Traction system fault detection method based on PPDM and expert prompt learning

Abstract

The invention provides a traction system fault detection method based on PPDM and expert prompt learning. The method comprises the following steps: S1) obtaining a fault-free data set UN and a noisy fault data set UF from a simulation platform; s2) building a PPDM-based reconstruction model, wherein the model comprises a PDM and an ICM; s3) establishing a domain generalization target network # imgabs0 #, and performing clustering to obtain Pmeta; s4) inputting UN into PDM, obtaining Pi based on Pmeta by # imgabs1 #, reconstructing extracted features and Pi into Uf1, inputting Uf1 into ICM, and calculating to obtain Jth; s5) taking Pi as' K 'and' V ', and taking the UF feature as' Q '; s6) inputting the UF into the PDM and the cross attention network to generate Pm, extracting features and reconstructing the features and the Pm into Uf2, inputting the Uf2 into the ICM, calculating to obtain j, comparing j with Jth, if jgt; if Jth, a fault exists, if jlt; and Jth, no fault exists.

Description

Technical Field

[0001] The embodiments of the present invention relate to the technical field of train fault detection, and in particular to a traction system fault detection method based on PPDM and expert prompt learning. Background Art

[0002] As an important part of modern transportation, high-speed trains provide people with a fast, convenient and comfortable way of travel. The traction control system (TCS) of high-speed trains is one of the core devices to ensure the normal operation of the train. The main function of this system is to control the traction speed of the train by controlling the motor speed and motor output torque.

[0003] With the development of technology, the traction control system of high-speed trains has been continuously improved and perfected. However, due to the complex operating environment and various reasons, high-speed trains may still encounter transient faults during operation. The symptoms are very minor and can be easily masked by external noise and interference. These faults may evolve into major faults, leading to emergency shutdowns or other serious accidents.

[0004] There are many reasons for minor faults in the traction control system of high-speed trains, including manufacturing defects in the equipment itself, poor manufacturing processes, mechanical damage and physical damage caused by long-term wear and use, and aging, damage, looseness and other problems of key components of the system. These factors may lead to minor faults in the traction control system, thus affecting the safety and travel experience of passengers.

[0005] In recent years, data-driven fault detection (FD) has been widely used in the safe operation of traction control systems due to its unique advantages. However, data-driven weak fault detection technology usually requires a large amount of historical data to establish an unsupervised fault detection network and training, which is difficult to obtain in the actual operation of high-speed trains. Currently, data-driven high-speed train fault detection mostly only considers the steady-state operation of the train, and does not perform well for faults under traction and braking conditions. However, traction and braking conditions are important links in train operation and also affect the safe operation of the train. Summary of the invention

[0006] The embodiment of the present invention provides a traction system fault detection method based on PPDM and expert prompt learning, which can effectively improve the model's detection capability for weak faults in complex scenarios and provide strong support for the safe operation of the traction system.

[0007] In a first aspect, an embodiment of the present invention provides a traction system fault detection method based on PPDM and expert prompt learning, comprising the following steps:

[0008] S1), respectively obtain the fault-free data set U from the simulation platform N and noisy fault dataset U F , and the fault-free data set U N and noisy fault dataset U F Divided into training set and test set, set the fault-free data set U N is the source domain, the noisy fault dataset U F is the target domain;

[0009] S2), building a data reconstruction model based on the Pre-Pyramid Deformation Module, wherein the model includes two modules, namely a pyramid deformation module (PDM module) and an information compression module (ICM module);

[0010] S3) Build a domain generalization model target network based on expert prompts The source domain dataset is clustered to obtain the initial meta-prompt P meta = {P1……P N};

[0011] S4), the fault-free data set U N Input to PDM module, PDM module N Perform feature extraction, target network Based on the meta prompt P meta Train to get expert prompt P i = {P E1 ...P EN}, change U N The extracted features and expert prompts are reconstructed into dataset U f1 , input to the ICM module to realize the compression of abnormal information of the data set, remove noise and redundant information, and after the ICM module, use the evaluation function to calculate the reconstruction error of the data set to obtain the threshold J th ;

[0012] S5) Design a network based on the cross attention mechanism and extract the fault-free dataset U N The trained expert prompt P i As the "key" (Key, K) and "value" (Value, V), the fault data set U F The features of are used as “query” (Query, Q);

[0013] S6), the fault data set U FThe PDM module extracts the features of the fault data set, and the cross-attention network generates attention weights by calculating the similarity between "Q" and "V". The expert prompts are weighted and combined according to the weights, and these weighted prompts are input into the target network. In the example above, generate the expert prompt P m = {P FE1 ...P FEN}, extract the features of the fault dataset and the expert prompt P m Reconstructed dataset U f2 , input to the ICM module to remove the noise of the data set, calculate the reconstruction error of the data set through the evaluation function to get the output value j, and compare it with the threshold Jth obtained in the training stage. If j>J th , then the system is judged to have a fault. If j <J th , then the system is judged to be fault-free.

