A high-speed rail wheel-rail sensing adversarial learning damage identification method
Through the high-speed rail wheel and rail perception and anti-learning damage recognition method, the high-dimensional features of the high-speed rail wheel and rail system are extracted using pre-trained models and open damage classifiers, solving the problem that deep learning is difficult to identify new damage, and improving the damage of the high-speed rail wheel and rail system Identify accuracy and security.
Patent Information
- Application Number
- CN202411165214.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-23
- Publication Date
- 2024-12-13
- Estimated Expiration
- 2044-08-23
AI Technical Summary
Deep learning methods are difficult to accurately identify new types of damage in the identification of damage of high-speed rail wheel and rail systems, resulting in a decrease in accuracy and affecting the safety and efficiency of high-speed rail operation.
The high-speed rail wheel-rail perceived adversarial learning damage recognition method is adopted to extract the high-dimensional features of the vibration signal through pre-training models, and use open damage classifiers and feature projectors to calculate boundary adversarial losses and open contrast perceived losses to establish decision boundaries to distinguish them. There are already injuries and new injuries.
It effectively improves the accuracy of online diagnosis of high-speed rail wheel and rail systems, breaks through the problem that traditional deep learning is difficult to distinguish new damage from existing damage, improves the ability to extract cross-working conditions status, and ensures the safety and efficiency of high-speed rail operation.
Smart Images
Figure CN118690276B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of high-speed rail wheel-rail intelligent operation and maintenance, and in particular to a high-speed rail wheel-rail perception adversarial learning damage identification method. Background Art
[0002] With the increase of high-speed train mileage, the damage identification of high-speed rail track system becomes more and more important. On high-speed trains, the damage of track system may have a serious impact on the safety and operation efficiency of the train.
[0003] In the era of industrial big data, traditional inspection methods are time-consuming and labor-intensive, and can no longer meet the needs of accurate identification of high-speed rail wheel and rail damage. Therefore, by combining machine learning, artificial intelligence and other technologies, it has become a new application direction to achieve real-time monitoring and identification of damage to the high-speed rail track system, thereby preventing accidents in a timely manner and improving the safety and efficiency of high-speed rail operations.
[0004] However, the high-speed rail wheel-rail system damage identification method based on deep learning is limited in two aspects in actual engineering applications. On the one hand, the operating conditions of high-speed rail are changeable, and the high-speed rail wheel-rail system status data collected under different conditions have different distributions. The distribution difference poses a challenge to the extraction of high-speed rail wheel-rail system status characteristics. On the other hand, the historical loss type data collected cannot cover all loss types. When the model is used for online identification, new types of damage encountered will be mistakenly identified as existing damage types. This reduces the accuracy of high-speed rail wheel-rail system damage identification through deep learning, and the safety and efficiency of high-speed rail operation are difficult to guarantee.
[0005] Therefore, a high-speed rail wheel-rail sensing adversarial learning 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 to provide a high-speed rail wheel-rail perception adversarial learning damage identification method, which can solve the technical problem that when high-speed rail wheel-rail system damage identification is performed through deep learning, new types of damage will be mistakenly identified as existing damage types, resulting in a decrease in accuracy, making it difficult to ensure the safety and efficiency of high-speed rail operation.
[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 rail wheel-rail sensing adversarial learning damage identification method, comprising:
[0009] Acquire a vibration signal to be identified;
[0010] Inputting the vibration signal into a pre-trained high-speed railway wheel-rail damage recognition model, wherein the high-speed railway wheel-rail damage recognition model includes a deep feature extractor and an open damage classifier;
[0011] Using the high-speed rail wheel-rail damage identification model, high-dimensional features of the vibration signal are extracted by a deep feature extractor, multi-dimensional prediction probability data corresponding to multiple damage types are obtained by an open damage classifier based on the high-dimensional features, and damage type identification results are determined according to the multi-dimensional prediction probability data corresponding to the multiple damage types;
[0012] The dimension of the predicted probability obtained by the open damage classifier based on the high-dimensional features is greater than the total number of damage type labels of the training samples, and the multiple damage types include the damage types of the training samples and new damage types.
