Digital twin-driven steel rail state transition diagnosis method

Abstract

The invention discloses a digital twinning driven steel rail state transition diagnosis method, which belongs to the field of rail traffic fault diagnosis, and comprises the following steps: establishing a digital twinning model of a vehicle-rail coupling system; performing simulation by using a digital twin model to obtain a plurality of groups of vibration response signals; performing time-frequency analysis on each group of vibration response signals and each group of actually collected axle box vibration signals through wavelet transform to obtain a time-frequency diagram of each group of vibration response signals and a time-frequency diagram of each group of axle box vibration signals; taking the time-frequency diagram of each group of vibration response signals as source domain data of a sub-domain alignment migration network, and taking the time-frequency diagram of each group of axle box vibration signals as target domain data of the sub-domain alignment migration network to obtain a trained sub-domain alignment migration network; and utilizing the trained sub-domain alignment migration network to predict a to-be-detected axle box vibration signal, and completing diagnosis and identification of different steel rail states. The problem that the steel rail state diagnosis precision is insufficient under the condition that sample labels are scarce is solved.

Description

Technical Field

[0001] The present invention belongs to the field of rail transit fault diagnosis, and in particular relates to a rail state migration diagnosis method driven by a digital twin. Background Art

[0002] At present, the methods for diagnosing the status of railway tracks in my country are mainly divided into static diagnosis and dynamic diagnosis. The static detection method of tracks refers to the detection of the geometric shape and position of the track and turnout without wheel load. It is mainly composed of static detection trolleys, high-precision total stations, data processing software, etc., and can detect the elevation, track gauge, superelevation and smoothness of the track. Although this type of diagnostic method can accurately diagnose the status of rails, they must be carried out during the rail transit window period and consume a lot of human resources.

[0003] Dynamic detection refers to the installation of detection equipment on the train to reflect the track status and track geometry through the smoothness of driving. At present, dedicated track inspection vehicles are the most widely used dynamic detection method, which is to install detection equipment on the track inspection vehicle train or bogie to obtain the health status of the track structure. However, the research and development, repair and maintenance of dedicated track inspection vehicles require a very high amount of expenses; more importantly, when using dedicated track inspection vehicles to diagnose track status, it is necessary to formulate a special scheduling table for it, which is very difficult for railway departments with high operating density; and the track status changes with the load of the train and other different working conditions. Dedicated track inspection vehicles cannot complete the fitness assessment of different trains. Therefore, it is difficult to achieve real-time track status diagnosis.

[0004] The train vibration response data collected by sensors installed on operating vehicles can reflect the status of the train during operation. The health of the track structure can be effectively reflected based on the train operation status. The monitoring system based on the train vibration response characteristics has the advantages of simple deployment and low cost. Therefore, the method of rapid identification and positioning of high-speed railway track structures and real-time monitoring based on the vibration response information of operating vehicles can provide a new idea for track structure health detection. However, traditional digital signal analysis technology has problems such as the analysis cannot be adaptive, the time-frequency resolution cannot achieve the optimal results at the same time, and modal aliasing, which have a great impact on the results.

[0005] With the development of deep learning technology, there are also some studies focusing on the diagnosis of key components of the vehicle-track system using neural networks. However, these deep learning-based methods require manual collection of vibration signals that can reflect the track status and labeling, which is often impractical, otherwise it will consume a lot of human resources. In addition, the collected vibration signals may be of low quality, and the fault information may be submerged, resulting in negative effects in the process of training the network.

[0006] With the rapid development of intelligent manufacturing technology, digital twins have emerged, providing a new way to obtain the physical world by connecting the physical world and the virtual world. Collecting simulated working condition data through twin models and enriching label scarce data sets are effective ways to improve the reliability of fault diagnosis. Although digital twin technology has developed rapidly, the data simulated by the twin model cannot completely maintain the same characteristics as the actual data. How to overcome this difference and promote reliable diagnostic models from twin models to physical models is a key issue to ensure diagnostic accuracy.

[0007] The prior art discloses a GAN-based interpolation method for missing data of wind measurement in a fan hub. The method applies a generative adversarial network to the field of fault diagnosis of a fan, inputs training samples into a generative model and a discriminative model for iterative training until a Nash equilibrium is reached, and completes the data supplementation work through the trained model. However, this method only uses a generative adversarial network to generate virtual data and add it to a data set, and only considers the correlation between real data and generated data, but does not consider the fault classification problem, and also does not design a fault diagnosis model that matches it. Summary of the invention

[0008] In view of the above-mentioned deficiencies in the prior art, the present invention provides a digital twin-driven rail state migration diagnosis method that solves the problem of insufficient rail state diagnosis accuracy under the condition of scarce sample labels.

