Method for Identifying Missing Damage of Track Shear Strands and Dampers Based on Self-Supervised Contrastive Learning

Abstract

The present invention proposes a method for identifying the missing damage of track shear keys and dampers based on self-supervised contrastive learning. This method uses a Chebyshev filter to filter the vibration response data collected by vibration, and uses the grid method and the double-threshold truncation method to extract the vibration response signal during train passing. A time series representation learning model based on self-supervised contrastive learning is built to accurately identify the damage of the track. Self-supervised contrastive learning can extract valuable information from a large amount of unlabeled data, and can achieve effective training and excellent performance with only a small amount of labeled data. At the same time, the mechanism of contrastive learning is used to better learn the data characteristics, so as to provide better recognition results. The present invention can achieve effective training and excellent track damage recognition performance based on a small amount of labeled data, thereby reducing the dependence of the model on manually labeled data, and is applicable to the efficient recognition of track damage in cases where it is inconvenient to manually label a large amount of data.

Description

Technical Field

[0001] The present invention belongs to the technical fields of machine learning and track damage identification, and particularly relates to a method for identifying the missing damage of track shear keys and dampers based on self-supervised contrast learning. Background Art

[0002] Under the action of long-term wheel friction, impact and various environmental factors, various types of diseases will inevitably occur in the track structure, resulting in the degradation of service performance. Under the action of the train cyclic load, the damper will bear a large repeated vertical load and gradually fatigue and break, resulting in the weakening of the ballast bed vibration damping performance, the aggravation of the dynamic response of the track system and the serious endangerment of the vehicle operation safety. Since the damper is encapsulated in the metal sleeve under the track slab, it cannot be directly observed visually. The shear key will also gradually deform under the influence of various factors, the deformation will occur between the ballast slabs, and the vertical deformation of the track cannot change continuously, increasing the displacement mutation between the ballast beds, which will seriously affect the driving safety. Therefore, it is particularly important to identify the damage of the damper and the shear key. However, the existing track detection methods still have great limitations and cannot meet the actual detection needs. Therefore, in order to ensure the service safety and operation efficiency of rail transit, it is essential to develop high-efficiency and high-precision structural health detection and monitoring technologies. Self-supervised contrast learning reduces the dependence on labeled data, reduces the cost of data collection and labeling, and at the same time has good robustness, and still has high performance in the face of environmental changes and data noise. Summary of the Invention

[0003] The purpose of the present invention is to solve the problems in the prior art to meet the actual detection needs, and a method for identifying the missing damage of track shear keys and dampers based on self-supervised contrast learning is proposed. This method is suitable for realizing effective training and excellent track damage identification performance in the case where it is inconvenient to manually label a large amount, and reduces the dependence of the model on manually labeled data.

[0004] The present invention is realized by the following technical solutions. The present invention proposes a method for identifying the missing damage of track shear keys and dampers based on self-supervised contrast learning, and the method includes the following steps:

[0005] Step 1: Collect the vibration monitoring response data corresponding to the missing damage of the shear key and the damper under the action of the train load on the urban track;

[0006] Step 2: Extract the vibration response data during train passing. Use the Chebyshev filter to filter the vibration response data collected in the experiment, select the optimal parameters of the double-threshold truncation method of short-time energy and short-time low-frequency energy by the grid method, and use the double-threshold truncation method to extract the vibration response data during train passing from a large amount of vibration monitoring response data; preprocess the vibration response data during train passing, and use the moving average method to process the data X of every five pointst , X t+1 , X t+2 , X t+3 , X t+4 Average to X' t =(X t + X t+1 + X t+2 + X t+3 + X t+4 ) / 5, and then randomly intercept the vibration response signal with a length of 5120 to form a model training dataset matrix

[0007] Step 3: Divide the matrix X obtained in Step 2 into two datasets according to the ratio of 90% and 10%. Substitute 90% of the dataset into the time series representation learning model based on time and context contrast for pre-training, and save the pre-training results, weights, and model parameters;

[0008] Step 4: Divide the 10% dataset obtained in Step 3 into a training set, a test set, and a validation set, and substitute them into the pre-trained model for retraining. Analyze the test results to accurately identify the missing damages of the track shear connectors and dampers.