[0014] Preferably, in step S1), the source domain training set U N and the target domain test dataset U F Normalization is performed between [-1,1] to speed up the convergence of the training network. After processing, a one-dimensional real vector is obtained:

[0015]

[0016] Where a is the normal operation data training set U N The number of sample data in U N (i) represents the i-th group of normal operation data training set; n represents the dimension of the signal; b is the fault operation test data set U F The number of sample data in U F (i) represents the i-th signal of the fault running test data set, Indicates signal U F (i) The signal amplitude in the 1st,…,nth dimensions, where n represents the dimension of the signal.

[0017] Preferably, in step S2), an ICM information compression module is designed:

[0018] The information compression module ICM adopts the VQ-Layer layer design method. The reconstructed data set is passed through the encoder to extract high-dimensional feature embedding and obtain the feature representation z e , by finding e recent memory items and quantize them into a single memory feature cube z qThe ICM information compression module learns the reconstructed data set output by the data reconstruction diversity perception module according to the embedded noise data information, and finds the characteristics of the noise data by calculating the L2 minimum distance; the ICM information compression module includes two steps: search "Search" and compression "Comp", searching for noise information in the data set and limiting the compression of the noise information; (including a hyperparameter β and a stop gradient operator SG, and a compression loss Lcom);

[0019] The two steps of searching and compressing in the ICM information compression module are as follows:

[0020] S21) The ICM module receives the input data set and identifies the noise information in the data by finding the feature vector with the minimum L2 distance to the memory unit. The calculation formula is as follows:

[0021]

[0022] Among them, z eh,w is the feature embedded in the dataset, z n is the feature vector in the memory cell;

[0023] S22) Calculate the compression loss through VQ-Layer to reduce the influence of noise data. The compression loss is calculated as follows:

[0024] L com =||SG(Z e )-Z q ||2+β||ze-SG(Z q )||2;

[0025] Among them, SG(·) represents the stop gradient operation, Z e is the embedding representation, Z q is the quantized feature. The hyperparameter β is used to balance the compression effect and the noise removal effect.

[0026] Preferably, step S3) includes the following steps:

[0027] S31) The source domain dataset U N , input into the PDM module to extract a set of feature vectors, which represent the main characteristics of the data;

[0028] S32) clustering the extracted feature vectors to obtain subsets with multiple similar features, using a meta-learning method to find commonalities from these clustered subsets and generate meta-prompts.

[0029] Preferably, step S4) includes the following steps:

[0030] The PDM module performs a fault-free data set UN and noisy fault dataset U F For feature extraction and reconstruction, the PDM module calculates the feature difference between normal data and fault data by setting a specific deformation function φ(x) and reconstructs it. The following function is used for feature extraction:

[0031] φ(x)=f deforn (x)+f norm (x);

[0032] Among them, f deforn represents the convolution operation, which is used to extract local features in time series data. norm Data normalization is performed to ensure the stability of the training network and accelerate convergence. φ(x) represents the process of converting the original input X into feature representation.

[0033] As a preferred embodiment, the PDM module performs a non-fault data set U N and noisy fault dataset U F Perform feature extraction and reconstruction. The PDM module reconstruction process is expressed as:

[0034] The PDM module is based on a given training set sample X = {X1...X N}, the normal data set U obtained by the digital simulation platform is learned N and noisy fault dataset U F The diversity between the data sets obtained by the digital simulation platform and the reconstructed data sets is expressed as The underestimation error caused by the reconstruction process is Compensate again.

[0035] Preferably, in step S4) and step S6), training the expert prompt comprises the following steps:

[0036] S41) Initialize the expert prompt, for each source domain X i , first change the expert prompt P i Initialized to the pre-trained meta-prompt P meta , ensuring that the expert prompts for each source domain are tuned and optimized based on the same initial prompts:

[0037] P i ←P meta ;

[0038] S42) Optimize the expert prompt by minimizing the KL and divergence loss L KL , for each sample in the source domain data (x i,j ,y i,j ) to perform gradient descent optimization:

[0039]

[0040] where α A is the learning rate, L KL Measuring the difference between the model output and the target label;

[0041] S43) Each optimized expert prompt P i Normalize to keep it consistent:

[0042]

[0043] This normalization step ensures that the expert prompts maintain the same scale across different source domains.

[0044] Preferably, the method for training the expert prompt using the cross-attention mechanism method in step S5) is:

[0045] The weighted average method is used to combine the prompts that match the subsets; the weights and the required number of expert prompts are assigned to the subsets of the training data set through the cross-attention mechanism. The expert prompts become the "key" (K) and "value" (V), and the normal operation data set obtained by the digital simulation platform becomes the target attention of the "query" (Q). The training generates the expert prompt P m = {P FE1 ...P FEN}, and the same number of attention weights q1…q corresponding to the expert prompts of the training data subset is calculated. n ; (Weight calculation method q n =VQK T ).

