A rail damage identification method based on ultrasonic B-display data

Abstract

A rail damage identification method for ultrasonic B-display data comprises the following steps: 1) collecting damage B-display data by using an ultrasonic flaw detection vehicle; 2) drawing a damage B-display image to obtain a cropped damage B-display image; 3) setting a label for the cropped damage B-display image to establish a typical damage B-display image sample library; 4) training an image classification model based on the damage B-display image samples to construct a damage identification model; 5) performing ultrasonic information combination division on the damage B-display data of step 1 to obtain ultrasonic information combination data; 6) dividing the ultrasonic information combination data into damage ultrasonic information groups to realize complete damage data extraction; 7) redrawing the damage B-display image, wherein the redrawn damage B-display image only contains one damage ultrasonic information group; 8) using the redrawn B-display image as the input of a damage identification model to perform intelligent damage identification; the method has the characteristics of improving damage judgment efficiency and reducing false alarm rate and missed alarm rate.

Description

Technical Field

[0001] The invention belongs to the technical field of rail damage identification, and in particular relates to a rail damage identification method for ultrasonic B-display data. Background Art

[0002] Rails carry trains and are an important part of the railway transportation system. Their service status directly affects the safety of train operations. Rails will be damaged due to squeezing and impact by trains during long-term service. If these damages are not repaired in time, the damage will accumulate and expand, causing the rails to be seriously damaged. When a train passes through a severely damaged line section, it is very likely to cause the rails to break, leading to serious railway traffic accidents. Therefore, it is of great significance to inspect and maintain rail damage in a timely manner to ensure the normal and reliable service of the rails. my country mainly uses ultrasonic flaw detection vehicles to detect damage to railways.

[0003] There are two types of detection data from ultrasonic flaw detection vehicles: A display data and B display data. The amount of A display data is huge. Due to technical reasons, A display data cannot be stored at present and is mainly used for on-site equipment status adjustment and monitoring. The amount of B display data is small and can be stored in real time. At present, the damage classification of most ultrasonic flaw detection vehicles is to manually judge the damage by manually replaying the B display data screen by screen. This mode of manual damage judgment has the problems of large workload, low judgment efficiency, and easy to cause missed judgment and misjudgment. In view of the above situation, it is necessary to study an efficient and intelligent method for rail damage. Summary of the invention

[0004] In order to overcome the deficiencies of the above-mentioned prior art, the purpose of the present invention is to provide a rail damage identification method for ultrasonic B-display data, so as to solve the problems of low efficiency and low intelligence in the current manual damage judgment. The method includes two stages: training and application. In the training stage, the image classification model is trained based on typical damage B-display image samples to construct a damage identification model. In the application stage, the complete damage data is first extracted from the original B-display data stream. The complete damage data refers to a set of data representing a single damage. The damaged B-display image is redrawn based on the complete damage data. Each damaged B-display image contains only one damage. The redrawn B-display image is classified for damage using the trained damage identification model.

[0005] To achieve the above object, the technical solution adopted by the present invention is: a rail damage identification method for ultrasonic B-display data, comprising the following steps:

[0006] Step 1: Use an ultrasonic flaw detection vehicle to collect damage B-display data in the national calibration section;

[0007] Step 2, using playback software to draw the damaged B-display image, and cropping the damaged B-display image with the damaged B-display image as the center to obtain a cropped damaged B-display image;

[0008] Step 3, setting labels for the cropped damaged B-display images and establishing a sample library of typical damaged B-display images;

[0009] Step 4: training the image classification model based on typical damage B-display image samples to build a damage recognition model;

[0010] Step 5, performing ultrasonic information combination division on the damage B-display data of step 1 to obtain ultrasonic information combination data;

[0011] Step 6: Using a clustering algorithm to combine adjacent ultrasound information combinations in physical locations to complete the division of the damage ultrasound information group, thereby achieving complete damage data extraction;

[0012] Step 7, redrawing the injury B-display image based on the complete injury data, wherein the redrawn injury B-display image only includes one injury ultrasound information group;

[0013] Step 8: Use the redrawn B-display image as input to the damage recognition model for intelligent damage recognition.