[0009] Furthermore, the specific content of Step 2 is as follows:

[0010] Step 2.1: Splice the vibration responses corresponding to each train passing, and obtain a complete vibration response of the train passing from the start time to the end time;

[0011] Step 2.2: Use the Chebyshev filter to filter the original time history curve to filter out the low-frequency drift and obtain more accurate data;

[0012] Step 2.3: Use the VAD model and the grid method to determine the thresholds of the short-time energy and the short-time low-frequency energy for extracting the vibration response data;

[0013] Step 2.4: Use the double thresholds of the short-time energy and the short-time low-frequency energy to extract and save the train passing signal;

[0014] Step 2.5: Perform the moving average method on the extracted train passing vibration response, randomly intercept a signal with a length of 5120 from the vibration response signal, and splice it into a dataset

[0015] Furthermore, the specific content of Step 2.3 is as follows:

[0016] Step 2.3.1: Select 1000 pieces of train signal data from the vibration response signal and label them. For one time history curve, the part with signal β is marked as 1, and the part without signal α is marked as 0. The train signal vector v = [α, β, α], and the label vector v = [0, 0, ……, 0, 1, 1, 1, ……, 1, 1, 1, 0, ……, 0, 0];

[0017] Step 2.3.2: Determine the grid parameter Z = γ * γ T , where γ = [0.04, 0.08, 0.12, 0.16, 0.20, 0.24, 0.28, 0.32, 0.36, 0.40];

[0018] Step 2.3.3: Determine the short-time energy E f and the short-time low-frequency energy E n . For the Hanning window W(n), the frame length is L = 50 ms, the translation distance each time is 25 ms, and the short-time energy frequency value range v = [5, 100]. The short-time energy is the square of the amplitude of each frame in the time domain, and the short-time low-frequency energy is to perform Fourier transform on each frame and calculate the total energy within the frequency value range. The short-time energy and short-time low-frequency energy of the nth frame are calculated according to the following formulas;

[0019]

[0020]

[0021] Step 2.3.4: Substitute 1000 pieces of vibration response data and 1000 labels into the VAD model for training, and calculate the maximum intersection over union IoU of the short-time energy and the short-time low-frequency energy;

[0022] Step 2.3.5: Use the threshold parameter corresponding to the maximum intersection over union to test 500 pieces of data and labels to verify the threshold parameter.

[0023] Furthermore, in Step 3, the matrix X obtained after preprocessing is divided into two data sets according to the ratio of 90% and 10%. The 90% data set is substituted into the time series representation learning model TS-TCC based on time and context contrast for pre-training. Through continuous optimization of the loss function, the extracted feature vectors that are similar are made closer, and those that are dissimilar are made farther apart. Save the pre-training results, weights, and model parameters.

[0024] Furthermore, the specific steps of Step 3 are as follows:

[0025] Step 3.1: Divide the preprocessed matrix X into two data sets according to the ratio of 90% and 10%;

[0026] Step 3.2: Perform two different data augmentations, jittering and permutation, on one piece of vibration response data;

[0027] Step 3.3: Substitute the two data after data augmentation into the encoder to extract latent features and The encoder is a convolutional neural network with four convolutional blocks. The initial layer performs one-dimensional convolution on the specified number of input channels to generate 32 output channels. Then, preliminary data normalization is performed through batch normalization operation, and non-linear features are added through the ReLU activation function. Finally, max pooling and the Dropout function are used for feature dimensionality reduction and overfitting prevention. The remaining three convolutional modules deepen the network structure in the same way successively, and increase the number of channels, adjust the convolutional kernel size and stride at each step to continuously improve the network capacity and feature extraction ability, and finally output a feature vector with a length of 128;

[0028] Step 3.4: Substitute the extracted latent features and into the Transformer to merge them into two feature vectors and

[0029] Step 3.5: Use and to predict the time history curves after k seconds respectively and W k is a linear function,

[0030] used to map c t to the same dimension as z;

[0031]

[0032] Step 3.6: Maximize the similarity of similar samples and minimize the similarity between dissimilar samples, that is, make the distance between the feature vectors c t of similar samples closer and the distance between the feature vectors c t of dissimilar samples farther;

[0033]

[0034] Step 3.7: Save the pre-training results, weights and model parameters.

[0035] Furthermore, in Step 3.2, jittering is achieved by adding random noise to the original data; permutation is achieved by splitting the time series of the original data into several parts and then randomly reordering these parts to create new data while keeping the data order within each split part unchanged.

[0036] Furthermore, the specific content of Step Four is as follows:

[0037] Step 4.1: Substitute the parameters and weights saved in Step 3 into the pre-trained model;

[0038] Step 4.2: Label 10% of the dataset matrix X, with the health condition being 0, shear link missing being 1, and damper missing being 2, and divide it into a training set, a validation set, and a test set according to the ratios of 60%, 20%, and 20%;

[0039] Step 4.3: Substitute the training set and the validation set into the pre-trained model, fine-tune the pre-trained parameters, and use the best training results to test the test set to diagnose the damage category corresponding to each vibration response data.