[0046] Preferably, step S6) specifically includes the following steps:

[0047] The expert prompt P generated by the expert prompt module i = {P E1 ...P EN The training data set is obtained from the digital simulation platform as the input of the fault detection framework. The training data set enters the diversity perception module to realize feature amplification, extraction and data set reconstruction. Expert prompts are assigned to the reconstructed data set. The similarity between the reconstructed data set and the training data set is controlled by expert prompts. The reconstructed training data set and the matched expert prompts are used as the input of the ICM information compression module. The information compression module restricts and compresses the abnormal information in the reconstructed data set. The compressed abnormal information is compressed through the loss L. com It indicates that the fault detection threshold j is calculated by (statistical method);

[0048] The test data set obtained by the digital simulation platform is used as the input of the data reconstruction fault detection framework, and the output is the evaluation criteria J th , the evaluation criteria J th Compared with the fault detection threshold j, the fault detection result is as follows:

[0049]

[0050] Preferably, in step S2), an ICM information compression module is designed:

[0051] The information compression module ICM adopts the VQ-Layer layer design method. The reconstructed data set is passed through the encoder to extract high-dimensional feature embedding and obtain the feature representation z e , by finding e recent memory items and quantize them into a single memory feature cube z q The ICM information compression module learns the reconstructed data set output by the data reconstruction diversity perception module according to the embedded noise data information, and finds the characteristics of the noise data by calculating the L2 minimum distance; the ICM information compression module includes two steps: search "Search" and compression "Comp", searching for noise information in the data set and limiting the compression of the noise information; (including a hyperparameter β and a stop gradient operator SG, and a compression loss Lcom);

[0052] The two steps of searching and compressing in the ICM information compression module are as follows:

[0053] S21) The ICM module receives the input data set and identifies the noise information in the data by finding the feature vector with the minimum L2 distance to the memory unit. The calculation formula is as follows:

[0054]

[0055] in, is the feature embedded in the dataset, z n is the feature vector in the memory cell;

[0056] S22) Calculate the compression loss through VQ-Layer to reduce the influence of noise data. The compression loss is calculated as follows:

[0057] L com =||SG(z e )-z q ||2+β||z e -SG(z q )||2;

[0058] Among them, SG(·) represents the stop gradient operation, Z eis the embedding representation, Z q is the quantized feature. The hyperparameter β is used to balance the compression effect and the noise removal effect.

[0059] Preferably, in step S6), the fault detection threshold is calculated, and the specific method is as follows:

[0060] The residual mean calculation formula is:

[0061]

[0062] where x i is the original data, is the reconstructed data, N is the number of data points;

[0063] The formula for calculating the standard deviation of the residuals is:

[0064]

[0065] The threshold calculation formula is:

[0066] j=μ res +η·σ res +βL com ;

[0067] Among them, μ res is the residual mean, σ res is the standard deviation of the residual, η is the confidence parameter, and β is the adjustment hyperparameter.

[0068] The beneficial effects of the present invention are:

[0069] 1. The present invention is based on a traction system fault detection method based on PPDM and expert prompt learning. PPDM realizes efficient feature extraction and processing of fault-free data and noisy fault data. This method effectively solves the problem of insufficient processing of noise data in traditional methods, thereby significantly improving the accuracy and robustness of fault detection.

[0070] 2. The present invention innovatively proposes a domain generalization model based on expert prompts. The expert prompts obtained through training are combined with the cross-attention mechanism to perform weighted combination of the "Q" and "K" features of fault-free data and the "V" features of fault data. The expert prompts are used as auxiliary information and are fused with the fault features to achieve accurate fault location and classification. This method effectively improves the model's ability to detect weak faults in complex scenarios and provides strong support for the safe operation of the traction system. BRIEF DESCRIPTION OF THE DRAWINGS

[0071] Figure 1 is a flow chart of a traction system fault detection method based on PPDM and expert prompt learning provided by an embodiment of the present invention;

[0072] Figure 2 A pyramid deformation module and a target network provided by an embodiment of the present invention Structural diagram;

[0073] Figure 3 This is a structural diagram of an ICM module in the data reconstruction phase provided by an embodiment of the present invention;

[0074] Figure 4 It is a graph of the total loss of overall framework detection provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0075] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0076] In the embodiments of the present invention, the words "further", "exemplarily" or "preferably" are used to indicate examples, illustrations or descriptions, and should not be interpreted as being more preferred or more advantageous than other embodiments or designs. The use of the words "further", "exemplarily" or "optionally" is intended to present related concepts in a specific way.

[0077] In order to more conveniently describe the working principle of the embodiment of the present invention later, an introduction to relevant technical scenarios is first given below.