[0009] In order to achieve the above-mentioned invention object, the technical solution adopted by the present invention is: a rail state migration diagnosis method driven by digital twin, comprising the following steps:

[0010] S1. Establish a digital twin model of the vehicle-track coupling system through multi-body dynamics;

[0011] S2. Using the digital twin model of the vehicle-rail coupling system, the vibration response signal of the axle box is simulated when the vehicle passes through sections with different rail states, and several groups of vibration response signals are obtained;

[0012] S3, collecting axle box vibration signals of axle boxes when a vehicle passes through sections with different rail states, obtaining several groups of axle box vibration signals that are actually collected, and performing time-frequency analysis on each group of vibration response signals and each group of axle box vibration signals by wavelet transform, respectively, to obtain time-frequency diagrams of each group of vibration response signals and each group of axle box vibration signals;

[0013] S4, using the time-frequency diagram of each group of vibration response signals as the source domain data of the subdomain alignment transfer network, and using the time-frequency diagram of each group of axle box vibration signals as the target domain data of the subdomain alignment transfer network, training the subdomain alignment transfer network, and obtaining a trained subdomain alignment transfer network;

[0014] S5. Obtain the time-frequency diagram of the axle box vibration signal to be tested, and input the trained subdomain alignment transfer network to obtain the true state label of the axle box vibration signal to be tested, and complete the diagnosis and identification of different rail states.

[0015] Furthermore, before performing time-frequency analysis on each group of axle box vibration signals by wavelet transform in step S3, each group of axle box vibration signals is firstly low-pass filtered.

[0016] Furthermore, the subdomain alignment migration network in step S4 includes a feature extractor, a state diagnostic device and N domain discriminators; the feature extractor includes a first convolutional layer, a first pooling layer, a second convolutional layer, a second pooling layer, a third convolutional layer, a flattening layer and a first fully connected layer connected in sequence; a batch normalization operation is performed after each pooling layer;

[0017] The state diagnostic device comprises a second fully connected layer, a third fully connected layer and a fourth fully connected layer connected in sequence;

[0018] Each of the domain discriminators includes a gradient reversal layer, a fifth fully connected layer, a sixth fully connected layer and a seventh fully connected layer connected in sequence;

[0019] The fourth fully connected layer and the seventh fully connected layer in the subdomain alignment migration network both use the SoftMax activation function, and the remaining network layers all use the ReLU activation function.

[0020] Furthermore, the step S4 is specifically as follows:

[0021] S401, using the time-frequency diagram of each group of vibration response signals as the source domain data of the subdomain alignment migration network, and using the time-frequency diagram of each group of axle box vibration signals as the target domain data of the subdomain alignment migration network;

[0022] S402, using a subdomain alignment transfer network to extract features of source domain data and features of target domain data respectively;

[0023] S403, according to the characteristics of the source domain data and the characteristics of the target domain data, the classification probability of the source domain data, the pseudo label of the target domain data and the domain classification probability of all data are obtained by using the subdomain alignment migration network;

[0024] S404, aligning the migration network to update network parameters according to the classification probability of the source domain data, the pseudo-label of the target domain data, and the domain classification probability of all data;

[0025] S405, judging whether the network parameters of the subdomain alignment transfer network have reached convergence, if so, obtaining a trained subdomain alignment transfer network, otherwise, returning to step S401.

[0026] Furthermore, the expression for updating the network parameters of the subdomain alignment migration network in step S404 is:

[0027]

[0028] in, is the network parameter of the updated feature extractor; The network parameters of the updated status diagnostic device; To make Minimize the corresponding θ F and θ C ; is the loss function of the subdomain alignment transfer network; θ F is the network parameter of the feature extractor; θ C is the network parameter of the state diagnostics; θ D,k is the network parameter of the k-th domain discriminator; is a set of network parameters of each domain discriminator; k is the domain discriminator number; |C s | is the number of domain discriminators; is the updated set of network parameters of the discriminator in each field; To make Maximize the corresponding θ F and L c is the classification loss of the source domain data; λ is the importance weight between the domain alignment target and the classification target; L d is the domain discrimination loss for all data; L w is the Wasserstein loss between source domain data and target domain data of the same category; L v It is the intra-class variance loss of the target domain data of the same category.