[0040] The present invention also proposes an identification system for the damage of track shear links and damper missing based on self-supervised contrastive learning, and the system includes:

[0041] A data acquisition module, on a certain track health monitoring system, collects the vibration monitoring response data of the track missing shear links and dampers under the action of train loads;

[0042] A data preprocessing module, filters the vibration response data collected from experiments using a Chebyshev filter, selects the best parameters of the short-time energy and short-time low-frequency energy double-threshold truncation method using the grid method, and uses the double-threshold truncation method to extract the vibration response data when the train passes from a large amount of vibration monitoring response data; preprocesses the vibration response data when the train passes, and uses the moving average method to average the data X of every five points t ,X t+1 ,X t+2 ,X t+3 ,X t+4 to X' t =(X t +X t+1 +X t+2 +X t+3 +X t+4 ) / 5, and then randomly intercepts a vibration response signal with a length of 5120 to form a model training dataset matrix

[0043] A model training module, divides the preprocessed dataset matrix X into two datasets according to the ratios of 90% and 10%, substitutes 90% of the dataset into the TS-TCC model for pre-training, and divides the 10% dataset into a training set, a test set, and a validation set to fine-tune the pre-training results;

[0044] A damage diagnosis module, uses the parameters obtained from the pre-trained model training to test the test set and diagnose the damage type corresponding to each monitoring data.

[0045] The present invention also provides an electronic device, including a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, the steps of the method for identifying the missing damage of the track shear key and damper based on self-supervised contrastive learning are implemented.

[0046] The present invention also provides a computer-readable storage medium for storing computer instructions. When the computer instructions are executed by a processor, the steps of the method for identifying the missing damage of the track shear key and damper based on self-supervised contrastive learning are implemented.

[0047] Advantages of the present invention:

[0048] 1. The present invention provides a vibration-based track damage monitoring method, which has the advantages of global and real-time compared with traditional track condition detection methods, and can detect hidden or unobvious damage in time and effectively.

[0049] 2. The present invention provides a self-supervised track damage identification method, which only requires a small amount of data to be labeled compared with supervised learning methods, alleviating the problem of high data annotation cost for damage identification.

[0050] 3. The time series representation learning model based on time contrast and context contrast used in the present invention is applicable to time history curves with time dependence, and has good generalization ability and the ability to identify different damage categories. Description of the Drawings

[0051] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are only the embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained according to the provided drawings without creative efforts.

[0052] Figure 1 It is a flow chart of the method for identifying the missing damage of the track shear key and damper based on self-supervised contrastive learning according to the present invention;

[0053] Figure 2 It is a representative time history curve graph collected at a certain measuring point in the embodiment; among them, (a) is the curve graph of the healthy state, (b) is the curve graph of the missing damper, (c) is the curve graph of the missing shear key. In the three graphs of (a), (b), and (c), the abscissa is time and the ordinate is the disturbance response;

[0054] Figure 3 It is a schematic diagram of the IoU for VAD model training in the embodiment;

[0055] Figure 4Flow chart of the TS-TCC model used in the embodiments;

[0056] Figure 5 Loss function graph and confusion matrix graph of the recognition results of randomly selected data damage in the embodiments. Detailed implementation manners

[0057] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0058] In the problem of damage identification, considering that track damage does not occur frequently and there are various types, it is not easy to label. The time series representation learning model based on time and context contrast can learn more robust and generalizable feature representations by contrasting data to enhance positive and negative samples, and can improve the accuracy of track damage detection in the case of data sparsity. Therefore, the present invention studies using this model to extract features and train the preprocessed time history curve data to achieve track damage identification. First, filter the vibration monitoring response data during train passing to make the train passing signal more obvious, and then use the VAD model and the grid method to find the most appropriate short-time energy and short-time frequency energy thresholds. Extract and preprocess the data, and substitute it into the model for pre-training and re-training, so as to achieve the goal of self-supervised track damage identification of the present invention.

