A progressive self-supervised damage identification method for high-speed rail wheels and rails

By adopting a progressive self-supervision method in the identification of wheel and rail damage of high-speed rail, establishing a teacher-student network and self-attention block, the problem of insufficient identification performance and interpretation in the existing technology is solved, and the problem of less labels and variable working conditions is realized. efficient damage identification.

CN118690146BActive Publication Date: 2024-11-22SUZHOU UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411195521.7
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-08-29
Publication Date
2024-11-22
Estimated Expiration
2044-08-29

AI Technical Summary

Technical Problem

The existing self-supervised learning methods are limited to designing predefined agent tasks in high-speed rail damage recognition, resulting in weak recognition performance and interpretation, and poor recognition effect under less labels and variable operating conditions.

Method used

Adopting an incremental self-supervision method, by establishing a teacher-student network for training, using mask operations and self-attention blocks for characterization learning and reconstruction, generating masks from easy to difficult, optimizing reconstruction losses and predicting losses, and enhancing the network characterization and identification performance.

Benefits of technology

The actual scenario performance of high-speed rail damage recognition is improved, the type recognition ability under less labels and variable operating conditions is improved, and the interpretability and recognition performance of the model is enhanced.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118690146B_ABST
    Figure CN118690146B_ABST
Patent Text Reader

Abstract

The present invention discloses a method for progressive self-supervised damage identification of high-speed rail wheels and rails in the field of high-speed rail wheel and rail operation and maintenance technology, the method comprising: obtaining a vibration signal to be identified; inputting the vibration signal into a pre-constructed teacher-student network for training until the reconstruction loss function and the prediction loss function constructed based on the vibration signal converge, and obtaining encoder parameters in the trained teacher-student network; migrating the encoder parameters to a pre-constructed recognition network, inputting training samples into the recognition network, and obtaining a trained recognition network; inputting the vibration signal into the trained recognition network, and outputting the prediction probability of the damage type. The present invention can solve the technical problem that the existing self-supervised learning method is limited to designing pre-defined proxy tasks to use unlabeled samples for pre-training, and its effectiveness and rationality can only be verified by the performance of downstream recognition tasks, resulting in weak recognition performance and interpretability.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of high-speed railway wheel and rail intelligent operation and maintenance, and in particular to a high-speed railway wheel and rail progressive self-supervision damage identification method. Background Art

[0002] High-speed railway is one of the main modes of transportation for long-distance travel. The interaction between wheels and rails is the basis of EMU operation. However, with the increase of vehicle speed and axle load, the working condition of the interaction between wheels and rails deteriorates and various damages frequently occur.

[0003] With the development of technologies such as the Internet of Things and sensors, the industry has entered the era of big data. Deep learning has revolutionized many fields with its advantages in processing big data and automatically learning representations, but when it is actually applied to high-speed rail wheels and rails, it often faces two problems. On the one hand, the training of recognition models based on deep learning relies on a large number of labeled samples, and labeling these vibration signals is time-consuming and laborious. On the other hand, the operating conditions of high-speed rail are changeable, resulting in differences in the distribution of training data and test data, which will lead to a decrease in the performance of the method in actual scenarios.

[0004] Self-supervised learning methods are usually used to solve the above two problems, which can learn useful representations from unlabeled samples and enhance the representation and discrimination capabilities of the model. However, existing self-supervised damage recognition methods are limited to designing pre-defined proxy tasks to use unlabeled samples for pre-training, and their effectiveness and rationality can only be passively verified by the performance of downstream recognition tasks. There is a lack of direct explanation support, resulting in weak recognition performance and interpretability.

[0005] Therefore, a high-speed rail wheel-rail progressive self-supervised damage identification method is urgently needed to solve the above technical problems. Summary of the invention

[0006] The purpose of the present invention is to overcome the deficiencies in the prior art and provide a method for progressive self-supervised damage identification of high-speed rail wheels and rails, which can solve the technical problems that the existing self-supervised learning methods are limited to designing pre-defined proxy tasks to use unlabeled samples for pre-training, and their effectiveness and rationality can only be passively verified by the performance of downstream recognition tasks, resulting in weak recognition performance and interpretability.