[0078] In recent years, data-driven fault detection (FD) has been widely used in the safe operation of traction control systems due to its unique advantages. However, data-driven weak fault detection technology usually requires a large amount of historical data to establish an unsupervised fault detection network and training, which is difficult to obtain in the actual operation of high-speed trains. Currently, data-driven high-speed train fault detection mostly only considers the steady-state operation of the train, and does not perform well for faults under traction and braking conditions. However, traction and braking conditions are important links in train operation and also affect the safe operation of the train.

[0079] Based on this, the present invention provides a traction system fault detection method based on PPDM and expert prompt learning, comprising the following steps: S1), respectively obtaining a fault-free data set U from a digital simulation platform N and noisy fault dataset u F , and divide the data set into training set and test set according to a certain ratio, and set the fault-free data set U N is the source domain, the noisy fault dataset U Fas the target domain; S2), building a data reconstruction model based on the Pre-Pyramid Deformation Module, wherein the model includes two modules, the PDM module and the ICM module; S3), building a domain generalization model target network based on expert prompts The source domain dataset is clustered to obtain the initial meta-prompt P meta = {P1……P N}; S4), the fault-free data set U N Input to PDM module, PDM module N Perform feature extraction, target network Based on the meta prompt P meta Train to get expert prompt P i = {P E1 ...P EN}, change U N The extracted features and expert prompts P i Reconstructed into dataset U f1 , input into the ICM module to realize the compression of abnormal information of the data set, remove noise and redundant information, and after the ICM module, use the evaluation function to calculate the reconstruction error of the data set to obtain the threshold J th ; S5), design a network based on the cross attention mechanism, and N The trained expert prompt P i As the "key" (Key, K) and "value" (Value, V), the fault data set U F The features of are used as “query” (Query, Q); S6), the fault data set U F The PDM module extracts the features of the fault data set, and the cross-attention network generates attention weights by calculating the similarity between "Q" and "V". The expert prompts are weighted and combined according to the weights, and these weighted prompts are input into the target network. In the example above, generate the expert prompt P m = {P fE1 ...P FEN}, extract the features of the fault dataset and the expert prompt P m Reconstructed dataset U f2 , input to the ICM module to remove the noise of the data set, calculate the reconstruction error of the data set through the evaluation function to get the output value j, and compare it with the threshold Jth obtained in the training stage. If j>J th , then the system is judged to have a fault. If j <J thBased on this, the embodiment of the present invention can effectively improve the model's ability to detect weak faults in complex scenarios, and provide strong support for the safe operation of the traction system.

[0080] The embodiments of the present invention are further described below in conjunction with the accompanying drawings.

[0081] like Figure 1 As shown, this embodiment provides a traction system fault detection method based on PPDM and expert prompt learning, comprising the following steps:

[0082] S1), obtaining a fault-free data set and a noisy fault data set of a digital simulation platform, and dividing the fault-free data set and the noisy fault data set into a data training set and a test data set;

[0083] Among them, the fault-free data set and the noisy fault operation data set include the transformer secondary side voltage output U n And current output I n , the voltage output U of the supporting capacitor on the DC link cd1 , the voltage output U of the supporting capacitor under the DC link cd2 , inverter a, b, c three-phase current output {I sa ,I sb ,I sc}, motor speed signal V;

[0084] The normal operation data training set obtained by the digital simulation platform is defined as U N ,in

[0085] [U n ,I n ,U cd1 ,U cd2 ,I sa ,I sb ,I sc ,V]∈U N ;

[0086] The fault operation test data set obtained by the digital simulation platform is defined as U F ,in

[0087] [U n ,I n ,U cd1 ,U cd2 ,I sa ,I sb ,I sc ,V]∈U F ;

[0088] For the normal operation data training set U N Run the test dataset U with faults FNormalization is performed between [-1,1] to speed up the convergence of the training network. After processing, a one-dimensional real vector is obtained:

[0089]

[0090] Where a is the normal operation data training set U N The number of sample data in U N (i) represents the i-th group of normal operation data training set; n represents the dimension of the signal;

[0091] b is the fault running test data set U F The number of sample data in U F (i) represents the i-th signal of the fault running test data set, Indicates signal U F (i) The signal amplitude of the 1st,…,nth dimension, where n represents the dimension of the signal;

[0092] b is the fault running test data set U F The number of sample data in U F (i) represents the i-th signal of the fault running test data set, Indicates signal U F (i) The signal amplitude in the 1st,…,nth dimensions, where n represents the dimension of the signal.

[0093] Building a domain generalization model target network based on expert-oriented hints The source domain dataset is clustered to obtain the initial meta-prompt P meta = {P1……P N}.

[0094] A data reconstruction model based on the Pre-Pyramid Deformation Module is constructed, wherein the model includes two modules: a PDM module and an ICM module.