[0013] Furthermore, the training process of the high-speed railway wheel-rail damage identification model includes:
[0014] Performing non-overlap segmentation on the training samples according to a preset segmentation length to obtain state samples, and calibrating the state samples according to known damage types to obtain a labeled training data set;
[0015] Inputting the labeled training data set into a deep feature extractor to obtain high-dimensional features;
[0016] Inputting the high-dimensional features into the open damage classifier to obtain damage type prediction probabilities corresponding to multiple known damage types;
[0017] The boundary adversarial loss is calculated according to the damage type prediction probability, the cross entropy loss is calculated based on the damage type prediction probability and the actual damage type label in the labeled training data set, and the cross entropy loss and the boundary adversarial loss are minimized according to the back propagation and optimization algorithm to complete the training.
[0018] Furthermore, the calculation formula for calculating the boundary resistance loss according to the damage type prediction probability is:
[0019] ,
[0020] in, The predicted probability of damage type indicating that the high-dimensional feature belongs to a new damage type, K is the total number of known damage type labels, represents the boundary adversarial loss, represents the network parameters of the deep feature extractor; represents the network parameters of the open damage classifier;
[0021] The calculation formula for calculating the cross entropy loss based on the predicted probability of the damage type and the actual damage type label in the labeled training data set includes:
[0022] ,
[0023] ,
[0024] In the formula, is the cross entropy loss, The unique hot vector representing the true damage type label of the nth state sample, represents the predicted probability of the damage type of the nth state sample, Indicates the total number of state samples.
[0025] Furthermore, the high-speed railway wheel-rail damage recognition model further includes a feature projector; and the training process of the high-speed railway wheel-rail damage recognition model further includes:
[0026] Input the high-dimensional features and the weight vectors preset in the Softmax layer of the open damage classifier into the feature projector, and output the corresponding projection features, wherein the number of the weight vectors is equal to the total number of the known damage types;
[0027] An open contrast perceptual loss is calculated according to the projected features, the open contrast perceptual loss is minimized based on a back-propagation and optimization algorithm, and a deep feature extractor, an open damage classifier and a feature projector are optimized.
[0028] Further, calculating the open contrast perception loss according to the projection feature includes:
[0029] ,
[0030] In the formula, represents the state sample ordinal number, Indicates the total number of state samples; Represents the projection feature of the high-dimensional feature of the n-th state sample; express The corresponding damage category ordinal number, represents the projection feature of the cth weight vector in the Softmax layer corresponding to the projection feature of the high-dimensional feature of the nth state sample, T represents the transpose of the vector, represents the projection feature of the kth weight vector of the Softmax layer, express The negative state sample ordinal number, express No. negative state sample projection features, express The total number of negative state samples.
[0031] Furthermore, the optimization algorithm adopts one of an adaptive moment estimation algorithm, a stochastic gradient descent algorithm and a root mean square transfer algorithm.
[0032] In a second aspect, the present invention provides a high-speed rail wheel-rail sensing adversarial learning damage identification device, comprising:
[0033] An acquisition module, used for acquiring a vibration signal to be identified;
[0034] An input module, used for inputting the vibration signal into a pre-trained high-speed railway wheel-rail damage recognition model, wherein the high-speed railway wheel-rail damage recognition model includes a deep feature extractor and an open damage classifier;
[0035] An identification module is used to use a high-speed rail wheel-rail damage identification model to extract high-dimensional features of vibration signals through a deep feature extractor, obtain multi-dimensional prediction probability data corresponding to multiple damage types through an open damage classifier based on the high-dimensional features, and determine a damage type identification result based on the multi-dimensional prediction probability data corresponding to the multiple damage types;
[0036] The dimension of the predicted probability obtained by the open damage classifier based on the high-dimensional features is greater than the total number of damage type labels of the training samples, and the multiple damage types include the damage types of the training samples and new damage types.
[0037] 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 the method described in the first aspect are performed.
[0038] In a fourth aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method described in the first aspect.