[0007] To achieve the above object, the present invention is implemented by adopting the following technical solutions:

[0008] In a first aspect, the present invention provides a high-speed railway wheel-rail progressive self-supervised damage identification method, comprising:

[0009] Acquire a vibration signal to be identified;

[0010] Inputting the vibration signal into a pre-constructed teacher-student network for training until a reconstruction loss function and a prediction loss function constructed based on the vibration signal converge, and obtaining encoder parameters in the trained teacher-student network;

[0011] Migrating the encoder parameters to a pre-built recognition network, inputting training samples to the recognition network, and obtaining a trained recognition network;

[0012] Inputting the vibration signal into a trained recognition network and outputting a damage type prediction probability;

[0013] The teacher-student network includes a teacher model and a student model with the same structure, and both the teacher model and the student model include an encoder, a decoder and a predictor; the recognition network includes an encoder and a classifier, and the encoders in the teacher-student network and the recognition network are the same.

[0014] Furthermore, the process of obtaining the encoder parameters in the teacher-student network includes:

[0015] Based on the student model in the teacher-student network, the vibration signal is processed by a mask operation to obtain a mask signal;

[0016] Performing characterization learning and reconstruction on the mask signal to obtain a reconstructed signal, and optimizing a reconstruction loss function constructed based on the reconstructed signal and a vibration signal;

[0017] Obtaining a true reconstruction loss according to the reconstructed signal and the vibration signal, predicting the reconstructed signal to obtain a predicted reconstruction loss, and optimizing a predicted loss function constructed based on the predicted reconstruction loss and the true reconstruction loss;

[0018] Based on the teacher model in the teacher-student network, the reconstruction loss of the vibration signal is predicted, and a difficulty mask is generated by sorting the prediction results. The difficulty mask is input into the student model and optimized until the reconstruction loss function and the prediction loss function converge, and the encoder parameters in the trained teacher-student network are obtained.

[0019] Furthermore, the teacher model is updated synchronously with the student model through exponential sliding average.

[0020] Further, processing the vibration signal through a mask operation to obtain a mask signal includes:

[0021] ,

[0022] in, Indicates vibration signal, , The matrix representing the vibration signal, N express N part,L Indicates each segment L data points, Indicates mask signal, mask , is a binary vector whose elements are all 0 or 1;

[0023] Performing characterization learning and reconstruction on the mask signal to obtain a reconstructed signal includes:

[0024] The mask signal is input into the self-attention block in the encoder and decoder, representation learning and signal reconstruction are repeated, and a reconstructed signal is output, where the reconstructed signal is:

[0025] ,

[0026] in, To reconstruct the signal, F represents the encoder, s represents the student model, represents the encoder parameters of the student model, D Represents a decoder, represents the decoder parameters of the student model;

[0027] The reconstruction loss function constructed based on the reconstructed signal and the vibration signal is optimized, and the expression of the reconstruction loss function is:

[0028] ,

[0029] in, is the reconstruction loss, for N A positive integer in the range, b for L A positive integer in the range, Indicated in Dimension, coordinates are ( , b ), Indicated in Dimension, coordinates are ( , b ) vibration signal,

[0030] The reconstructed signal is predicted by the predictor, and the predicted reconstruction loss is expressed as:

[0031] ,

[0032] in, P represents the predictor, denotes the predictor parameters, To predict the reconstruction loss,

[0033] The true reconstruction loss is obtained according to the reconstructed signal and the vibration signal, and the expression of the true reconstruction loss is:

[0034] ,

[0035] in, is the true reconstruction loss,

[0036] The prediction loss function constructed based on the prediction reconstruction loss and the true reconstruction loss is optimized. The expression of the prediction loss function is:

[0037] ,

[0038] ,

[0039] ,

[0040] ,

[0041] in, represents the prediction loss, The coordinates are The average prediction result of The coordinates are The average prediction result of and are two indicator factors, The coordinates are The average true reconstruction loss is The coordinates are The average true reconstruction loss is for N A positive integer in the range, the dimension in the prediction loss function is , represents coordinates, Representation and Different coordinates, is the activation function, is an exponential function.