[0095] The fault-free dataset U N Input to PDM module, PDM module N Perform feature extraction, target network Based on the meta prompt P meta Train to get expert prompt P i = {P E1 ...P EN}, change U N The extracted features and expert prompts P i Reconstructed into dataset U f1 , input into the ICM module to realize the compression of abnormal information of the data set, remove noise and redundant information, and after the ICM module, use the evaluation function to calculate the reconstruction error of the data set to obtain the threshold Jth .

[0096] Design a network based on the cross-attention mechanism and N The trained expert prompt P i As the "key" (Key, K) and "value" (Value, V), the fault data set U F The features of are used as “Query” (Q).

[0097] The fault data set U F The PDM module extracts the features of the fault data set, and the cross-attention network generates attention weights by calculating the similarity between "Q" and "V". The expert prompts are weighted and combined according to the weights, and these weighted prompts are input into the target network. In the example above, generate the expert prompt P m = {P FE1 ...P FEN}, extract the features of the fault dataset and the expert prompt P m Reconstructed dataset U f2 , input to the ICM module to remove the noise of the data set, calculate the reconstruction error of the data set through the evaluation function to get the output value j, and compare it with the threshold Jth obtained in the training stage. If j>J th , then the system is judged to have a fault. If j <J th , then the system is judged to be fault-free.

[0098] like Figure 2 As shown, the pyramid module and the target network The effects are as follows:

[0099] As a preferred step S4) and step S6), the PDM module performs a non-fault data set U N and noisy fault dataset U F For feature extraction and reconstruction, the PDM module calculates the feature difference between normal data and fault data by setting a specific deformation function φ(x) and reconstructs it. The following function is used for feature extraction:

[0100] φ(x)=f deforn (x)+f norm (x);

[0101] Among them, f deforn represents the convolution operation, which is used to extract local features in time series data. norm Data normalization is performed to ensure the stability of the training network and accelerate convergence. φ(x) represents the process of converting the original input X into feature representation.

[0102] The PDM module reconstruction process is expressed as:

[0103] The PDM module is based on a given training set sample X = {X1...X N}, the normal data set U obtained by the digital simulation platform is learned N and noisy fault dataset U F The diversity between the data sets obtained by the digital simulation platform and the reconstructed data sets is expressed as The underestimation error caused by the reconstruction process is Again, compensation is made;

[0104] S31) The source domain dataset U N , input into the PDM module to extract a set of feature vectors, which represent the main characteristics of the data;

[0105] S32) clustering the extracted feature vectors to obtain subsets with multiple similar features, using a meta-learning method to find commonalities from the clustered subsets to generate meta-prompts;

[0106] S41) Initialize the expert prompt, for each source domain X i , first change the expert prompt P i Initialized to the pre-trained meta-prompt P meta , ensuring that the expert prompts for each source domain are tuned and optimized based on the same initial prompts:

[0107] P i ←[ meta ;

[0108] S42) Optimize the expert prompt by minimizing the KL and divergence loss L kL , for each sample in the source domain data (x i,j ,y i,j ) to perform gradient descent optimization:

[0109]

[0110] where α A is the learning rate, L KL Measuring the difference between the model output and the target label;

[0111] S43) Each optimized expert prompt P i Normalize to keep it consistent:

[0112]

[0113] This normalization step ensures that the expert prompts maintain the same scale across different source domains.

[0114] As a preferred step S5), the cross attention mechanism is applied as follows:

[0115] The weighted average method is used to combine the prompts that match the subsets; the weights and the required number of expert prompts are assigned to the subsets of the training data set through the cross-attention mechanism. The expert prompts become the "key" (K) and "value" (V), and the normal operation data set obtained by the digital simulation platform becomes the target attention of the "query" (Q). The training generates the expert prompt P m = {P FE1 ...P FEN}, and the same number of attention weights q1…q corresponding to the expert prompts of the training data subset is calculated. n ; (Weight calculation method q n =VQK T ).

[0116] like Figure 3 As shown, the functions of the ICM module are as follows:

[0117] The expert prompt P generated by the expert prompt module i = {P E1 ...P EN The training data set is obtained from the digital simulation platform as the input of the fault detection framework. The training data set enters the diversity perception module to realize feature amplification, extraction and data set reconstruction. Expert prompts are assigned to the reconstructed data set. The similarity between the reconstructed data set and the training data set is controlled by expert prompts. The reconstructed training data set and the matched expert prompts are used as the input of the ICM information compression module. The information compression module restricts and compresses the abnormal information in the reconstructed data set. The compressed abnormal information is compressed through the loss L. com Indicates that the calculated L com The weight value of the expert prompt is calculated by (statistical method) to obtain the fault detection threshold j;

[0118] The test data set obtained by the digital simulation platform is used as the input of the data reconstruction fault detection framework, and the output is the evaluation criteria J th , the evaluation criteria J th Compared with the fault detection threshold j, the fault detection result is as follows:

[0119]

[0120] The information compression module ICM adopts the VQ-Layer layer design method. The ICM information compression module learns the reconstructed data set output by the data reconstruction diversity perception module based on the embedded noise data information, and finds the characteristics of the noise data by calculating the L2 minimum distance; the ICM information compression module includes two steps: search "Search" and compression "Comp", which searches for noise information in the data set and limits the compression of the noise information; (including a hyperparameter β and a stopping gradient operator SG, and a compression loss Lcom).