[0039] Compared with the prior art, the present invention has the following beneficial effects:
[0040] The present invention inputs the vibration signal to be identified into a pre-built high-speed railway wheel-rail damage identification model, outputs multi-dimensional prediction probability data corresponding to multiple damage types, and determines the damage type identification result according to the multi-dimensional prediction probability data corresponding to the multiple damage types, and judges whether the vibration signal to be identified contains a new type of damage, thereby realizing effective online diagnosis of the high-speed railway wheel-rail system;
[0041] By minimizing the boundary adversarial loss obtained in the high-speed rail wheel-rail damage identification model, the decision boundary between existing damage types and new damage types can be established without the target working condition data set participating in the model training, thus overcoming the difficulty of establishing the boundary between existing damage types and new loss types in traditional deep learning damage identification methods.
[0042] It is able to learn highly clustered cross-operating condition features through the designed open contrast-aware loss, overcoming the defect that traditional deep learning damage identification methods are difficult to extract cross-operating condition features, and effectively improving the accuracy of high-speed rail wheel-rail damage identification. BRIEF DESCRIPTION OF THE DRAWINGS
[0043] Figure 1 It is a flowchart of a high-speed railway wheel-rail sensing adversarial learning damage identification method provided in the first embodiment of the present invention;
[0044] Figure 2 It is a schematic diagram of the structure of a high-speed rail wheel-rail damage recognition model according to a high-speed rail wheel-rail perception adversarial learning damage recognition method provided in the first embodiment of the present invention;
[0045] Figure 3 It is a schematic diagram of online recognition of an unlabeled test data set of a high-speed railway wheel-rail perception adversarial learning damage recognition method provided in the first embodiment of the present invention;
[0046] Figure 4 It is a schematic diagram of a confusion matrix of a health status result of an unlabeled test data set predicted by a high-speed railway wheel-rail sensing adversarial learning damage identification method provided in the first embodiment of the present invention;
[0047] Figure 5 It is a schematic diagram of visualized clustering results of high-dimensional features of a labeled training data set and an unlabeled test data set of a high-speed rail wheel-rail perception adversarial learning damage identification method provided in Example 1 of the present invention. DETAILED DESCRIPTION
[0048] 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.
[0049] 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.
[0050] One-Hot Vector is a method to convert categorical variables into numerical form, allowing machine learning models to process non-numerical data.
[0051] Gradient Reversal Layer (GRL) is a technique used in adversarial training, especially in domain adaptation problems. The core idea is to improve the performance of the open-impairment classifier by reversing the gradients of the classifier and feature extractor during training, while the performance of the deep feature extractor will not be harmed.
[0052] 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;
[0053] 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.
[0054] Embodiment 1:
[0055] Figure 1 This is a flow chart of the high-speed rail wheel-rail sensing adversarial learning 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.
[0056] The high-speed rail wheel-rail sensing adversarial learning 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-supervisory 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. The method of this embodiment specifically includes the following steps:
[0057] Step 1: Collect the vibration signal to be identified under the target working condition of high-speed railway wheel and rail;
[0058] Step 2: inputting the vibration signal into a pre-trained high-speed rail wheel-rail damage recognition model, wherein the high-speed rail wheel-rail damage recognition model includes a deep feature extractor and an open damage classifier; using the high-speed rail wheel-rail damage recognition model, extracting high-dimensional features of the vibration signal through the deep feature extractor, and obtaining multi-dimensional prediction probability data corresponding to multiple damage types through the open damage classifier based on the high-dimensional features;
[0059] like Figure 3 As shown, the vibration signal is input into a pre-trained high-speed railway wheel-rail damage recognition model, and multi-dimensional prediction probability data corresponding to various damage types are obtained through an open damage classifier based on the high-dimensional features;
[0060] like Figure 2 As shown, the training process of the high-speed railway wheel-rail damage identification model includes:
[0061] A gradient flip layer (GRL) is further arranged between the deep feature extractor and the open damage classifier. Specifically, the deep feature extractor and the open damage classifier constitute a feedforward neural network for predicting the damage type prediction probability of the input sample; the deep feature extractor and the feature projector constitute a feedforward neural network for outputting the projection features of the high-dimensional features; the open damage classifier and the feature projector constitute a feedforward neural network for outputting the projection features of the weight vector of the Softmax layer of the open damage classifier;
[0062] Regarding the weight vector of the Softmax layer of the open damage classifier, the following explanation is given:
[0063] The Softmax layer is commonly used in multi-classification problems. It converts a vector or a set of real numbers into a probability distribution. The Softmax function outputs a probability value between 0 and 1 for each category, and the sum of the probabilities of all categories is 1. For an open damage classifier, the weight vector of the Softmax layer is part of the model parameters and is used to represent the probability of mapping the feature vector to each category. In this embodiment, the weight vector is learned and updated through the back propagation algorithm during the training process so that the model can more accurately predict the category of damage.