[0042] Furthermore, the self-attention block includes a residual-connected layer normalization layer, a multi-head self-attention layer, a random dropout layer, and a multi-layer perceptron layer, and the output of the self-attention block is expressed as:

[0043] ,

[0044] ,

[0045] in, () indicates the normalization layer. represents the intermediate features of the self-attention block, represents the final feature of the self-attention block, MSA () represents a multi-head self-attention layer, Dropout () represents the random dropout layer, MLP () represents a multi-layer perceptron layer.

[0046] Furthermore, generating a difficulty mask by sorting the prediction results includes:

[0047] The prediction result is expressed as:

[0048] ,

[0049] in, Represents the prediction result, F represents the encoder, represents the encoder parameters, P represents the predictor, denotes the predictor parameters, t represents the teacher model, Indicates vibration signal;

[0050] The prediction results are sorted to generate a difficulty mask, and the generation process of the difficulty mask is expressed as:

[0051] ,

[0052] in, represents the index for generating the difficulty mask, Indicates the index of the maximum value of the predicted result to be returned. represents the proportion of difficult masks, N express N part, for N A positive integer in the range, The dimension is N , the coordinates are The average prediction result of Represents the number of coordinates. The proportion of difficult masks increases linearly during the training process, generating masks from easy to difficult.

[0053] Furthermore, the training process of the recognition network includes:

[0054] The encoder parameters are transferred to the recognition network, and the expression for inputting the training sample to obtain the extracted features is:

[0055] ,

[0056] in, C represents the classification layer, represents the encoder, represents the encoder parameters of the migration, Indicates that the dimension is N and the coordinates are i The extracted features, represents the training sample,

[0057] The predicted probability that the training sample belongs to different damage types is calculated based on the extracted features, and the calculation formula is:

[0058] ,

[0059] in, represents the predicted probability corresponding to the extracted feature, represents the total number of damage categories, represents the activation function of the extracted features, i and j are all positive integers representing numbers, The dimension is N , the coordinates are j Extraction features of

[0060] Calculate the cross entropy loss between the true probability and the predicted probability:

[0061]

[0062] in, A one-hot vector representing the true probability, is the cross entropy loss;

[0063] The gradient of the cross entropy loss with respect to the recognition network parameters is calculated, the cross entropy loss is optimized by back propagation and gradient descent algorithm, and the network parameters are updated until the loss converges or the set number of iterations is reached to obtain a trained recognition network.

[0064] In a second aspect, the present invention provides a high-speed railway wheel-rail progressive self-supervision damage identification device, comprising:

[0065] Acquisition module: obtains the vibration signal to be identified;

[0066] A teacher-student network training module is used to input the vibration signal into a pre-constructed teacher-student network for training until a reconstruction loss function and a prediction loss function constructed based on the vibration signal converge, thereby obtaining encoder parameters in the trained teacher-student network;

[0067] A recognition network training module, used to transfer the encoder parameters to a pre-built recognition network, input training samples to the recognition network, and obtain a trained recognition network;

[0068] The recognition module is used to input the vibration signal into the trained recognition network and output the damage type prediction probability.

[0069] In a third aspect, the present invention provides an electronic terminal, comprising a processor and a memory connected to the processor, wherein a computer program is stored in the memory, and when the computer program is executed by the processor, the steps of any of the above methods are performed.

[0070] In a fourth aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, which implements the steps of any of the above methods when executed by a processor.

[0071] Compared with the prior art, the present invention has the following beneficial effects:

[0072] The progressive self-supervised high-speed railway wheel-rail damage identification method described in the present invention enhances the representation ability of the network by establishing a teacher-student network, and predicts the reconstruction loss of the mask reconstruction task at the same time, further enhancing the network's understanding of the signal, and actively generates masks from easy to difficult according to the prediction results, thereby constructing an effective proxy task for pre-training, which can complete the type recognition of high-speed railway wheel-rail damage under few labels and variable working conditions, and improve the recognition performance in actual scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0073] Figure 1 It is a flowchart of a high-speed railway wheel-rail progressive self-supervisory damage identification method provided in the first embodiment of the present invention;

[0074] Figure 2 It is a model schematic diagram of an encoder, a decoder and a predictor of a high-speed railway wheel-rail progressive self-supervised damage identification method provided in the first embodiment of the present invention;