[0029] Furthermore, the classification loss of the source domain data is expressed as:

[0030]

[0031] Among them, n s is the number of source domain samples in the source domain data; i is the source domain sample number in the source domain data; l c () is the cross entropy loss; is the classification probability of each category of the i-th source domain sample; is the feature extracted by the i-th source domain sample after passing through the feature extractor; is the i-th source domain sample; is the category of the i-th source domain sample; is the classification probability set of the i-th source domain sample; It is a judgment function to determine whether the category represented by the k-th domain discriminator is the correct classification of the i-th source domain sample. If so, it is 1, otherwise, it is 0; log() is the logarithmic sign; is the classification probability that the i-th source domain sample belongs to the category represented by the k-th domain discriminator.

[0032] Furthermore, the expression of the domain discrimination loss of all the data is:

[0033]

[0034] Among them, n k is the total number of samples belonging to the category represented by the kth domain discriminator in the source domain data and the target domain data; r is the sample number; l c () is the cross entropy loss; F(x r ) is the rth sample x r Features extracted after passing through the feature extractor; D k (F(x r )) represents the feature F(x r ) Input the kth domain discriminator; d r For sample x r The domain label of x. If it is 0, then x r From the source domain, otherwise, x r From the target domain; I[y r = k] is used to determine the rth sample x r Whether it belongs to the category represented by the k-th domain discriminator, if so, it is 1, otherwise, it is 0; y r is the category label of the rth sample; is the rth sample x r The domain classification probability.

[0035] Furthermore, the expression of the Wasserstein loss between the source domain data and the target domain data of the same category is:

[0036]

[0037] Among them, n s is the number of source domain samples in the source domain data; i is the source domain sample number in the source domain data; y is the domain discriminator number corresponding to the category label; n t is the number of target domain samples in the target domain data; j is the target domain sample number in the target domain data; P s,y＝k is the probability distribution of the source domain sample in the source domain data belonging to the category represented by the kth domain discriminator; P t,y＝k is the probability distribution of the target domain sample in the target domain data belonging to the category represented by the kth domain discriminator; P s,y≠kis the probability distribution of the source domain sample in the source domain data not belonging to the category represented by the kth domain discriminator; Π(P s,y＝k ,P t,y＝k ) is P s,y＝k and P t,y＝k The joint probability distribution of s,y≠k ,P t,y＝k ) is P s,y≠k and P t,y＝k The joint probability distribution of When the i-th source domain sample When it belongs to the category represented by the k-th domain discriminator, Extracted features; For When it does not belong to the category represented by the k-th domain discriminator, Extracted features; is the jth target domain sample When it belongs to the category represented by the k-th domain discriminator, The features extracted from s,y＝k Appear in And the distribution P t,y＝k Appear in The joint probability distribution of s,y≠k Appear in And the distribution P t,y＝k Appear in The joint probability distribution of For the joint probability distribution Π(P s,y＝k ,P t,y＝k ) and The expectation of the distance, there exists γ1 that minimizes this expectation; for The infimum of For the joint probability distribution Π(P s,y≠k ,P t,y＝k ) and The expectation of the distance, there exists γ2 that minimizes this expectation; for The infimum of ; || || is the norm symbol.

[0038] Furthermore, the expression of the intra-class variance loss of the target domain data of the same category is:

[0039]

[0040] Among them, n tis the number of target domain samples in the target domain data; j is the target domain sample number in the target domain data; y is the domain discriminator number corresponding to the category label; is the jth target domain sample When it belongs to the category represented by the k-th domain discriminator, The features extracted from is the average value of all samples belonging to the category represented by the kth domain discriminator.

[0041] The beneficial effects of the present invention are as follows: the present invention uses the dynamic parameters of the vehicle-rail coupling system in the physical world to establish a digital twin model, and uses the model to simulate the axle box vibration response signal when the vehicle travels on rails in different states, thereby expanding the axle box signal data set with scarce labels; the multi-layer convolutional layer in the feature extractor is used to extract features from samples, adaptively learn fault features layer by layer, and realize feature learning of different rail states; an independent domain discriminator is equipped for each category of samples to ensure that the extracted source domain and target domain features are only aligned according to their own categories; a Wasserstein loss function is designed to shorten the The source domain and target domain data features in the same category are combined while maximizing the distance between data features in different categories; the intra-class variance loss function is used to shorten the distance between target domain data features in the same category to improve the accuracy of rail state diagnosis; the time domain characteristics and frequency domain characteristics of the signal are considered at the same time, and the wavelet time-frequency diagrams of the simulation signal and the collected signal are used as samples to improve the effect of network feature learning; the batch processing method used in the present invention can solve the problem of poor initialization effect, help the gradient propagate to each layer of the network, accelerate the convergence of the model, and effectively alleviate the problem of model overfitting, avoid the collapse of the generated model, and prevent the generated samples from converging to the same point. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 The present invention is a flow chart of the method.