[0059] Combined with Figure 1 , the present invention proposes a method for identifying the damage of track shear keys and damper missing based on self-supervised contrast learning. The method includes the following steps:

[0060] Step 1: Collect the vibration monitoring response data corresponding to the damage of track shear keys and damper missing under the action of train loads in urban rails;

[0061] Step 2: Extract the train passing vibration response data, filter the vibration response data collected through experiments using a Chebyshev filter, select the best parameters of the double-threshold truncation method of short-time energy and short-time low-frequency energy using the grid method, and use the double-threshold truncation method to extract the vibration response data when the train passes from a large amount of vibration monitoring response data. Preprocess the train passing vibration response data, and use the moving average method to average the data X t , X t+1 , X t+2 , X t+3 , X t+4 of every five points into X' t =(X t +X t+1 +Xt+2 +X t+3 +X t+4 ) / 5, and then randomly intercept a vibration response signal with a length of 5120 to form a model training dataset matrix

[0062] Step 3: Divide the matrix X obtained in Step 2 into two datasets according to the ratio of 90% and 10%. Substitute 90% of the dataset into the time series representation learning model based on time and context contrast for pre-training, and save the pre-training results, weights, and model parameters;

[0063] Step 4: Divide the 10% dataset obtained in Step 3 into a training set, a test set, and a validation set, and substitute them into the pre-trained model for retraining. Analyze the test results to accurately identify the missing damages of the track shear key and the damper.

[0064] The specific content of Step 2 is as follows:

[0065] Step 2.1: Concatenate the vibration responses corresponding to each train passing to obtain a complete vibration response of the train passing from the start time to the end time;

[0066] Step 2.2: Use a Chebyshev filter to filter the original time history curve to remove low-frequency drift and obtain more accurate data;

[0067] Step 2.3: Use the VAD model and the grid method to determine the thresholds of the short-time energy and the short-time low-frequency energy for extracting vibration response data;

[0068] Step 2.4: Use appropriate double thresholds of short-time energy and short-time low-frequency energy to extract and save the train passing signal;

[0069] Step 2.5: Perform a moving average method on the extracted train passing vibration response, randomly intercept a signal with a length of 5120 from the vibration response signal, and splice it into a dataset

[0070] The specific content of Step 2.3 is as follows:

[0071] Step 2.3.1: Select 1000 train passing signal data from the vibration response signal and label them. The part of a time history curve with signal β is 1, and the part without signal α is 0. The train passing signal vector v = [α, β, α], and the label vector v = [0, 0,..., 0, 1, 1, 1,..., 1, 1, 1, 0,..., 0, 0];

[0072] Step 2.3.2: Determine the grid parameter Z = γ * γ T, γ = [0.04, 0.08, 0.12, 0.16, 0.20, 0.24, 0.28, 0.32, 0.36, 0.40];

[0073] Step 2.3.3, determine the short-time energy E f and the short-time low-frequency energy E n , with a Hanning window W(n), a frame length of L = 50 ms, a shift distance of 25 ms each time, and a short-time energy frequency range v = [5, 100]. The short-time energy is the square of the amplitude of each frame in the time domain, and the short-time low-frequency energy is obtained by performing a Fourier transform on each frame and calculating the sum of the energies within the frequency range. The short-time energy and short-time low-frequency energy of the nth frame are calculated according to the following formulas;

[0074]

[0075] Step 2.3.4, substitute 1000 vibration response data and 1000 labels into the VAD model for training, and calculate the maximum intersection over union IoU of the short-time energy and short-time low-frequency energy;

[0076] Step 2.3.5, use the threshold parameter corresponding to the maximum intersection over union to test 500 data and labels to verify the threshold parameter.

[0077] The specific steps of step three are as follows:

[0078] Divide the matrix X obtained after preprocessing into two datasets according to the ratio of 90% and 10%. Substitute 90% of the dataset into the time series representation learning model based on time and context contrast (TS-TCC) for pre-training. Through continuous optimization of the loss function, make the extracted feature vectors that are similar closer and those that are dissimilar farther apart. Save the pre-training results, weights, model parameters, etc.; The specific process is as follows:

[0079] Step 3.1, divide the preprocessed matrix X into two datasets according to the ratio of 90% and 10%;

[0080] Step 3.2, perform two different data augmentations, jittering and permutation, on a vibration response data. Jittering is achieved by adding some very small random noise to the original data. Permutation is achieved by splitting the time series of the original data into several parts and then randomly reordering these parts to create new data while keeping the data order within each split part unchanged.

[0081] Step 3.3, substitute the two data after data augmentation into the encoder to extract latent features and The encoder is a convolutional neural network with four convolutional blocks. The initial layer performs one-dimensional convolution on the specified number of input channels to generate 32 output channels. Then, initial data normalization is carried out through batch normalization operation, followed by adding non-linear features through the ReLU activation function. Finally, max pooling and the Dropout function are used for feature dimensionality reduction and preventing overfitting. The remaining three convolutional modules deepen the network structure in the same way successively, and increase the number of channels, adjust the convolutional kernel size and stride at each step to continuously improve the network capacity and feature extraction ability, and finally output a feature vector with a length of 128.

