmentation attention module into the 1x1 convolutional layer of the Yol ov5 detection head model. The network structure of the YPResNet50 model formed by splicing the three together is shown in the attached figure. Figure 4 As shown in the figure, when building the YPResNet50 model, the weights of the PSA-Module layer in the YOLOv5 detection head model and the pyramid segmentation attention module are randomly initialized, and then the training options are adjusted according to the data.

[0049] Step 4: Train, optimize, and verify the YPResNet50 model to obtain the state evaluation model.

[0050] The data of the training set is divided into two parts, namely the training data set and the validation data set, which correspond to the training stage and validation stage of the YPResNet50 model.

[0051] Import the training data set into the YPResNet50 model, select the training options, adjust the parameters and training options according to the actual situation, optimize the weights, and obtain a more mature evaluation model.

[0052] The evaluation accuracy of the evaluation model is verified based on the validation data set, and the training options and weights of the model are optimized and adjusted to obtain the state evaluation model.

[0053] The training of the neural network model can obtain the evaluation results of the fatigue state of the rail surface and give the accuracy. In order to verify the superiority of the YPResNet50 model, the accuracy of the training results is compared with the more common neural network results on the market. The training process of each network can be understood in detail through the pictures of the training process. Figure 5 and attached Figure 6 .

[0054] Attached Figure 5 and attached Figure 6It is automatically generated by the model during the training process of the convolutional neural network. The data in the figure can scientifically and objectively reflect the training results of CNN. In the training process of this embodiment, 70% of the data is selected as the training data set and 30% of the data is selected as the verification data set. The wavy curve with burrs at the top of the figure represents the training accuracy of CNN, the wavy curve with burrs at the bottom represents the loss during the training process, and the wavy dotted line represents the verification results of the two curves after the verification set is introduced into the network. Compared with the training result of 71.93% of the ResNet50 model, the accuracy of the YPResNet50 model is as high as 96.49%, and the evaluation accuracy is significantly improved by 24.56%.

[0055] In addition, the YPResNet50 model in this embodiment is compared with other types of neural networks, and the final statistical results are shown in the attached Figure 7 As shown in the table, it can be seen that AlexNet has the lowest verification accuracy of 28.07%, VGG19Net can be improved by 43.86%, and NASNetMobile's accuracy is improved to more than 90%.

[0056] The above training results show that the accuracy of the 50% neural network in the YPResNet model after training reaches the best accuracy of more than 95%, accurately evaluating the fatigue state of the rail surface and reaching a practical level. In order to avoid errors, the CNN in YPResNet50 was retrained, and the result exceeded 93%.

[0057] After improvements and the addition of more data, the YPResNet50 neural network model can be directly embedded in the software to form an application that objectively evaluates the damage state of the rail surface.

[0058] The third stage (Note 3) and the fourth stage (Note 4) of the rail surface fatigue (damage) state are relatively difficult to distinguish. Therefore, this example selects some pictures of Note 3 and Note 4 for verification. The verification results are as follows: Figure 8 shown.

[0059] Figure 8 Six examples are shown. The annotations above each image are the actual rail surface damage stage and the objective evaluation results of the trained YPResNet50. The percentage expresses the accuracy of the neural network's opinion of the results. The accuracy of the results of (d), (e), and (f) exceeds 99%, which is consistent with the identification results of actual medium- and long-term rail damage detection experts. Only one of the six images (example (c)) has an evaluation error, giving a result of 4, but it should actually be 3.

[0060] As a warning, the CNN also only gave an accuracy of 58%, and the main reason for this error is that YPResNet50 displaces the rust on the surface of the rails and regards it as a detachment and defect in the surface.

[0061] This error can be corrected by optimizing and retraining the network using a large number of similar photos. Therefore, it is foreseeable that the training results of YPResNet50 are integer values ​​formed by transfer learning within the data range of 1 to 5 points, and its accuracy is sufficient to objectively evaluate the fatigue stage of rail surface defects, and its results can be applied in practice. In addition, when performing the effect check C (Check), it was found that the accuracy of YPResNet50 was as high as 96.49%, and it could reach 99% after improving the data, which is much higher than other neural networks.

[0062] Step 5: Status assessment model identification application.

[0063] The new rail surface image is imported into the condition assessment model and the damage condition grade is output.

[0064] This application uses an improved neural network to classify and identify the surface fatigue damage state of the entire rail life cycle, divides the evolution process of the fatigue state of the rail surface throughout the life cycle into five stages, and automatically identifies the rail state through a machine under artificial intelligence big data, which can assist manual detection of the fatigue damage stage of the rail and the rail state to make timely, rapid and effective judgments and assessments, so as to reasonably select maintenance plans and avoid excessive detection and maintenance. Increased maintenance costs and unnecessary waste. The promotion and application of the technology in the field of railway transportation will make railway transportation safer and more reliable, save resources and improve efficiency, and inject new vitality into the development of the railway industry.

[0065] The present invention also provides a rail state assessment device for implementing the above-mentioned neural network-based rail state assessment method, including a camera, a control processor, a storage medium, an intelligent control screen, and a rechargeable energy storage power supply.