[0075] Figure 3 It is a schematic diagram of an identification network model of a high-speed railway wheel-rail progressive self-supervised damage identification method provided in the first embodiment of the present invention;

[0076] Figure 4 It is a partial structural diagram of a high-speed railway wheel-rail progressive self-supervision damage identification device provided in Example 2 of the present invention. DETAILED DESCRIPTION

[0077] The technical solution of the present invention is described in detail below through the accompanying drawings and specific embodiments. It should be understood that the embodiments of the present application and the specific features in the embodiments are detailed descriptions of the technical solution of the present application, rather than limitations on the technical solution of the present application. In the absence of conflict, the embodiments of the present application and the technical features in the embodiments can be combined with each other.

[0078] The term "and / or" in this article is only a description of the association relationship of the associated objects, indicating that there can be three relationships. For example, A and / or B can mean: A exists alone, A and B exist at the same time, and B exists alone. In addition, the character " / " in this article generally indicates that the associated objects before and after are in an "or" relationship.

[0079] In machine learning and deep learning, "one-hot encoding" is a method of converting categorical variables into numerical form. This is often used to handle categorical features in models, because many algorithms and models can only process numerical inputs.

[0080] One-Hot Vector is a method to convert categorical variables into numerical form, allowing machine learning models to process non-numerical data.

[0081] Non-overlap segmentation: Non-overlap segmentation of vibration signals usually means that during signal processing, the signal is segmented into multiple consecutive segments with no overlapping parts between each segment. It allows to obtain higher resolution in the frequency domain while reducing signal distortion caused by window functions;

[0082] In practical applications, the method to be chosen depends on the characteristics of the signal and the analysis objectives. If non-overlapping segmentation is required, MODWT or maximum overlap discrete wavelet transform can be considered, which can process signals of arbitrary length and avoid a series of problems caused by overlap.

[0083] Embodiment 1:

[0084] Figure 1 This is a flow chart of the high-speed rail wheel-rail progressive self-supervision damage identification method in the first embodiment of the present invention. This flow chart only shows the logical sequence of the method described in this embodiment. Under the premise of no conflict, in other possible embodiments of the present invention, different Figure 1 The steps shown or described are accomplished in the order shown.

[0085] The high-speed rail wheel-rail progressive self-supervision damage identification method provided in this embodiment can be applied to a terminal and can be executed by a high-speed rail wheel-rail progressive self-supervision damage identification device, which can be implemented by software and / or hardware, and can be integrated in a terminal, for example: any smart phone, tablet computer or computer device with communication function. Figure 1 , Figure 2 and Figure 3 As shown, the method of this embodiment specifically includes the following steps:

[0086] Step 1: Obtain the vibration signal to be identified and use sensors to collect the vibration signals of high-speed rail wheels and rails under different working conditions.

[0087] Step 2: Segment the collected vibration signals without overlap according to the preset segmentation length to obtain an unlabeled dataset , the samples with known damage types in the unlabeled dataset are calibrated according to the damage type to obtain a labeled dataset ,in n , m Represents the number of samples in the unlabeled data set and the labeled data set respectively. The unlabeled data set can be used as a test set and a training set, and the labeled data set is used as a training set. When collecting data, the one-dimensional data can also be reshaped into two-dimensional data. , so that it matches the network input.

[0088] Step three: input the vibration signal into a pre-constructed teacher-student network for training until the reconstruction loss function and the prediction loss function constructed based on the vibration signal converge, and obtain the encoder parameters in the trained teacher-student network, wherein the teacher-student network includes a teacher model and a student model with the same structure, and both the teacher model and the student model include an encoder, a decoder and a predictor.

[0089] Based on the student model in the teacher-student network, processing the vibration signal through a mask operation to obtain a mask signal includes:

[0090] ,

[0091] in, Indicates vibration signal, , The matrix representing the vibration signal, N express N part, L Indicates each segment L data points, Indicates mask signal, mask , is a binary vector whose elements are all 0 or 1.

[0092] The mask signal is characterized and reconstructed to obtain a reconstructed signal, and a reconstruction loss function constructed based on the reconstructed signal and the vibration signal is optimized:

[0093] The mask signal is input into the self-attention blocks in the encoder and decoder.