[0064] The training samples are segmented without overlap according to a preset segmentation length, the training samples may be vibration signals collected under historical working conditions, state samples are obtained, and calibration is performed according to the damage type of the state samples to obtain labeled training data sets of multiple working conditions;
[0065] Inputting the labeled training data set into a deep feature extractor to obtain high-dimensional features;
[0066] Inputting the high-dimensional features and the weight vectors preset in the Softmax layer of the open damage classifier into the feature projector, and outputting corresponding projection features, wherein the total number of known damage types is equal to the total number of weight vectors;
[0067] The high-dimensional features are input into the open damage classifier to obtain damage type prediction probabilities corresponding to multiple known damage types. It should be noted that the damage type prediction probability appearing here is the same as the multi-dimensional prediction probability data appearing in the previous text, under the premise that the vibration signal to be identified is a vibration signal of a known damage type;
[0068] Calculating the open contrast perceptual loss based on the projected features includes:
[0069] ,
[0070] In the formula, represents the state sample ordinal number, Indicates the total number of state samples; Represents the projection feature of the high-dimensional feature of the n-th state sample; express The corresponding damage category ordinal number, represents the projection feature of the cth weight vector in the Softmax layer corresponding to the projection feature of the high-dimensional feature of the nth state sample, T represents the transpose of the vector, represents the projection feature of the kth weight vector of the Softmax layer, express The negative state sample ordinal number, express No. negative state sample projection features, express The total number of negative state samples;
[0071] Minimizing the open contrast perceptual loss based on a back-propagation and optimization algorithm;
[0072] The calculation formula for calculating the boundary confrontation loss based on the predicted probability of the damage type is:
[0073] ,
[0074] in, The predicted probability of damage type indicating that the high-dimensional feature belongs to a new damage type, K is the total number of known damage type labels, represents the boundary adversarial loss, represents the network parameters of the deep feature extractor; represents the network parameters of the open damage classifier;
[0075] The purpose of the above formula is to measure The difference between the value and 1, we want to reduce the loss during the training process to make the output The value is close to 1. For the classifier, we want to update its parameters by minimizing the boundary adversarial loss; for the feature extractor, we want to maximize the boundary adversarial loss to update the parameters of the feature extractor. Through this adversarial training process, we finally get the desired model.
[0076] The calculation formula for calculating the cross entropy loss based on the predicted probability of the damage type and the actual damage type label in the labeled training data set includes:
[0077] ,
[0078] ,
[0079] In the formula, is the cross entropy loss, The unique hot vector representing the true damage type label of the nth state sample, represents the predicted probability of the damage type of the nth state sample, Indicates the total number of state samples.
[0080] The training is completed by optimizing the deep feature extractor, open damage classifier, and feature projector by minimizing the cross entropy loss and boundary adversarial loss according to the back-propagation and optimization algorithm;
[0081] Step three: Determine the damage type identification result based on the multi-dimensional prediction probability data corresponding to the multiple damage types; wherein the dimension of the prediction probability obtained by the open damage classifier based on the high-dimensional features is greater than the total number of damage type labels of the training samples, and the multiple damage types include the damage types of the training samples and new damage types. In this embodiment, the dimension is the K+1 mentioned above, and the total number of damage type labels of the training samples is the K mentioned above. When the dimension is greater than the total number of damage type labels of the training samples, it can be determined that a new damage type exists and an identification conclusion is drawn.