[0043] Figure 2 It is a schematic diagram of the structure of the subdomain alignment migration network of the present invention;

[0044] Figure 3 It is a time domain and frequency domain comparison diagram of the simulation signal and the collected signal under the rail corrugation state in Example 2 of the present invention;

[0045] Figure 4 It is a time domain and frequency domain comparison diagram of the simulation signal and the collected signal in the rail-to-rail weld state in Example 2 of the present invention;

[0046] Figure 5 A confusion matrix diagram of the rail state diagnosis result in Example 2 of the present invention;

[0047] Figure 6This is a visualization diagram of the features extracted in Example 2 of the present invention;

[0048] Figure 7 This is a schematic diagram showing how the diagnostic effect varies with the number of source domain samples in Example 2 of the present invention. DETAILED DESCRIPTION

[0049] The specific implementation modes of the present invention are described below so that those skilled in the art can understand the present invention. However, it should be clear that the present invention is not limited to the scope of the specific implementation modes. For those of ordinary skill in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the attached claims, these changes are obvious, and all inventions and creations utilizing the concept of the present invention are protected.

[0050] Example 1

[0051] like Figure 1 As shown, in one embodiment of the present invention, a rail state migration diagnosis method driven by a digital twin comprises the following steps:

[0052] S1. Establish a digital twin model of the vehicle-track coupling system through multi-body dynamics;

[0053] S2. Using the digital twin model of the vehicle-rail coupling system, the vibration response signal of the axle box is simulated when the vehicle passes through sections with different rail states, and several groups of vibration response signals are obtained;

[0054] S3, collecting axle box vibration signals of axle boxes when a vehicle passes through sections with different rail states, obtaining several groups of axle box vibration signals that are actually collected, and performing time-frequency analysis on each group of vibration response signals and each group of axle box vibration signals by wavelet transform, respectively, to obtain time-frequency diagrams of each group of vibration response signals and each group of axle box vibration signals;

[0055] S4, using the time-frequency diagram of each group of vibration response signals as the source domain data of the subdomain alignment transfer network, and using the time-frequency diagram of each group of axle box vibration signals as the target domain data of the subdomain alignment transfer network, training the subdomain alignment transfer network, and obtaining a trained subdomain alignment transfer network;

[0056] S5. Obtain the time-frequency diagram of the axle box vibration signal to be tested, and input the trained subdomain alignment transfer network to obtain the true state label of the axle box vibration signal to be tested, and complete the diagnosis and identification of different rail states.

[0057] Before performing time-frequency analysis on each group of axle box vibration signals by wavelet transform in step S3, each group of axle box vibration signals is firstly low-pass filtered.

[0058] In this embodiment, the digital twin model includes a vehicle subsystem and a track subsystem, wherein the vehicle subsystem is composed of a vehicle body, a frame, a wheelset, and an axle box, and the track subsystem is composed of rails and a roadbed. The mathematical models of rail corrugation and rail welds are added to the track irregularity spectrum to simulate the axle box vibration response signal when the vehicle travels on rails in different states.

[0059] The signal simulated by the digital twin model does not contain fault information, so the real collected axle box vibration signal is low-pass filtered to remove some noise.

[0060] The subdomain alignment migration network in step S4 includes a feature extractor, a state diagnostic device and N domain discriminators;

[0061] The feature extractor is used to extract features of source domain data and target domain data;

[0062] The state diagnostic device is used to obtain the classification probability of the source domain data according to the characteristics of the source domain data, and to obtain the pseudo label of the target domain data according to the characteristics of the target domain data;

[0063] Each of the domain discriminators is used to obtain respective domain classification probabilities according to the characteristics of the source domain data and the target domain data;

[0064] Where N is the number of rail states.