[0082] Step 3.4: Substitute the extracted latent features and into the Transformer to merge them into two feature vectors and

[0083] Step 3.5: Use and to predict the time history curves and W k is a linear function used to map c t to the same dimension as z;

[0084]

[0085] Step 3.6: Maximize the similarity of similar samples and minimize the similarity between dissimilar samples, that is, make the distance between the feature vectors c t of similar samples closer and the distance between the feature vectors c t of dissimilar samples farther;

[0086]

[0087]

[0088] Step 3.7: Save the pre-training results, weights, model parameters, etc.

[0089] Specifically, the fourth step is as follows:

[0090] Step 4.1: Substitute the parameters, weights, etc. saved in step three into the pre-trained model;

[0091] Step 4.2: Label 10% of the dataset matrix X, with the health status being 0, shear key missing being 1, and damper missing being 2, and divide it into a training set, a validation set, and a test set according to the ratios of 60%, 20%, and 20%;

[0092] Step 4.3: Substitute the training set and the validation set into the pre-trained model, fine-tune the pre-trained parameters, and test the test set with the best training results to identify the damage category corresponding to each vibration response data.

[0093] The present invention proposes a method for identifying the missing damage of track shear keys and dampers based on self-supervised contrastive learning. This method uses a Chebyshev filter to filter the vibration response data collected by vibration, and uses the grid method and the double-threshold truncation method to extract the vibration response signal during train passing. A time series representation learning model based on self-supervised contrastive learning is built to accurately identify the damage of track shear keys and dampers. Self-supervised contrastive learning can extract valuable information from a large amount of unlabeled data, and can achieve effective training and excellent performance with only a small amount of labeled data. At the same time, the mechanism of contrastive learning is used to better learn the data characteristics, so as to provide better recognition results. The present invention can achieve effective training and excellent track damage recognition performance based on a small amount of labeled data, thereby reducing the dependence of the model on artificially labeled data, and is applicable to the efficient identification of track damage in the case where it is inconvenient to perform a large amount of artificial labeling, and has good engineering application value.

[0094] The present invention proposes a system for identifying the missing damage of track shear keys and dampers based on self-supervised contrastive learning, and the system includes:

[0095] A data acquisition module, on a certain track health monitoring system, collects the vibration monitoring response data of the track missing shear keys and dampers under the action of train loads;

[0096] A data preprocessing module, uses a Chebyshev filter to filter the vibration response data collected by the experiment, selects the optimal parameters of the double-threshold truncation method of short-time energy and short-time low-frequency energy by the grid method, and uses the double-threshold truncation method to extract the vibration response data during train passing from a large amount of vibration monitoring response data; preprocesses the vibration response data during train passing, and uses the moving average method to average the data X of every five points t ,X t+1 ,X t+2 ,X t+3 ,X t+4 to X' t =(X t +X t+1 +X t+2 +X t+3 +X t+4 ) / 5, and then randomly intercepts a vibration response signal with a length of 5120 to form a model training data set matrix

[0097] The model training module divides the preprocessed dataset matrix X into two datasets according to the ratio of 90% and 10%. 90% of the dataset is substituted into the TS-TCC model for pre-training, and 10% of the dataset is divided into a training set, a test set, and a validation set to fine-tune the pre-training results.

[0098] The damage diagnosis module uses the parameters obtained from training the pre-trained model to test the test set and diagnose the damage type corresponding to each monitoring data.

[0099] The method for identifying the missing damage of track shear keys and dampers based on self-supervised contrast learning proposed by the present invention realizes the identification of subway track damage by using the self-supervised contrast learning method, avoids a large amount of manual data annotation, and provides a new way for exploring track damage identification in complex environments. The present invention can realize the identification of track damage types, save human resources and time costs, and assist the maintenance of the track department.

[0100] Embodiment

[0101] Combined with Figures 2 - 5 , for a certain urban track health monitoring system in China, it is found that some track beds are missing shear keys and dampers, and the vibration monitoring response data of the track missing shear keys and dampers under the action of train loads is collected. This embodiment shows the damage identification results of the measurement data by using the method for identifying the missing damage of track shear keys and dampers based on self-supervised contrast learning described in the present invention.

[0102] Next, the method for identifying the missing damage of track shear keys and dampers based on self-supervised contrast learning in the present invention is used to identify the missing damage of track shear keys and dampers.

[0103] The specific content of step one is as follows: On a certain track health monitoring system in China, it is found that some tracks are missing shear keys and dampers, and the vibration response data of the track missing shear keys and dampers is collected, and the sampling frequency is 1000Hz.