[0014] In the step 4, the image classification model selects the EfficientNet network with the highest accuracy in the image classification task to train the damage recognition model; the sine-cosine algorithm SCA is used to optimize the scaling coefficients α, β, and γ of the EfficientNet baseline network on the damaged B-display image sample set; and the model parameters are optimized using the transfer learning strategy to improve the accuracy of the damage recognition model.

[0015] In the step 5, the original B-display data stream is divided into four groups of ultrasound information combinations according to the channel and position correlation and the channel combinations corresponding to different lesions.

[0016] In step 6, the density-based clustering algorithm DBSCAN is used to divide the ultrasound information combination data into damaged ultrasound information groups.

[0017] The beneficial effects of the present invention are:

[0018] The present invention adopts a clustering algorithm to divide the B-display data into ultrasonic information groups, thereby realizing automatic and effective extraction of complete damage data; the present invention constructs a damage recognition model based on a convolutional neural network model, thereby realizing intelligent and accurate recognition of damage data.

[0019] The present invention realizes intelligent damage identification on B-display data collected by a flaw detection vehicle, improves the efficiency of damage determination, and reduces the false alarm rate and missed alarm rate. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 A flow chart of a rail damage identification method based on ultrasonic B-display data of the present invention.

[0021] Figure 2 Flow chart of the scaling coefficient of the sine-cosine algorithm (SCA) optimization model of the present invention.

[0022] Figure 3 Optimization process diagram of the sine-cosine algorithm of the present invention for the scaling factor of the EfficientNet baseline network.

[0023] Figure 4 Diagram of the SCA-EfficientNet transfer learning process of the present invention.

[0024] Figure 5 Diagram of the SCA-EfficientNet model training process based on transfer learning in this invention.

[0025] Figure 6 Schematic diagram of the detection range of the ultrasonic sensor at different angles of the present invention.

[0026] Figure 7 Result diagram of ultrasound information group division according to the present invention.

[0027] Figure 8 Flowchart of the KMNN-DBSCAN algorithm of the present invention.

[0028] Fig. 9 Schematic diagram of the damaged B-display image redrawing process of the present invention. DETAILED DESCRIPTION

[0029] In order to further understand the content of the present invention, the present invention is described in detail in conjunction with the accompanying drawings and embodiments. It should be understood that the embodiments are only for explaining the present invention and are not intended to limit it.

[0030] like Figure 1 As shown, the purpose of this example is to provide a rail damage identification method based on ultrasonic B-display data, and its main steps are as follows:

[0031] Step 1: Use an ultrasonic flaw detection vehicle to collect damage B-display data in the national calibration section;

[0032] Step 2, using playback software to draw the damaged B-display image, and cropping the damaged B-display image with the damaged B-display image as the center to obtain a cropped damaged B-display image;

[0033] Step 3, setting labels for the cropped damaged B-display images and establishing a sample library of typical damaged B-display images;

[0034] Step 4: Select the EfficientNet network with the highest accuracy in image classification tasks to train the damage recognition model; use the sine-cosine algorithm (SCA) to optimize the scaling coefficients α, β, and γ of the EfficientNet baseline network. Perform transfer learning on the model on the ImageNet dataset, optimize the model weight parameters, and build a damage recognition model based on transfer learning;

[0035] Step 5, combining and dividing the ultrasound information of the B-display data of step 1 according to the correlation between the ultrasound sensor channel and the position and combining different injuries;

[0036] Step 6, dividing the ultrasound information combination data into damage ultrasound information groups based on the partitioned KMNN-DBCAN algorithm to achieve complete damage data extraction;

[0037] Step 7, redrawing the injury B-display image based on the complete injury data, wherein the redrawn injury B-display image only includes one injury ultrasound information group;

[0038] Step 8: Use the redrawn B-display image as input to the damage recognition model for intelligent damage recognition.

[0039] The damage recognition model is essentially a convolutional neural network (which can be understood as a black box). It inputs a redrawn damage B-display image and outputs a damage recognition result. The damage recognition standard is the damage recognition accuracy (i.e. Table 4), that is, the proportion of the number of damages correctly identified to the total number of damages.