[0094] The self-attention block includes a residual connection layer normalization layer, a multi-head self-attention layer, a random dropout layer and a multi-layer perceptron layer, and the output of the self-attention block is expressed as:

[0095] ,

[0096] ,

[0097] in, () indicates the normalization layer. represents the intermediate features of the self-attention block, represents the final feature of the self-attention block, MSA () represents a multi-head self-attention layer, Dropout () represents the random dropout layer, MLP () represents a multi-layer perceptron layer.

[0098] The multi-layer perceptron layer includes a GeLU activation function, and the calculation formula of the GeLU activation function is:

[0099] ,

[0100] ,

[0101] Wherein, k is the intermediate feature in the multi-layer perceptron layer;

[0102] That is, the mask signal is repeatedly represented and reconstructed through multiple self-attention blocks, and finally the reconstructed signal is output. The reconstructed signal is:

[0103] ,

[0104] in, To reconstruct the signal, F represents the encoder, represents the encoder parameters, D Represents a decoder, Represents decoder parameters.

[0105] The reconstruction loss function constructed based on the reconstructed signal and the vibration signal is optimized, and the expression of the reconstruction loss function is:

[0106] ,

[0107] in, is the reconstruction loss, for N A positive integer in the range, b for L A positive integer in the range, Indicated in Dimension, coordinates are ( , b ), Indicated in Dimension, coordinates are ( , b ) vibration signal.

[0108] The reconstructed signal is predicted by the predictor, and the predicted reconstruction loss is expressed as:

[0109] ,

[0110] in, P represents the predictor, denotes the predictor parameters, Reconstruct the loss for the prediction.

[0111] The true reconstruction loss is obtained according to the reconstructed signal and the vibration signal, and the expression of the true reconstruction loss is:

[0112] ,

[0113] in, is the true reconstruction loss.

[0114] The prediction loss function constructed based on the prediction reconstruction loss and the true reconstruction loss is optimized. The expression of the prediction loss function is:

[0115] ,

[0116] ,

[0117] ,

[0118] ,

[0119] in, represents the prediction loss, The coordinates are The average prediction result of The coordinates are The average prediction result of and are two indicator factors, The coordinates are The average true reconstruction loss is The coordinates are The average true reconstruction loss is for N A positive integer in the range, the dimension in the prediction loss function is , represents coordinates, Representation and Different coordinates, is the activation function, As for the indicator factor, it can be found from the formula that one of the two indicator factors must be 0, which can ensure that when the predicted reconstruction loss is consistent with the true reconstruction loss, the predicted loss is small, and when they are inconsistent, the loss is large.

[0120] Generating difficulty masks by sorting the prediction results includes:

[0121] The prediction result is expressed as:

[0122] ,

[0123] in, Represents the prediction result, F represents the encoder, represents the encoder parameters, P represents the predictor, denotes the predictor parameters, t represents the teacher model, Indicates vibration signal.

[0124] The prediction results are sorted to generate a difficulty mask, and the generation process of the difficulty mask is expressed as:

[0125] ,

[0126] in, represents the index for generating the difficulty mask, Indicates the index of the maximum value of the predicted result to be returned. represents the proportion of difficult masks, N express N part, for N A positive integer in the range, The dimension is N , the coordinates are i The average prediction result of Represents the number of coordinates. The proportion of difficult masks increases linearly during the training process, generating masks from easy to difficult.

[0127] The difficulty mask is input into the student model and optimized until the reconstruction loss function and the prediction loss function converge, thereby obtaining the encoder parameters in the trained teacher-student network.

[0128] The teacher model is updated synchronously with the student model through exponential sliding average, and other processes are consistent with the student model. The synchronous update process is expressed as:

[0129] ,

[0130] Among them, m=0.999 represents the momentum coefficient, which represents the rate of change of training parameters over time. In order to make the network update smoother, it is usually taken to be close to 1. are the teacher model parameters, are the student model parameters.

[0131] Step 4: The recognition network includes an encoder and a classifier. The encoder in the teacher-student network is the same as that in the recognition network. The encoder parameters are transferred to the pre-constructed recognition network, and the training samples are input into the recognition network to obtain a trained recognition network.