[0082] The optimization algorithm adopts one of an adaptive moment estimation algorithm, a stochastic gradient descent algorithm and a root mean square transfer algorithm, all of which belong to the prior art and will not be described in detail here.
[0083] The back propagation refers to:
[0084] The gradient of the calculated loss is calculated, and then the optimization algorithm updates the network parameters according to the calculated gradient, which can be expressed as follows:
[0085] ,
[0086] ,
[0087] ,
[0088] In the formula, represents the calculation of gradient, denoted as the network parameters of the deep feature extractor, open damage classifier, and feature projector, q represents the number of training iterations, , is the weight parameter, is the learning rate parameter.
[0089] The present embodiment is further described below with reference to the accompanying drawings and experimental cases:
[0090] Taking the bearing damage dataset collected on the wheel-rail test bench as an example, the test bench can simulate the structural and motion relationship between the vehicle wheelset and the track. The bearing vibration signal is divided into 2048 points without overlap. According to the four working conditions described in Table 1, four working condition datasets are obtained. The working condition A, B, and C data are used as training datasets, including 7 damage types: normal state (label 0), inner ring damage (label 1), outer ring damage (label 2), roller damage (label 3), inner ring and roller combined damage (label 4), inner ring and outer ring combined damage (label 5), and outer ring and roller combined damage (label 6). The working condition D dataset is used as the target working condition test dataset. In addition to these 7 damage types, it also includes the 8th outer ring, inner ring and roller combined damage type (label 7). Each damage type contains 200 samples.
[0091] Table 1:
[0092]
[0093] In this embodiment, the deep feature extractor is composed of four one-dimensional convolutional layers and four maximum pooling layers, and each one-dimensional convolutional layer is connected to a ReLU activation function; the open damage classifier is composed of two fully connected layers, ReLU and Softmax activation functions, the output dimension of the Softmax classifier is 8, and the last dimension represents the eighth new type of damage in the target working condition D that is different from the training data set; the feature projector is composed of one fully connected layer.
[0094] It should be noted that the deep feature extractor takes the labeled training data set of working conditions A, B, and C as input, and outputs high-dimensional features with a length of 256; the open damage classifier takes the 256-dimensional features extracted by the deep feature extractor as input, and outputs the prediction probability of the damage type; the feature projector takes the 256-dimensional features extracted by the deep feature extractor and the 8-dimensional weight vector of the Softmax layer of the open damage classifier as input, and outputs the corresponding 256-dimensional projection features respectively.
[0095] The training process of the high-speed rail wheel-rail depth state feature perception learning model specifically includes:
[0096] For the labeled training data sets of working conditions A, B, and C, the cross entropy of the open damage classifier is minimized to optimize the deep feature extractor and the open damage classifier;
[0097] Minimize the open contrast perceptual loss for labeled training data sets of conditions A, B, and C To optimize the deep feature extractor, open damage classifier, and feature projector:
[0098] ,
[0099] In the formula, Indicates the damage type ordinal number, =1 indicates the total number of damage type labels in the training set samples; represents the state sample ordinal number, Indicates the total number of state samples; Represents the projection feature of the high-dimensional feature of the n-th state sample, represents the projection feature of the cth weight vector in the Softmax layer corresponding to the projection feature of the high-dimensional feature of the nth state sample, T represents the transpose of the vector, represents the projection feature of the kth weight vector of the Softmax layer, express The negative state sample ordinal number, express No. negative state sample projection features, express The total number of negative state samples;
[0100] Furthermore, the boundary adversarial learning strategy minimizes the boundary adversarial loss To optimize the deep feature extractor and open damage classifier:
[0101] ,
[0102] In the formula, The probability that the input sample belongs to a new type of damage in the last dimension of the open damage classifier output. In this embodiment, the total number of iterative training is set to 200, and the optimization algorithm adopts the stochastic gradient descent algorithm. When the model reaches the number of iterative training, the training is stopped. Finally, a trained high-speed rail wheel-rail damage recognition model is obtained, which includes a trained deep feature extractor and an open damage classifier.
[0103] In this embodiment, the sample set of the target operating condition D is input as follows: Figure 3 The trained deep feature extractor and open damage classifier shown in the figure can identify the damage type of the sample online.