[0121] The two steps of searching and compressing in the ICM information compression module are as follows:

[0122] S21) The ICM module receives the input data set and identifies the noise information in the data by finding the feature vector with the minimum L2 distance to the memory unit. The calculation formula is as follows:

[0123]

[0124] Among them, z eh,w is the feature embedded in the dataset, z n is the feature vector in the memory cell;

[0125] S22) Calculate the compression loss through VQ-Layer to reduce the influence of noise data. The compression loss is calculated as follows:

[0126] L com =||SG(Z e )-Z q ||2+β||ze-SG(Z q )||2;

[0127] Among them, SG(·) represents the stop gradient operation, Z e is the embedding representation, Z q is the quantized feature. The hyperparameter β is used to balance the compression effect and the noise removal effect.

[0128] As a preferred step S6), the fault detection threshold is calculated, and the specific method is as follows:

[0129] The residual mean calculation formula is:

[0130]

[0131] where x i is the original data, is the reconstructed data, N is the number of data points;

[0132] The formula for calculating the standard deviation of the residuals is:

[0133]

[0134] The threshold calculation formula is:

[0135] j=μ res +η·σ res +βL com ;

[0136] Among them, μ res is the residual mean, σ res is the standard deviation of the residual, η is the confidence parameter, and β is the adjustment hyperparameter.

[0137] like Figure 4 As shown, the detection performance of the model is as follows:

[0138] The validation loss (Vali Loss) of the model tends to be stable during the training process. This loss includes all the losses of the training. The validation loss decreases rapidly in the initial stage, indicating that the model has effectively learned the data features in the early stage. With the increase of training rounds (epochs), the validation loss tends to be stable at about 10 times and finally remains at about 1.65, indicating that the generalization ability of the model after training is relatively stable. Such results verify the robustness and reliability of the model in handling fault detection tasks and are suitable for practical application scenarios.

[0139] In addition, an embodiment of the present invention also discloses an electronic device, comprising: at least one processor; at least one memory for storing at least one program; when the at least one program is executed by the at least one processor, a traction system fault detection method based on PPDM and expert prompt learning as in any of the previous embodiments is implemented.

[0140] In addition, an embodiment of the present invention further discloses a computer-readable storage medium, in which computer-executable instructions are stored, and the computer-executable instructions are used to execute the traction system fault detection method based on PPDM and expert prompt learning as in any of the previous embodiments.

[0141] The system architecture and application scenarios described in the embodiments of the present invention are intended to more clearly illustrate the technical solutions of the embodiments of the present invention, and do not constitute a limitation on the technical solutions provided by the embodiments of the present invention. Those skilled in the art will appreciate that with the evolution of the system architecture and the emergence of new application scenarios, the technical solutions provided by the embodiments of the present invention are equally applicable to similar technical problems.

[0142] Those skilled in the art will appreciate that all or some of the steps in the methods disclosed above, and the functional modules / units in the systems and devices may be implemented as software, firmware, hardware, or a suitable combination thereof.

[0143] In hardware implementation, the division between the functional modules / units mentioned in the above description does not necessarily correspond to the division of physical components; for example, a physical component may have multiple functions, or a function or step may be performed by several physical components in cooperation. Some physical components or all physical components may be implemented as software executed by a processor, such as a central processing unit, a digital signal processor or a microprocessor, or implemented as hardware, or implemented as an integrated circuit, such as an application-specific integrated circuit. Such software may be distributed on a computer-readable medium, which may include a computer storage medium (or non-transitory medium) and a communication medium (or temporary medium). As known to those of ordinary skill in the art, the term computer storage medium includes volatile and non-volatile, removable and non-removable media implemented in any method or technology for storing information (such as computer-readable instructions, data structures, program modules or other data). Computer storage media include, but are not limited to, RAM, ROM, EEPROM, flash memory or other memory technology, CD-ROM, digital versatile disk (DVD) or other optical disk storage, magnetic cassettes, magnetic tapes, disk storage or other magnetic storage devices, or any other medium that can be used to store desired information and can be accessed by a computer. Furthermore, it is well known to those skilled in the art that communication media typically embodies computer readable instructions, data structures, program modules, or other data in a modulated data signal such as a carrier wave or other transport mechanism, and may include any information delivery media.