[0132] The training process of the recognition network includes:

[0133] The encoder parameters are transferred to the recognition network, and the expression for inputting the training sample to obtain the extracted features is:

[0134] ,

[0135] in, C represents the classification layer, represents the encoder, represents the encoder parameters of the migration, The dimension is N , the coordinates are i The extracted features, Represents training samples.

[0136] The predicted probability that the training sample belongs to different damage types is calculated based on the extracted features, and the calculation formula is:

[0137] ,

[0138] in, represents the predicted probability corresponding to the extracted feature, represents the total number of damage categories, represents the activation function of the extracted features, i and j are all positive integers representing numbers, The dimension is N , the coordinates are j The extracted features.

[0139] Calculate the cross entropy loss between the true probability and the predicted probability:

[0140] ,

[0141] in, A one-hot vector representing the true probability, is the cross entropy loss.

[0142] The gradient of the cross entropy loss with respect to the recognition network parameters is calculated, the cross entropy loss is optimized by back propagation and gradient descent algorithm, and the network parameters are updated until the loss converges or the set number of iterations is reached to obtain a trained recognition network. The specific process of training the recognition network according to the cross entropy loss belongs to the existing mature technology and will not be repeated here.

[0143] Step 5: Input the vibration signal into the trained recognition network and output the predicted probability of the damage type.

[0144] The present embodiment is further described below with reference to the accompanying drawings and experimental cases:

[0145] 1. Experimental data

[0146] The data used in the experiment comes from the wheelset bearing damage simulation test bench, which can simulate the structural and motion relationship between the vehicle wheelset and the track. The data set contains seven health states: normal state, outer ring damage with damage size of 0.2mm and 0.4mm, inner ring damage with damage size of 0.2mm and 0.4mm, and roller damage with damage size of 0.2mm and 0.4mm. The data of each health state was collected at a speed of 600r / min, and three different load conditions were simulated, namely 0.8kN, 1.6kN and 2.4kN. Three working conditions were defined and represented by A, B and C respectively, as shown in Table 1. Each health state under each working condition contains 200 samples, and the length of each sample is set to 4096 data points to ensure the amount of information for at least one rotation of the bearing.

[0147] Table 1: Sample information for different working conditions

[0148]

[0149] 2. Method Validation

[0150] In order to verify the effectiveness of the proposed invention, this experimental case implemented three wheel-rail damage identification tasks under few labels and variable working conditions. To simulate the variable working condition, the data under two working conditions are set as training data, and the data under the remaining working condition is set as test data. To simulate the few-label situation, the data under one working condition in the training data is set to have 10% labeled data, which is indicated in bold. In order to verify the effectiveness of the progressive self-supervised high-speed rail wheel-rail damage identification method and system proposed in the present invention, it is compared with four other advanced wheel-rail damage identification methods based on self-supervised learning, including SimMIM, contextual autoencoder (CAE), masked autoencoder (MAE), and denoising autoencoder (DAE). The experimental settings of each method are consistent, where the initial learning rate of the teacher-student network pre-training is 0.001, the batch size is 64, and the training cycle is 50; the initial learning rate of the recognition network training is 0.0001, the batch size is 32, and the training cycle is 200.

[0151] Compared with other wheel-rail damage identification methods, the identification accuracy and standard deviation of the proposed method have achieved the best results in different identification tasks. The comparison results are shown in Table 2.

[0152] Table 2:

[0153]

[0154] Embodiment 2:

[0155] like Figure 4 As shown, the second embodiment of the present invention provides a high-speed railway wheel-rail progressive self-supervision damage identification device, comprising:

[0156] Acquisition module: obtains the vibration signal to be identified;

[0157] A teacher-student network training module is used to input the vibration signal into a pre-constructed teacher-student network for training until a reconstruction loss function and a prediction loss function constructed based on the vibration signal converge, thereby obtaining encoder parameters in the trained teacher-student network;

[0158] In this embodiment, the teacher-student network training module includes a progressive mask generation module for sorting the prediction results to generate a progressive difficulty mask;

[0159] A recognition network training module, used to transfer the encoder parameters to a pre-built recognition network, input training samples to the recognition network, and obtain a trained recognition network;

[0160] The recognition module is used to input the vibration signal into the trained recognition network and output the damage type prediction probability.