[0066] The camera, alarm unit, storage medium and intelligent control screen are electrically connected to the control processor respectively, and the rechargeable energy storage power supply is used to power the camera, control processor, storage medium and intelligent control screen; the camera is used to collect rail images, and the logic program corresponding to the YPResNet50 model is stored in the storage medium or configured in the control processor, and the images collected by the camera are stored in the storage medium. The control processor calls the logic program corresponding to the YPResNet50 model, analyzes and evaluates the damage status of the rail surface image collected by the camera, and displays the rail image and the assessed damage status level on the intelligent control screen.

[0067] The device has automatic detection and analysis functions. If a dangerous fault is found during the operation (defect detection), it can also issue a prompt alarm. The rail image and damage rating are displayed on the same screen, synchronously and in different zones. The image displays two areas and two parameter areas. It has good environmental adaptability and can work in a temperature environment of -30 to 50°C. It has a fully rainproof structure and is equipped with a GPS defect positioning function.

[0068] It has the function of automatically recording the flaw detection data (operation route, flaw detection speed, time, alarm status, sensitivity, etc.) throughout the whole process, forming an independent file every day, and can continuously record and save the flaw detection data for 60 days. It is equipped with microcomputer playback software. The data recorded by the flaw detection operation can be continuously played and analyzed through the playback software.

[0069] The above embodiments are only for illustrating the technical concept and features of the present invention, and their purpose is to enable people familiar with this technology to understand the content of the present invention and implement it accordingly, and they cannot be used to limit the scope of protection of the present invention. For those skilled in the art, it is obvious that the present invention is not limited to the details of the above exemplary embodiments, and the present invention can be implemented in other specific forms without departing from the spirit or basic features of the present invention. Therefore, no matter from which point of view, the embodiments should be regarded as exemplary and non-restrictive. The scope of the present invention is limited by the attached claims rather than the above description, and it is intended to include all changes within the meaning and scope of the equivalent elements of the claims in the present invention.

Claims

1. A rail condition assessment method based on neural network, characterized in that: include: Collect rail surface image data, analyze rail surface damage status, and classify damage status levels; Classify the data according to the damage status level, and expand the data in the classification database to obtain a training set; Build a ResNet50 pre-trained model and a Yolo v5 detection head model, and introduce a pyramid segmentation attention module. The three are spliced ​​and chimerized to build a new training model, which is defined as the YPResNet50 model; The YPResNet50 model is trained by importing the training set, the evaluation accuracy of the YPResNet50 model is verified, and the training options and weights of the model are optimized and adjusted to obtain a state evaluation model; The new rail surface image is imported into the condition assessment model, and the damage condition grade is output.

2. A rail condition assessment method based on neural network according to claim 1, characterized in that: The YOLOv5 detection head model includes a 1x1 convolution layer, a batch normalization layer, a ReLU layer, a fully connected layer, a Softmax layer, and a classification layer.

3. A rail condition assessment method based on neural network according to claim 2, characterized in that: The PSA-Module model includes an input layer, a PSA-Module layer, a 1x1 convolution layer, a batch normalization layer, and a ReLU layer.

4. A rail condition assessment method based on neural network according to claim 3, characterized in that: The construction process of the YPResNet50 model is: Get all the layers before the last convolutional layer of the ResNet50 pre-trained model and use them as input, concatenate and embed them into the PSA-Module layer in the pyramid segmentation attention module, and concatenate and embed the ReLU layer of the pyramid segmentation attention module into the 1x1 convolutional layer of the Yo lov5 detection head model; When building the YPResNet50 model, the weights of the PSA-Module layer in the YOLOv5 detection head model and the pyramid segmentation attention module are randomly initialized.

5. The rail condition assessment method based on neural network according to claim 1, characterized in that: The damage status level includes five levels, namely: Note1 indicates a brand new rail, Note3 indicates a rail that needs to be polished and maintained, Note4 indicates a rail with a major safety hazard, and Note5 indicates a rail that needs to be replaced immediately.

6. A rail condition assessment method based on neural network according to claim 1, characterized in that: The data of the training set is divided into two parts, namely a training data set and a validation data set, corresponding to the training stage and validation stage of the YPResNet50 model.

7. The rail condition assessment method based on neural network according to claim 1, characterized in that: By rotating the acquired images by 90, 180, and 270 degrees, new data sets were obtained to expand the data set capacity and reduce the difference in the number of images in the classification database under different damage states.

8. The rail condition assessment method based on neural network according to claim 1, characterized in that: When collecting data, the most severely damaged area on each rail surface is captured as the object of collection and preservation.

9. The rail condition assessment method based on neural network according to claim 1, characterized in that: The size of the captured images is standardized and the image size is uniformly changed to 224x224.

10. A rail condition assessment device, used to implement the rail condition assessment method based on neural network according to claims 1-9, characterized in that: Including camera, control processor, storage medium, intelligent control screen, and rechargeable energy storage power supply; The camera, the alarm unit, the storage medium and the intelligent control screen are electrically connected to the control processor respectively, and the rechargeable energy storage power supply is used to power the camera, the control processor, the storage medium and the intelligent control screen; the camera is used to collect rail images, the logic program corresponding to the YPResNet50 model is stored in the storage medium or configured in the control processor, and the images collected by the camera are stored in the storage medium; The control processor calls the logic program corresponding to the YPResNet50 model, analyzes and evaluates the damage status of the rail surface image collected by the camera, and displays the rail image and the evaluated damage status level on the intelligent control screen.

Images