[0040] In step 1, specifically: this embodiment uses a double-track ultrasonic rail flaw detection vehicle to continuously collect ten times of flaw B-display data as data samples on an artificial flaw test line calibrated in a certain country.

[0041] In step 2, specifically: this embodiment uses playback software to open the damage B-display data to draw the damage B-display image, selects the damage with the most obvious features through manual judgment, and uses the screenshot function of the playback software to crop and save it with the damage B-display image as the center.

[0042] In step 3, specifically: in this example, labels are set for the cropped damaged B-display images by manual labeling to establish a sample set of typical damaged B-display images. 80% of the sample set is selected as a training set, and the remaining 20% ​​is selected as a test set.

[0043] In step 4, if Figure 2The figure shows the scaling coefficient process of the sine-cosine algorithm SCA optimization model in this example, which is specifically: configure the population size, variable dimension, and maximum number of iterations; initialize the position of the first generation of solutions in the population, that is, the initial values ​​of α, β, and γ; calculate the fitness of all solutions, take the largest fitness value as the global optimal fitness value, and take the solution with the largest fitness as the initial value of the global optimal solution; use the position update formula of the sine-cosine algorithm SCA to update the position of the solution; calculate the fitness value of each solution after the update, update the global optimal fitness value and the position of the global optimal solution; determine whether to end, if the end condition is met, end the iteration, otherwise continue to update the position of the global optimal solution;

[0044] The test accuracy of the baseline network model, i.e. the fitness function of the sine-cosine algorithm SCA optimization model scaling coefficient, is expressed as:

[0045] f i =test_accuracy(N i (d Si ,w Si ,r Si ))

[0046]

[0047] Where f i is the fitness value under the i-th group of scaling factors, that is, the test accuracy of the model under the current conditions. Obviously, f i The larger the value, the higher the quality of the scaling factor; i is the network model under the i-th group of scaling factors. Si 、w Si 、r Si is the depth, width and image resolution of the network under the i-th set of scaling factors. k is the basic network of the kth stage of the model, L k represents the number of repetitions of the basic network in the kth stage. X is the input feature matrix of the kth stage. H k , W k , C k The height, width, and number of channels of the input feature matrix for the kth stage.

[0048] The position update formula of the solution of the sine-cosine algorithm SCA is:

[0049]

[0050] Where t is the current iteration number; X *is the current optimal individual position; r2, r3 and r4 are random parameters; r2 has a value range of [0,2π], which defines the distance of the current solution toward or away from the optimal solution; r3 has a value range of [0,2], and r3 defines a random weight for the optimal solution. When r3 is greater than 1, it emphasizes the role of the optimal solution in calculating the motion distance of the candidate solution; when r3 is less than 1, it weakens the role of the optimal solution in calculating the motion distance of the candidate solution. r4 has a value range of [0,1]. When r4 is less than 0.5, the sine model is selected to update the position of the current solution; when r4 is greater than 0.5, the cosine model is selected to update the position of the current solution, realizing equal probability switching of the sine and cosine models. r1 decreases linearly from 2 to 0 as the number of iterations increases. A larger r1 enables the algorithm to have a higher global exploration capability, and a smaller r1 enables the algorithm to have a higher local development capability.

[0051] like Figure 3 The figure shows the optimization process of the scaling factor of the EfficientNet baseline network by the sine-cosine algorithm in this example. After 55 iterations, the model test accuracy reaches the optimal value of 81.7%. When the model test accuracy is optimal, the corresponding optimal scaling factor values ​​are 1.25, 1.15, and 1.1. The model optimized by the sine-cosine algorithm is called the SCA-EfficientNet model.

[0052] Figure 4 , which is the SCA-EfficientNet transfer learning process of this example, firstly, it is trained on the large-scale dataset ImageNet in the source domain to obtain the SCA-EfficientNet pre-trained model, and then the SCA-EfficientNet pre-trained model is migrated to the target domain; finally, the weight parameters learned by the pre-trained model in the source domain are fixed on the target dataset, i.e., the small sample damaged B-display image training sample set, and the fully connected layer is modified to fine-tune the pre-trained model, so as to realize the transfer learning of SCA-EfficientNet based on shared parameters and construct the SCA-EfficientNet damage recognition model based on transfer learning.