[0144] The terms "component", "module", "system", etc. used in this specification are used to represent computer-related entities, hardware, firmware, a combination of hardware and software, software, or software in execution. For example, a component can be, but is not limited to, a process running on a processor, a processor, an object, an executable file, an execution thread, a program, or a computer. By way of illustration, both applications running on a computing device and a computing device can be components. One or more components may reside in a process or an execution thread, and a component may be located on a computer or distributed between two or more computers. In addition, these components may be executed from various computer-readable media having various data structures stored thereon. Components may communicate, for example, through local or remote processes based on signals having one or more data packets (e.g., data from two components interacting with another component between a local system, a distributed system, or a network, such as the Internet interacting with other systems through signals).

Claims

1. A traction system fault detection method based on PPDM and expert prompt learning, characterized in that: The following steps are involved: S1), respectively obtain the fault-free data set U from the digital simulation platform N and noisy fault dataset U F , and the fault-free data set U N and noisy fault dataset U F Divided into training set and test set, set the fault-free data set U N is the source domain, the noisy fault dataset U F is the target domain; S2), building a data reconstruction model based on the Pre-Pyramid Deformation Module, wherein the data reconstruction model includes a pyramid deformation module and an information compression module; S3) Build a domain generalization model target network based on expert prompts The source domain dataset is clustered to obtain the initial meta-prompt P meta = {P1……P N }; S4), the fault-free data set U N Input to the pyramid deformation module, the pyramid deformation module U N Perform feature extraction, target network Based on the meta prompt P meta Train to get expert prompt P i = {P E1 ......P EN }, change U N The extracted features and expert prompts P i Reconstructed into dataset U f1 , the dataset U f1 Input into the information compression module to remove noise and redundant information. After the information compression module, the reconstruction error of the data set is calculated using the evaluation function to obtain the threshold J th ; S5) Design a network based on the cross attention mechanism and extract the fault-free dataset U N The trained expert prompt P i As "key" and "value", the fault dataset U F The features of as "query"; S6), the fault data set U F The pyramid deformation module extracts the features of the fault data set, and the cross attention network calculates the similarity between the "query" and the "value" to generate the attention weight, and performs a weighted combination of the expert prompts according to the weight, and inputs the weighted combination of the expert prompts into the target network. In the example above, generate the expert prompt P m = {P FE1 ......P FEN }, extract the features of the fault dataset and the expert prompt P m Reconstructed dataset U f2 , input to the information compression module to remove the noise of the data set, calculate the reconstruction error of the data set through the evaluation function to get the output value j, and compare it with the threshold Jth obtained in the training stage. If j>J th , then the system is judged to have a fault, if j<J th , then the system is judged to be fault-free.

2. A traction system fault detection method based on PPDM and expert prompt learning according to claim 1, characterized in that: In step S1), the fault-free data set U N and the noisy fault dataset U F Including transformer secondary side voltage output U n And current output I n , the voltage output U of the supporting capacitor on the DC link cd1 , the voltage output U of the supporting capacitor under the DC link cd2 , inverter a, b, c three-phase current output {I sa , I sb , I sc }, motor speed signal V; The normal operation data training set obtained by the digital simulation platform is defined as U N , where [U n , I n , U cd1 , U cd2 , I sa , I sb , I sc , V]∈U N ; The fault operation test data set obtained by the digital simulation platform is defined as U F , where [U n , I n , U cd1 , U cd2 , I sa , I sb , I sc , V]∈U F ; For the normal operation data training set U N Run the test dataset U with faults F Normalize between [-1, 1] to get a one-dimensional real vector: Where a is the normal operation data training set U N The number of sample data in U N (i) represents the i-th group of normal operation data training set; n represents the dimension of the signal; b is the fault running test data set U F The number of sample data in ; U F (i) represents the i-th signal of the fault running test data set, Indicates signal U F (i) The signal amplitude of the 1st, ..., nth dimension, where n represents the dimension of the signal.

3. The traction system fault detection method based on PPDM and expert prompt learning according to claim 1 is characterized in that: In step S3), the source domain data set is clustered to obtain the initial meta-prompt, which includes the following steps: S31) The source domain dataset U N , input into the pyramid deformation module to extract a set of feature vectors, which represent the main characteristics of the data; S32) clustering the extracted feature vectors to obtain subsets with multiple similar features, using a meta-learning method to find commonalities from the clustered subsets to generate meta-prompts.

4. The traction system fault detection method based on PPDM and expert prompt learning according to claim 1 is characterized in that: In step S4) and step S6), the pyramid deformation module transforms the fault-free data set U N and noisy fault dataset U F Perform feature extraction and reconstruction. The pyramid deformation module calculates the feature difference between normal data and fault data by setting a specific deformation function φ(x) and reconstructs it. The feature extraction uses the following function: φ(x)=f conv (x)+f deform (x);(1) Among them, f conv represents the convolution operation, which is used to extract local features in time series data. deform Used for data normalization, φ(x) represents the process of converting the original input X into feature representation.