[0161] The high-speed railway wheel-rail progressive self-supervision damage identification device provided in the second embodiment of the present invention can execute the high-speed railway wheel-rail progressive self-supervision damage identification method provided in the first embodiment of the present invention, and has the corresponding functional modules and beneficial effects of the execution method.

[0162] Embodiment three:

[0163] Embodiment 3 of the present invention further provides an electronic terminal, comprising a processor and a memory connected to the processor, wherein a computer program is stored in the memory, and the processor is configured to operate according to the instructions to execute the steps of the method described in embodiment 1.

[0164] The electronic terminal provided in the third embodiment of the present invention can execute the high-speed railway wheel-rail progressive self-supervision damage identification method provided in the first embodiment of the present invention, and has the corresponding functional modules and beneficial effects of the execution method.

[0165] Embodiment 4:

[0166] Embodiment 4 of the present invention further provides a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, the steps of the method described in Embodiment 1 are implemented, and the computer program has functional modules and beneficial effects corresponding to the execution method.

[0167] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, devices, or computer program products. Therefore, the present application may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present application may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program codes.

[0168] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (apparatus), and computer program products according to the embodiments of the present application. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0169] These computer program instructions may also be stored in a computer readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture including an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0170] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0171] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the technical principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A progressive self-supervised damage identification method for high-speed rail wheels and rails, It is characterized in that include: Acquire a vibration signal to be identified; Inputting the vibration signal into a pre-constructed teacher-student network for training until a reconstruction loss function and a prediction loss function constructed based on the vibration signal converge, and obtaining encoder parameters in the trained teacher-student network; The process of obtaining the encoder parameters in the teacher-student network includes: Based on the student model in the teacher-student network, the vibration signal is processed by a mask operation to obtain a mask signal; Performing characterization learning and reconstruction on the mask signal to obtain a reconstructed signal, and optimizing a reconstruction loss function constructed based on the reconstructed signal and a vibration signal; Obtaining a true reconstruction loss according to the reconstructed signal and the vibration signal, predicting the reconstructed signal to obtain a predicted reconstruction loss, and optimizing a predicted loss function constructed based on the predicted reconstruction loss and the true reconstruction loss; Based on the teacher model in the teacher-student network, the reconstruction loss of the vibration signal is predicted, and a difficulty mask is generated by sorting the prediction results, and the difficulty mask is input into the student model and optimized until the reconstruction loss function and the prediction loss function converge, so as to obtain the encoder parameters in the trained teacher-student network; Migrating the encoder parameters to a pre-built recognition network, inputting training samples to the recognition network, and obtaining a trained recognition network; Inputting the vibration signal into a trained recognition network and outputting a damage type prediction probability; The teacher-student network includes a teacher model and a student model with the same structure, and both the teacher model and the student model include an encoder, a decoder and a predictor; the recognition network includes an encoder and a classifier, and the encoders in the teacher-student network and the recognition network are the same; Processing the vibration signal through a mask operation to obtain a mask signal includes: , in, Indicates vibration signal, , The matrix representing the vibration signal, N express N part, L Indicates each segment L data points, Indicates the mask signal, mask , is a binary vector whose elements are all 0 or 1; Performing characterization learning and reconstruction on the mask signal to obtain a reconstructed signal includes: The mask signal is input into the self-attention block in the encoder and decoder, representation learning and signal reconstruction are repeated, and a reconstructed signal is output, where the reconstructed signal is: , in, To reconstruct the signal, F represents the encoder, s represents the student model, represents the encoder parameters of the student model, D Represents a decoder, represents the decoder parameters of the student model; The reconstruction loss function constructed based on the reconstructed signal and the vibration signal is optimized, and the expression of the reconstruction loss function is: , in, is the reconstruction loss, for N A positive integer in the range, b for L A positive integer in the range, Indicated in Dimension, coordinates are ( , b ), Indicated in Dimension, coordinates are ( , b )’s vibration signal; The reconstructed signal is predicted by the predictor, and the predicted reconstruction loss is expressed as: , in, P represents the predictor, denotes the predictor parameters, Reconstruct loss for prediction; The true reconstruction loss is obtained according to the reconstructed signal and the vibration signal, and the expression of the true reconstruction loss is: , in, is the true reconstruction loss, The prediction loss function constructed based on the prediction reconstruction loss and the true reconstruction loss is optimized. The expression of the prediction loss function is: , , , , in, represents the prediction loss, The coordinates are The average prediction result of The coordinates are The average prediction result of and are two indicator factors, The coordinates are The average true reconstruction loss is The coordinates are The average true reconstruction loss is for N A positive integer in the range, the dimension in the prediction loss function is , represents coordinates, Representation and Different coordinates, is the activation function, is an exponential function.