[0053] In order to further improve the performance of the model, transfer learning is performed on SCA-EfficientNet based on the ImageNet dataset, and a SCA-EfficientNet damage recognition model based on transfer learning is constructed. Figure 4 The figure shows the SCA-EfficientNet transfer learning process in this example. Figure 5 The figure shows the training process of the SCA-EfficientNet model based on transfer learning in this example. Figure 5It can be seen that the initial accuracy of SCA-EfficientNet based on transfer learning is 45.8% during training, and after 25 iterations, it reaches the highest training accuracy of 99.9%. The model trained by SCA-EfficientNet based on transfer learning is tested using the damage B-image test sample set. Finally, the test accuracy of the SCA-EfficientNet damage recognition model based on transfer learning in this example is 98.7%.

[0054] Step 5, specifically: Figure 6 The following are the detection ranges of ultrasonic sensors at different angles. Figure 6 The detection range of ultrasonic sensors at different angles is different. There are some relevant channels (such as 0°, 37°) and some irrelevant channels (such as 37°, oblique 70°). In order to shield the influence of irrelevant channels on the subsequent ultrasonic information group division, it is necessary to divide the ultrasonic information combination of the B display data. Table 1 shows the ultrasonic sensor combination corresponding to common rail damage. Figure 6 The contents of Table 1 are both the reason and the basis for the present invention to carry out the combination and division of ultrasonic information.

[0055] Table 1

[0056]

[0057]

[0058] According to the correlation between channels and positions, and combined with the corresponding ultrasonic sensor channel combinations of various types of damage, the original B-display data is divided into four ultrasonic information combinations, such as Figure 7 The figure shows the result of ultrasound information group division in this example. It is just to show how ultrasound information combination is divided in this example, that is, this example divides B display data into Figure 7 Four combinations of ultrasound information of injuries.

[0059] In step 6, specifically:

[0060] This example uses the K-median nearest neighbor (KMNN) algorithm to improve the DBSCAN algorithm and proposes the KMNN-DBSCAN algorithm. Figure 8The flowchart of the KMNN-DBSCAN algorithm in this example is shown. The specific steps of the KMNN-DBSCAN algorithm are: first, the K-median nearest neighbor algorithm and the median method are used to generate the Eps parameter list; then, the median method and the given Eps parameter value list are used to generate the Minpts parameter value list; finally, the optimal Eps and Minpts values ​​are selected based on the Eps and Minpts parameter value lists; the optimization process of the optimal Eps and Minpts parameter values ​​is specifically as follows: sequentially select the Eps and Minpts parameter value pairs corresponding to different K values ​​as the clustering parameters of the DBSCAN algorithm, and obtain the number of clustering results under the Eps and Minpts parameter values ​​corresponding to different K values; when the number of clustering results does not change for three consecutive times, the number of clustering results reaches stability, and the current number of clustering results is recorded as the optimal number of clustering results. After the number of clustering results stabilizes, continue the above operations until the number of clustering results changes; the interval from the time when the number of clustering results reaches the optimal number to the time when the number of clustering results changes is called the stable zone, and the expectations of all Eps and Minpts parameter value pairs corresponding to the stable zone are obtained. Will As the optimal Eps and Minpts value.

[0061] The KMNN-DBSCAN algorithm is used to divide the ultrasonic information combination data into damaged ultrasonic information groups. In this example, the F-Measure and SC (Silhouette Coefficient) indicators are selected to evaluate the ultrasonic information group division results of the KMNN-DBSCAN algorithm. Among them, F-Measure reflects the ability of the algorithm to identify valid data points, and the larger the F value, the stronger the ability of the algorithm to identify valid data points. The SC (Silhouette Coefficient) silhouette coefficient comprehensively reflects the compactness within the class and the separation between classes. The value range of SC is [-1, 1]. Within the value range, the larger the SC value, the better the clustering result.