5. A traction system fault detection method based on PPDM and expert prompt learning according to claim 4, characterized in that: The pyramid deformation module is used for the fault-free dataset U N and noisy fault dataset U F Perform feature extraction and reconstruction. The reconstruction process of the pyramid deformation module is expressed as: The pyramid deformation module is based on a given training set sample X = {X1...X N }, the normal data set U obtained by the digital simulation platform is learned N and noisy fault dataset U F The diversity between the data sets obtained by the digital simulation platform and the reconstructed data sets is expressed as The underestimation error in the reconstruction process is Compensate again.

6. A traction system fault detection method based on PPDM and expert prompt learning according to claim 4, characterized in that: In step S4) and step S6), the training expert prompt P i The steps include: S41) Initialize the expert prompt, for each source domain X i , first change the expert prompt P i Initialized to the pre-trained meta-prompt P meta , ensuring that the expert prompts for each source domain are tuned and optimized based on the same initial prompts: P i ←P meta ;(2) S42) Optimize the expert prompt by minimizing the KL and divergence loss L KL , for each sample in the source domain data (x i，j ,y i，j ) to perform gradient descent optimization: where α A is the learning rate, L KL Measuring the difference between the model output and the target label; S43) Each optimized expert prompt P i Normalize to keep it consistent: The normalization process is performed to ensure that the expert prompts maintain the same scale across different source domains.

7. The traction system fault detection method based on PPDM and expert prompt learning according to claim 1 is characterized in that: The method of training the expert prompt by the cross-attention mechanism method in step S5) is: The weighted average method is used to combine the prompts that match the subsets; the weights and the required number of expert prompts are assigned to the subsets of the training data set through the cross-attention mechanism. The expert prompts become the "key" and "value", and the normal operation data obtained by the digital simulation platform becomes the target attention of the "query". The training generates the expert prompt P m = {P FE1 ...P FEN }, and the same number of attention weights q1…q corresponding to the expert prompts of the training data subset is calculated. n ; Among them, the weight calculation method q n =VQK T .

8. The traction system fault detection method based on PPDM and expert prompt learning according to claim 6 is characterized in that: In step S6), the expert prompt P generated by the expert prompt module is i = {P E1 ...P EN The training data set is obtained from the digital simulation platform as the input of the fault detection framework. The training data set enters the diversity perception module to realize feature amplification, extraction and data set reconstruction. Expert prompts are assigned to the reconstructed data set. The similarity between the reconstructed data set and the training data set is controlled by expert prompts. The reconstructed training data set and the matched expert prompts are used as the input of the information compression module. The information compression module restricts and compresses the abnormal information in the reconstructed data set. The compressed abnormal information is compressed through the loss L. com It indicates that the fault detection threshold j is calculated by statistical method; The test data set obtained by the digital simulation platform is used as the input of the data reconstruction fault detection framework, and the output is the evaluation criteria J th , the evaluation criteria J th Compared with the fault detection threshold j, the fault detection result is as follows:

9. A traction system fault detection method based on PPDM and expert prompt learning according to claim 8, characterized in that: In step S2), design an information compression module: The information compression module adopts the VQ-Layer layer design method. The reconstructed data set is passed through the encoder to extract high-dimensional feature embedding and obtain the feature representation z e , by finding e recent memory items and quantize them into a single memory feature cube z q The information compression module learns the reconstructed data set output by the data reconstruction diversity perception module according to the embedded noise data information, and finds the characteristics of the noise data by calculating the L2 minimum distance; the information compression module includes two steps, search "Search" and compression "Comp", to search for noise information in the data set and limit and compress the noise information; the noise information includes a hyperparameter β and a stop gradient operator SG, and a compression loss Lcom; The two steps of searching and compressing in the information compression module are as follows: S21) The information compression module receives the input data set and identifies the noise information in the data by finding the feature vector with the minimum L2 distance to the memory unit. The calculation formula is as follows: in, is the feature embedded in the dataset, z n is the feature vector in the memory cell; S22) Calculate the compression loss through VQ-Layer. The compression loss is calculated as follows: L com =||SG(z e )-With q ||2+β||z e -SG(from q )||2; (6) Among them, SG(·) represents the stop gradient operation, Z e is the embedding representation, Z q is the quantized feature, and the hyperparameter β is used to balance the compression effect and noise removal effect.

10. The traction system fault detection method based on PPDM and expert prompt learning according to claim 8, characterized in that: In step S6), the fault detection threshold is calculated, and the specific method is as follows: The residual mean calculation formula is: where x i is the original data, is the reconstructed data, N is the number of data points; The formula for calculating the standard deviation of the residuals is: The threshold calculation formula is: j=μ res +η·s res +βL com ;(9) Among them, μ res is the residual mean, σ res is the standard deviation of the residual, η is the confidence parameter, and β is the adjustment hyperparameter.

Images