2. According to claim 1, a high-speed railway wheel-rail progressive self-supervised damage identification method, It is characterized in that The teacher model is updated synchronously with the student model through exponential sliding average.

3. According to the method for progressive self-supervised damage identification of high-speed railway wheels and rails according to claim 1, It is characterized in that The self-attention block includes a residual connection layer normalization layer, a multi-head self-attention layer, a random dropout layer and a multi-layer perceptron layer. The output of the self-attention block is expressed as: , , in, () indicates the normalization layer. represents the intermediate features of the self-attention block, represents the final feature of the self-attention block, MSA () represents a multi-head self-attention layer, Dropout () represents the random dropout layer, MLP () represents a multi-layer perceptron layer.

4. According to claim 1, a high-speed railway wheel-rail progressive self-supervised damage identification method, It is characterized in that Generating difficulty masks by sorting the prediction results includes: The prediction result is expressed as: , in, Represents the prediction result, F represents the encoder, represents the encoder parameters, P represents the predictor, denotes the predictor parameters, t represents the teacher model, Indicates vibration signal; The prediction results are sorted to generate a difficulty mask, and the generation process of the difficulty mask is expressed as: , in, represents the index for generating the difficulty mask, Indicates the index of the maximum value of the predicted result to be returned. represents the proportion of difficult masks, N express N part, for N A positive integer in the range, The dimension is N , the coordinates are The average prediction result of Represents the number of coordinates. The proportion of difficult masks increases linearly during the training process, generating masks from easy to difficult.

5. According to claim 1, a high-speed railway wheel-rail progressive self-supervised damage identification method, It is characterized in that The training process of the recognition network includes: The encoder parameters are transferred to the recognition network, and the expression for inputting the training sample to obtain the extracted features is: , in, C represents the classification layer, represents the encoder, represents the encoder parameters of the migration, Indicates that the dimension is N and the coordinates are i The extracted features, represents training samples; The predicted probability that the training sample belongs to different damage types is calculated based on the extracted features, and the calculation formula is: , in, represents the predicted probability corresponding to the extracted feature, represents the total number of damage categories, represents the activation function of the extracted features, i and j are all positive integers representing numbers, The dimension is N , the coordinates are j Extraction features of Calculate the cross entropy loss between the true probability and the predicted probability: , in, A one-hot vector representing the true probability, is the cross entropy loss; The gradient of the cross entropy loss with respect to the recognition network parameters is calculated, the cross entropy loss is optimized by back propagation and gradient descent algorithm, and the network parameters are updated until the loss converges or the set number of iterations is reached to obtain a trained recognition network.

6. A high-speed railway wheel-rail progressive self-supervised damage identification device, based on the high-speed railway wheel-rail progressive self-supervised damage identification method according to any one of claims 1 to 5, It is characterized in that include: Acquisition module: obtains the vibration signal to be identified; A teacher-student network training module is used to input the vibration signal into a pre-constructed teacher-student network for training until a reconstruction loss function and a prediction loss function constructed based on the vibration signal converge, thereby obtaining encoder parameters in the trained teacher-student network; A recognition network training module, used to transfer the encoder parameters to a pre-built recognition network, input training samples to the recognition network, and obtain a trained recognition network; The recognition module is used to input the vibration signal into the trained recognition network and output the damage type prediction probability.

7. An electronic terminal, It is characterized in that comprising a processor and a memory connected to the processor, The memory stores a computer program, and when the computer program is executed by the processor, the steps of the method according to any one of claims 1 to 5 are performed.

8. A computer-readable storage medium having a computer program stored thereon, It is characterized in that When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 5 are implemented.