[0062] The calculation formula of F-Mesure is:

[0063]

[0064] Where Recall is the recall rate and Precision is the accuracy rate.

[0065] The calculation formula of Recall is:

[0066]

[0067] The calculation formula of Precision is:

[0068]

[0069] Where TP is the number of data element points predicted as positive and actually positive. FP is the number of data element points predicted as positive and actually negative. FN is the number of data element points predicted as negative and actually positive.

[0070] The calculation formula of SC(i) of SC of an element point i is:

[0071]

[0072] Where a(i) is the average value of the distance between an element point i and all other element points in its cluster, and b(i) is the minimum value of the distance between an element point i and all other element points in different clusters.

[0073] The average SC of all data element points is the SC of the clustering result.

[0074] Table 2 shows the comparison of ultrasound information group division results between the KMNN-DBSCAN algorithm in this example and other parameter adaptive DBSCAN algorithms.

[0075] Table 2

[0076]

[0077] It can be seen from Table 2 that the KMNN-DBSCAN algorithm not only has a strong ability to identify effective element points, but also performs better in the accuracy and rationality of ultrasonic information group division.

[0078] In order to solve the problem that the ultrasonic information group division result of the KMNN-DBSCAN algorithm is inconsistent with the actual number of classes, this example introduces the idea of ​​partitioning to improve the KMNN-DBSCAN algorithm and proposes a partitioned KMNN-DBSCAN algorithm. The specific steps of the partitioned KMNN-DBSCAN algorithm in this example are: first, the element points of the input data set are projected in the X-axis and Y-axis directions respectively to obtain the distribution of the input data set in the X-axis and Y-axis. The partition point is selected at the sparsest density between different density centers to obtain sub-datasets with different densities; then, KMNN-DBSCAN is used to perform parameter adaptation on the sub-datasets respectively, and the sub-datasets are clustered using local optimal parameters to obtain the clustering results of the sub-datasets; finally, the clustering results of the sub-datasets are merged to finally obtain the clustering results of the original data set. Table 3 shows the comparison of the ultrasonic information group division results of the partitioned KMNN-DBSCAN algorithm in this example and the KMNN-DBSCAN algorithm.

[0079] Table 3

[0080]

[0081] According to the correct number of clustering, F value and SC index in Table 3, the partitioned KMNN-DBSCAN algorithm can achieve the division of all damaged ultrasound information groups, further improving the correctness and rationality of the KMNN-DBSCAN algorithm's ultrasound information group division.

[0082] Step (7) is specifically:

[0083] like Fig. 9 D 037 Dataset damage B display image redrawing process, Fig. 9 (a) is D 037 The dataset corresponds to the damaged B-display image, and the partitioned KMNN-DBSCAN algorithm is used to identify the D 037 The complete damage data of the dataset is extracted, and the extraction results are as follows Fig. 9 Finally, the damaged B-display image is redrawn based on the complete data. The redrawn result is shown in Fig. 9 (c) shows that Fig. 9 (c) It can be seen that the redrawn B-image is consistent with the B-image of the playback software; the redrawn injury B-image only contains one injury ultrasound information group.

[0084] Furthermore, in order to verify the effectiveness of the rail damage identification method for ultrasonic B-display data in this example, an ultrasonic rail flaw detector was used to collect damage B-display data at another artificial damage test line in a national calibration section, including 1500 types of damage. The 1500 damage data collected were identified using the method in this example, and the accuracy of various types of damage identification is shown in Table 4.

[0085] Table 4

[0086]

[0087] It can be seen from Table 4 that the rail damage recognition method for ultrasonic B-display data in this example can achieve a recognition rate of more than 95% for different types of rail damage, and the average recognition rate of all damage is 97.0%, realizing intelligent damage recognition for ultrasonic B-display data.