[0104] The specific content of step two is as follows: Based on the data collected in step one, the vibration responses corresponding to each train passing are spliced to obtain a complete vibration response of a train passing from the start time to the end time. The original time history curve is filtered by a Chebyshev filter to filter out the low-frequency drift and retain the signal data within 5Hz - 100Hz. 1000 signals are selected and labeled and substituted into the VAD model and the grid method to determine the thresholds of the short-time energy and the short-time low-frequency energy for extracting the vibration response data. The grid parameter Z = γ * γ T , γ = [0.04, 0.08, 0.12, 0.16, 0.20, 0.24, 0.28, 0.32, 0.36, 0.40], and the highest intersection over union IoU is as Figure 3As shown, another 500 data are selected and tagged to test the dual thresholds corresponding to the highest IoU. The training and test results are shown in Table 1 below. The appropriate dual thresholds are used to extract the passing vehicle signals, and the extracted passing vehicle vibration responses are processed by the moving average method. A signal with a length of 5120 is randomly intercepted from the vibration response signal and spliced into a dataset.

[0105] Table 1 Training and Test Results of Dual Thresholds

[0106] F1 Score (%) Recall Rate (%) Sample Precision (%) Optimal Intersection over Union (%) Training Results 95.96 99.80 95.93 0.8914 Test Results 97.59 96.85 98.34 0.8656

[0107] Specifically, Step 3 is as follows: The matrix X obtained after preprocessing is divided into two datasets according to the ratio of 90% and 10%. The 90% dataset is divided into a training set and a test set. The training set is substituted into the time series representation learning model based on time and context contrast (TS-TCC) for pre-training. The model framework is as Figure 4 shown. Through continuous optimization of the loss function, the extracted feature vectors that are similar are made closer, and those that are dissimilar are made farther apart. The pre-training results, weights, model parameters, etc. are saved.

[0108] Specifically, Step 4 is as follows: The 10% dataset obtained in Step 4 is divided into a training set and a test set and substituted into the pre-trained model for re-training, and the test set in Step 3 is used for testing. The confusion matrix is as Figure 5 shown. By analyzing the test results, the track shear connectors and damper missing damages can be accurately identified.

[0109] In the above embodiments, self-supervised contrast learning demonstrates its advantages in training using limited labeled data. The recognition effect is good, about 90% on average. Self-supervised contrast learning extracts valuable information from a large amount of unlabeled data, effectively makes up for the shortage of labeled data, reduces the dependence on labeled data, and improves the generalization ability of the model.

[0110] The present invention proposes an electronic device, including a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, the steps of the method for identifying track shear connectors and damper missing damages based on self-supervised contrast learning are implemented.

[0111] The present invention proposes a computer-readable storage medium for storing computer instructions, and when the computer instructions are executed by a processor, the steps of the method for identifying track shear connectors and damper missing damages based on self-supervised contrast learning are implemented.

[0112] The memory in the embodiments of the present application may be a volatile memory or a non-volatile memory, or may include both volatile and non-volatile memories. Among them, the non-volatile memory may be a read only memory (ROM), a programmable ROM (PROM), an erasable PROM (EPROM), an electrically erasable PROM (EEPROM), or a flash memory. The volatile memory may be a random access memory (RAM), which is used as an external cache. By way of example but not limitation, many forms of RAM are available, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDR SDRAM), enhanced SDRAM (ESDRAM), synchlink DRAM (SLDRAM), and direct rambus RAM (DRRAM). It should be noted that the memory of the method described in the present invention is intended to include but not limited to these and any other suitable types of memories.

[0113] In the above embodiments, it can be implemented in whole or in part by software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer instructions are loaded and executed on a computer, the processes or functions described in the embodiments of the present application are generated in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wire (such as coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (such as infrared, wireless, microwave, etc.). The computer-readable storage medium can be any available medium that can be accessed by a computer or a data storage device such as a server, data center, etc. that includes one or more integrated available media. The available media can be magnetic media (such as floppy disks, hard disks, magnetic tapes), optical media (such as high-density digital video discs (DVDs)), or semiconductor media (such as solid state discs (SSDs)), etc.

[0114] In the implementation process, each step of the above method can be completed by the integrated logic circuit of the hardware in the processor or the instructions in the form of software. The steps of the method disclosed in combination with the embodiments of the present application can be directly embodied as being executed and completed by the hardware processor, or executed and completed by the combination of the hardware and software modules in the processor. The software module can be located in a mature storage medium in the art such as random access memory, flash memory, read-only memory, programmable read-only memory, or electrically erasable programmable memory, registers, etc. This storage medium is located in the memory, and the processor reads the information in the memory and combines its hardware to complete the steps of the above method. To avoid repetition, it will not be described in detail here.