[0104] Specifically, the target working condition D test data set does not participate in the training of the constructed high-speed railway wheel-rail damage identification model, but is only used to verify the accuracy of the model in identifying sample damage types.
[0105] The confusion matrix of the diagnosis results of the target working condition D test data set by the method of the present invention is as follows: Figure 4 As shown, Figure 4 The horizontal axis is the damage type prediction label, that is, the maximum value of the predicted probability obtains the corresponding damage type ordinal number. It can be seen that the diagnostic accuracy of this method for the 8 damage types of the target working condition data has reached more than 95%, especially the recognition accuracy of the 8th new type of damage (labeled as 7) that is not in the training data set has also reached 90%, indicating that this method has superior performance in cross-working condition open damage identification. Figure 5 It is a visualization of the features extracted from different working condition data samples through t-SNE technology. Figure 5 It can be seen that the features of the same type between the three working condition data involved in the training and the target working condition data can be effectively aligned, the boundaries between different types of features are relatively obvious, and the new damage types are also well separated from the existing damage types. This proves that the method of the present invention can not only obtain cross-working condition generalization knowledge and learn distribution-independent representations, but also learn feature representations that are more useful for classification, thereby extracting more clustered state features.
[0106] Embodiment 2:
[0107] Embodiment 2 of the present invention provides a high-speed railway wheel-rail sensing adversarial learning damage identification device, comprising:
[0108] An acquisition module, used for acquiring a vibration signal to be identified;
[0109] An input module, used for inputting the vibration signal into a pre-trained high-speed railway wheel-rail damage recognition model, wherein the high-speed railway wheel-rail damage recognition model includes a deep feature extractor and an open damage classifier;
[0110] An identification module is used to use a high-speed rail wheel-rail damage identification model to extract high-dimensional features of vibration signals through a deep feature extractor, obtain multi-dimensional prediction probability data corresponding to multiple damage types through an open damage classifier based on the high-dimensional features, and determine a damage type identification result based on the multi-dimensional prediction probability data corresponding to the multiple damage types;
[0111] The dimension of the predicted probability obtained by the open damage classifier based on the high-dimensional features is greater than the total number of damage type labels of the training samples, and the multiple damage types include the damage types of the training samples and new damage types.
[0112] The high-speed rail wheel-rail perception adversarial learning damage identification device provided in an embodiment of the present invention can execute the high-speed rail wheel-rail perception adversarial learning damage identification method provided in an embodiment of the present invention, and has functional modules and beneficial effects corresponding to the execution method.
[0113] Embodiment three:
[0114] 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.
[0115] The electronic terminal provided in the third embodiment of the present invention can execute the high-speed rail wheel-rail perception adversarial learning 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.
[0116] Embodiment 4:
[0117] 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.
[0118] 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 adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present application may adopt the form of a computer program product implemented on one or more computer-usable storage media (using disk storage, CD-ROM, optical storage, etc.) containing computer-usable program codes.
[0119] 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.
[0120] 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 comprising 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.
[0121] 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 in the computer or other programmable device. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.
[0122] 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 high-speed rail wheel-rail sensing adversarial learning damage identification method, characterized in that: include: Acquire a vibration signal to be identified; Inputting the vibration signal into a pre-trained high-speed railway wheel-rail damage recognition model, wherein the high-speed railway wheel-rail damage recognition model includes a deep feature extractor and an open damage classifier; Using the high-speed rail wheel-rail damage identification model, high-dimensional features of the vibration signal are extracted by a deep feature extractor, multi-dimensional prediction probability data corresponding to multiple damage types are obtained by an open damage classifier based on the high-dimensional features, and damage type identification results are determined according to the multi-dimensional prediction probability data corresponding to the multiple damage types; The dimension of the predicted probability obtained by the open damage classifier based on the high-dimensional features is greater than the total number of damage type labels of the training samples, and the multiple damage types include the damage types of the training samples and new damage types.