Claims

1. A rail damage identification method for ultrasonic B-display data, characterized in that: The following steps are involved: Step 1: Use an ultrasonic flaw detection vehicle to collect damage B-display data in the national calibration section; Step 2, using playback software to draw the damaged B-display image, and cropping the damaged B-display image with the damaged B-display image as the center to obtain a cropped damaged B-display image; Step 3, setting labels for the cropped damaged B-display images and establishing a sample library of typical damaged B-display images; Step 4: training the image classification model based on typical damage B-display image samples to build a damage recognition model; In the step 4, the image classification model selects the EfficientNet network with the highest accuracy in the image classification task to train the damage recognition model; the sine-cosine algorithm SCA is used to optimize the scaling coefficients α, β, and γ of the EfficientNet baseline network on the damaged B-display image sample set; the model parameters are optimized using the transfer learning strategy to improve the accuracy of the damage recognition model; The scaling coefficient process of the sine-cosine algorithm SCA optimization model is as follows: configure the population size, variable dimension, and maximum number of iterations; initialize the position of the first generation of solutions in the population, that is, the initial values ​​of α, β, and γ; calculate the fitness of all solutions, take the maximum fitness value as the global optimal fitness value, and take the solution with the maximum fitness as the initial value of the global optimal solution; use the position update formula of the sine-cosine algorithm SCA to update the position of the solution; Calculate the fitness value of each solution after the update, update the global optimal fitness value and the position of the global optimal solution; determine whether to end, if the end condition is met, end the iteration, otherwise continue to update the position of the global optimal solution; The test accuracy of the baseline network model, i.e. the fitness function of the sine-cosine algorithm SCA optimization model scaling coefficient, is expressed as: f i =test_accuracy(N i (d Si ,w Si ,r Si )) Where f i is the fitness value under the i-th group of scaling factors, that is, the test accuracy of the model under the current conditions. Obviously, f i The larger the value, the higher the quality of the scaling factor; i is the network model under the i-th group of scaling factors, d Si 、w Si 、r Si is the depth, width and image resolution of the network under the i-th set of scaling factors, F k is the basic network of the kth stage of the model, L k represents the number of repetitions of the basic network in the kth stage, X is the input feature matrix of the kth stage, and H k , W k , C k Enter the height, width and number of channels of the feature matrix for the kth stage; Step 5, performing ultrasonic information combination division on the damage B-display data of step 1 to obtain ultrasonic information combination data; Step 6: Using a clustering algorithm to combine adjacent ultrasound information combinations in physical locations to complete the division of the injury ultrasound information group, thereby achieving complete injury data extraction; Step 7, redrawing the injury B-display image based on the complete injury data, wherein the redrawn injury B-display image only includes one injury ultrasound information group; Step 8: Use the redrawn B-display image as input to the damage recognition model for intelligent damage recognition.

2. A rail damage identification method for ultrasonic B-display data according to claim 1, characterized in that: In the step 5, the original B-display data stream is divided into four groups of ultrasound information combinations according to the channel and position correlation and the channel combinations corresponding to different lesions.

3. A rail damage identification method for ultrasonic B-display data according to claim 1, characterized in that: In step 6, the density-based clustering algorithm DBSCAN is used to divide the ultrasound information combination data into damaged ultrasound information groups.

4. The rail damage identification method for ultrasonic B-display data according to claim 1 is characterized in that: The clustering algorithm comprises the following steps: First, the K-median nearest neighbor algorithm and the median method are used to generate the Eps parameter list; Secondly, the Minpts parameter value list is generated using the median method and the given Eps parameter value list; Finally, the optimal Eps and Minpts values ​​are selected based on the Eps and Minpts parameter value lists; the optimization process of the optimal parameter values ​​of Eps and Minpts is as follows: the Eps and Minpts parameter value pairs corresponding to different K values ​​are selected in turn as the clustering parameters of the DBSCAN algorithm, and the number of clustering results under the Eps and Minpts parameter values ​​corresponding to different K values ​​is obtained; when the number of clustering results does not change for three consecutive times, the number of clustering results reaches stability, and the current number of clustering results is recorded as the optimal number of clustering results. After the number of clustering results stabilizes, the above operation is continued until the number of clustering results changes; from the optimal number of clustering results to the number of clustering results changing ized district Intermediate Stablize district, right Stable Zone For all pairs of Eps and Minpts parameter values, the expected EPS and Minpts are calculated respectively, and EPS and Minpts are taken as the optimal Eps and Minpts values.

Images