[0115] It should be noted that the processor in the embodiments of the present application may be an integrated circuit chip with signal processing capabilities. In the implementation process, the steps of the above method embodiments can be completed by the integrated logic circuit in the hardware of the processor or instructions in the form of software. The above-mentioned processor may be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components. It can implement or execute the various methods, steps, and logic block diagrams disclosed in the embodiments of the present application. The general-purpose processor may be a microprocessor or the processor may also be any conventional processor, etc. The steps of the method disclosed in combination with the embodiments of the present application can be directly embodied as being executed and completed by a hardware decoding processor, or executed and completed by a combination of hardware and software modules in the decoding processor. The software module may be located in a mature storage medium in the art such as a random access memory, a flash memory, a read-only memory, a programmable read-only memory, or an electrically erasable programmable memory, a register, etc. The storage medium is located in the memory, and the processor reads the information in the memory and combines its hardware to complete the steps of the above method.

[0116] The above has introduced in detail the method for identifying the missing damage of the track shear hinge and damper based on self-supervised contrast learning proposed by the present invention. Specific examples are used in this article to elaborate on the principle and implementation manner of the present invention. The description of the above embodiments is only used to help understand the method and its core idea of the present invention; at the same time, for those of ordinary skill in the art, according to the idea of the present invention, there will be changes in the specific implementation manner and application scope. In summary, the content of this specification should not be construed as a limitation to the present invention.

Claims

1. A method for identifying missing damages of track shear keys and dampers based on self-supervised contrastive learning, characterized in that, The method includes the following steps: Step 1: Collect vibration monitoring response data corresponding to the missing damage of shear keys and dampers under the action of train loads on urban rails; Step 2: Extract the vibration response data of the passing train. Filter the vibration response data collected in the experiment using a Chebyshev filter. Select the optimal parameters of the double-threshold truncation method for short-time energy and short-time low-frequency energy using the grid method. Use the double-threshold truncation method to extract the vibration response data when the train passes from a large amount of vibration monitoring response data. Preprocess the vibration response data of the passing train. Use the moving average method to average the data of every five points X t ,X t+1 ,X t+2 ,X t+3 ,X t+4 to X' t =(X t +X t+1 +X t+2 +X t+3 +X t+4 ) / 5. Then randomly intercept the vibration response signal with a length of 5120 to form the model training dataset matrix Step 3: Divide the matrix X obtained in Step 2 into two data sets according to the ratio of 90% and 10%. Substitute 90% of the data set into the time series representation learning model based on time and context contrast for pre-training, and save the pre-training results, weights, and model parameters; Step 4: Divide the 10% data set obtained in Step 3 into a training set, a test set, and a validation set, and substitute them into the pre-trained model for retraining. Analyze the test results to accurately identify the missing damage of track shear keys and dampers; The specific content of Step 2 is as follows: Step 2.1: Stitch the vibration responses corresponding to each train operation period to obtain a complete vibration response of passing trains from the start time to the end time; Step 2.2: Use a Chebyshev filter to filter the original time history curve to remove low-frequency drift and obtain more accurate data; Step 2.3: Use the VAD model and the grid method to determine the thresholds of short-time energy and short-time low-frequency energy for extracting vibration response data; Step 2.4: Extract and save the passing train signals using the double thresholds of short-time energy and short-time low-frequency energy; Step 2.5: Apply the moving average method to the extracted vehicle passing vibration response. Randomly intercept a signal with a length of 5120 from the vibration response signal and splice it into a data set The specific content of Step 2.3 is as follows: Step 2.3.1: Select 1000 passing train signal data from the vibration response signal and label them. The part with signal β in one time history curve is 1, and the part without signal α is 0. The passing train signal vector v = [α, β, α], and the label vector v = [0, 0, ……, 0, 1, 1, 1, ……, 1, 1, 1, 0, ……, 0, 0]; Step 2.3.2, determine the grid parameter Z = γ * γ T , γ = [0.04, 0.08, 0.12, 0.16, 0.20, 0.24, 0.28, 0.32, 0.36, 0.40]; Step 2.3.3, determine the short-time energy E f and the short-time low-frequency energy E n , with a Hanning window W(n), a frame length of L = 50 ms, a translation distance of 25 ms each time, and a short-time energy frequency value range v = [5, 100]. The short-time energy is the square of the amplitude of each frame in the time domain. The short-time low-frequency energy is obtained by performing a Fourier transform on each frame and calculating the sum of the energies within the frequency value range. The short-time energy and short-time low-frequency energy of the nth frame are calculated according to the following formulas; Step 2.3.4: Substitute 1000 vibration response data and 1000 labels into the VAD model for training, and calculate the maximum intersection over union IoU of short-time energy and short-time low-frequency energy; Step 2.3.5: Use the threshold parameter corresponding to the maximum intersection over union to test 500 data and labels to verify the threshold parameter.