2. According to claim 1, a high-speed railway wheel-rail perception adversarial learning damage identification method is characterized in that: The training process of the high-speed rail wheel-rail damage identification model includes: Performing non-overlap segmentation on the training samples according to a preset segmentation length to obtain state samples, and calibrating the state samples according to known damage types to obtain a labeled training data set; Inputting the labeled training data set into a deep feature extractor to obtain high-dimensional features; Inputting the high-dimensional features into the open damage classifier to obtain damage type prediction probabilities corresponding to multiple known damage types; The boundary adversarial loss is calculated according to the damage type prediction probability, the cross entropy loss is calculated based on the damage type prediction probability and the actual damage type label in the labeled training data set, and the cross entropy loss and the boundary adversarial loss are minimized according to the back propagation and optimization algorithm to complete the training.
3. A high-speed railway wheel-rail sensing adversarial learning damage identification method according to claim 2, characterized in that: The calculation formula for calculating the boundary confrontation loss based on the predicted probability of the damage type is: ， in, The predicted probability of damage type indicating that the high-dimensional feature belongs to a new damage type, K is the total number of known damage type labels, represents the boundary adversarial loss, represents the network parameters of the deep feature extractor; represents the network parameters of the open damage classifier; The calculation formula for calculating the cross entropy loss based on the predicted probability of the damage type and the actual damage type label in the labeled training data set includes: ， ， In the formula, is the cross entropy loss, The unique hot vector representing the true damage type label of the nth state sample, represents the predicted probability of the damage type of the nth state sample, Indicates the total number of state samples.
4. According to claim 2, a high-speed railway wheel-rail perception adversarial learning damage identification method is characterized in that: The high-speed railway wheel-rail damage identification model also includes a feature projector; The training process of the high-speed rail wheel-rail damage identification model also includes: Input the high-dimensional features and the weight vectors preset in the Softmax layer of the open damage classifier into the feature projector, and output the corresponding projection features, wherein the number of the weight vectors is equal to the total number of the known damage types; An open contrast perceptual loss is calculated according to the projected features, the open contrast perceptual loss is minimized based on a back-propagation and optimization algorithm, and a deep feature extractor, an open damage classifier and a feature projector are optimized.
5. A high-speed railway wheel-rail sensing adversarial learning damage identification method according to claim 4, characterized in that: Calculating the open contrast perceptual loss based on the projected features includes: ， In the formula, represents the state sample ordinal number, Indicates the total number of state samples; Represents the projection features of the high-dimensional features of the n-th state sample; express The corresponding damage category ordinal number, represents the projection feature of the cth weight vector in the Softmax layer corresponding to the projection feature of the high-dimensional feature of the nth state sample, T represents the transpose of the vector, represents the projection feature of the kth weight vector of the Softmax layer, express The negative state sample ordinal number, express No. negative state sample projection features, express The total number of negative state samples.
6. The high-speed railway wheel-rail sensing adversarial learning damage identification method according to claim 2 is characterized in that: The optimization algorithm adopts one of an adaptive moment estimation algorithm, a stochastic gradient descent algorithm and a root mean square transfer algorithm.
7. A high-speed rail wheel-rail sensing adversarial learning damage identification device, characterized in that: include: An acquisition module, used for acquiring a vibration signal to be identified; An input module, used for inputting the vibration signal into a pre-trained high-speed railway wheel-rail damage recognition model, wherein the high-speed railway wheel-rail damage recognition model includes a deep feature extractor and an open damage classifier; An identification module is used to use a high-speed rail wheel-rail damage identification model to extract high-dimensional features of vibration signals through a deep feature extractor, obtain multi-dimensional prediction probability data corresponding to multiple damage types through an open damage classifier based on the high-dimensional features, and determine a damage type identification result based on the multi-dimensional prediction probability data corresponding to the multiple damage types; The dimension of the predicted probability obtained by the open damage classifier based on the high-dimensional features is greater than the total number of damage type labels of the training samples, and the multiple damage types include the damage types of the training samples and new damage types.
8. An electronic terminal, characterized in that: The method comprises 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 the method according to any one of claims 1 to 6 are executed.
9. A computer-readable storage medium having a computer program stored thereon, 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 6 are implemented.