2. The method according to claim 1, wherein In Step 3, divide the matrix X obtained after preprocessing into two data sets according to the ratio of 90% and 10%. Substitute 90% of the data set into the time series representation learning model TS-TCC based on time and context contrast for pre-training. Through continuous optimization of the loss function, make the extracted feature vectors that are similar closer and those that are dissimilar farther apart. Save the pre-training results, weights, and model parameters.

3. The method according to claim 1, wherein The specific content of Step 3 is as follows: Step 3.1: Divide the preprocessed matrix X into two data sets according to the ratio of 90% and 10%; Step 3.2: Perform two different data augmentations, jittering and permutation, on one vibration response data; Step 3.3: Substitute the two data after data augmentation into the encoder respectively to extract latent features and The encoder is a convolutional neural network with four convolutional blocks. The initial layer performs one-dimensional convolution on the specified number of input channels to generate 32 output channels. Then, preliminary data normalization is performed through batch normalization operation, and non-linear features are added through the ReLU activation function. Finally, max pooling and the Dropout function are used for feature dimensionality reduction and preventing overfitting. The remaining three convolutional modules deepen the network structure in the same way successively, and increase the number of channels, adjust the convolutional kernel size and the sliding step size at each step to continuously improve the network capacity and feature extraction ability, and finally output a feature vector with a length of 128; Step 3.

4. Substitute the extracted potential features and into the Transformer to merge them into two feature vectors and Step 3.5, using and respectively predict the time history curves after k seconds and W k is a linear function used to map c t to the same dimension as ; Step 3.6: Maximize the similarity of similar samples and minimize the similarity between dissimilar samples, that is, make the feature vectors c of similar samples t get closer and the feature vectors c of dissimilar samples t get farther apart; Step 3.7: Save the pre-training results, weights, and model parameters.

4. The method according to claim 3, characterized in that, In Step 3.2, jittering is achieved by adding random noise to the original data; permutation is achieved by splitting the time series of the original data into several parts and then randomly reordering these parts to create new data while keeping the data order within each split part unchanged.

5. The method according to claim 1, wherein The specific content of Step 4 is as follows: Step 4.1: Substitute the parameters and weights saved in Step 3 into the pre-trained model; Step 4.2: Label the 10% dataset matrix X, with the health status being 0, shear link missing being 1, and damper missing being 2, and divide it into a training set, a validation set, and a test set according to the ratios of 60%, 20%, and 20%. Step 4.3: Substitute the training set and the validation set into the pre-trained model, fine-tune the pre-trained parameters, and use the best training results to test the test set to diagnose the damage category corresponding to each vibration response data.

6. The track shear key and damper missing damage identification system based on self-supervised contrastive learning is characterized in that: The system is implemented based on the identification method described in Claim 1, and the system includes: A data acquisition module, on a certain track health monitoring system, collects the vibration monitoring response data of the track missing shear links and dampers under the action of train loads. Data preprocessing module, which filters the vibration response data collected from experiments using a Chebyshev filter, selects the optimal parameters of the short-time energy and short-time low-frequency energy double-threshold truncation method using the grid method, and uses the double-threshold truncation method to extract the vibration response data when the train passes from a large amount of vibration monitoring response data; preprocesses the vibration response data when the train passes, and uses the moving average method to average the data of every five points X t ,X t+1 ,X t+2 ,X t+3 ,X t+4 to X' t =(X t +X t+1 +X t+2 +X t+3 +X t+4 ) / 5, and then randomly intercepts the vibration response signal with a length of 5120 to form the model training dataset matrix A model training module divides the pre-processed dataset matrix X into two datasets according to the ratios of 90% and 10%. The 90% dataset is substituted into the TS-TCC model for pre-training, and the 10% dataset is divided into a training set, a test set, and a validation set to fine-tune the pre-training results. A damage diagnosis module uses the parameters obtained from training the pre-trained model to test the test set and diagnose the damage type corresponding to each monitoring data.

7. An electronic device, comprising a memory and a processor, the memory storing a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method described in any one of Claims 1-5.

8. A computer-readable storage medium for storing computer instructions, characterized in that, When the computer instructions are executed by the processor, it implements the steps of the method described in any one of Claims 1-5.

Images