[0065] The subdomain alignment migration network in step S4 includes a feature extractor, a state diagnostic device and N domain discriminators; the feature extractor includes a first convolutional layer, a first pooling layer, a second convolutional layer, a second pooling layer, a third convolutional layer, a flattening layer and a first fully connected layer connected in sequence; a batch normalization operation is performed after each pooling layer;

[0066] The state diagnostic device comprises a second fully connected layer, a third fully connected layer and a fourth fully connected layer connected in sequence;

[0067] Each of the domain discriminators includes a gradient reversal layer, a fifth fully connected layer, a sixth fully connected layer and a seventh fully connected layer connected in sequence;

[0068] The fourth fully connected layer and the seventh fully connected layer in the subdomain alignment migration network both use the SoftMax activation function, and the remaining network layers all use the ReLU activation function.

[0069] In this embodiment, during the training process, the subdomain alignment migration network includes a feature extractor, a state diagnostic device and N domain discriminators, such as Figure 2As shown in the figure. In the feature extractor, the size of the input signal wavelet frequency graph is 28×28×3, the size of the convolution kernel of each convolution layer is set to 5×5, and the number of convolution kernels is set to 16, 16, and 32 respectively. There is a pooling layer after the first two convolution layers, and the size of the convolution kernel in the pooling layer is set to 3×3; in the state diagnostics, the number of neurons in each fully connected layer is 64, 32, and 3 respectively; in each domain discriminator, the number of neurons in each fully connected layer is 64, 32, and 2 respectively.

[0070] During the training process, the wavelet time-frequency diagram of the simulated signal is used as the source domain data, and the wavelet time-frequency diagram of the real collected signal is used as the target domain data. The source domain data features and the target domain data features are extracted by the feature extractor. The source domain data features are input into the state diagnostic device to obtain the classification probability of the source domain data; the target domain data features are input into the state diagnostic device to obtain the pseudo label corresponding to the target domain data, thereby obtaining the category of the target domain sample.

[0071] In the structure of the subdomain alignment migration network, the output layers of the state diagnostics and domain discriminators use the SoftMax activation function, and the rest use the ReLU activation function to solve the problem of gradient disappearance and speed up the convergence.

[0072] The step S4 is specifically as follows:

[0073] S401, using the time-frequency diagram of each group of vibration response signals as the source domain data of the subdomain alignment migration network, and using the time-frequency diagram of each group of axle box vibration signals as the target domain data of the subdomain alignment migration network;

[0074] S402, using a subdomain alignment transfer network to extract features of source domain data and features of target domain data respectively;

[0075] S403, according to the characteristics of the source domain data and the characteristics of the target domain data, the classification probability of the source domain data, the pseudo label of the target domain data and the domain classification probability of all data are obtained by using the subdomain alignment migration network;

[0076] S404, aligning the migration network to update network parameters according to the classification probability of the source domain data, the pseudo-label of the target domain data, and the domain classification probability of all data;

[0077] S405, judging whether the network parameters of the subdomain alignment transfer network have reached convergence, if so, obtaining a trained subdomain alignment transfer network, otherwise, returning to step S401.

[0078] The expression for updating the network parameters of the subdomain alignment migration network in step S404 is:

[0079]

[0080] in, is the network parameter of the updated feature extractor; The network parameters of the updated status diagnostic device; To make Minimize the corresponding θ F and θ C ; is the loss function of the subdomain alignment transfer network; θ F is the network parameter of the feature extractor; θ C is the network parameter of the state diagnostics; θ D,k is the network parameter of the k-th domain discriminator; is a set of network parameters of each domain discriminator; k is the domain discriminator number; |C s | is the number of domain discriminators; is the updated set of network parameters of the discriminator in each field; To make Maximize the corresponding θ F and L c is the classification loss of the source domain data; λ is the importance weight between the domain alignment target and the classification target; L d is the domain discrimination loss for all data; L w is the Wasserstein loss between source domain data and target domain data of the same category; L v It is the intra-class variance loss of the target domain data of the same category.

[0081] The expression of the classification loss of the source domain data is:

[0082]

[0083] Among them, n s is the number of source domain samples in the source domain data; i is the source domain sample number in the source domain data; l c () is the cross entropy loss; is the classification probability of each category of the i-th source domain sample; is the feature extracted by the i-th source domain sample after passing through the feature extractor; is the i-th source domain sample; is the category of the i-th source domain sample; is the classification probability set of the i-th source domain sample; It is a judgment function to determine whether the category represented by the k-th domain discriminator is the correct classification of the i-th source domain sample. If so, it is 1, otherwise, it is 0; log() is the logarithmic sign; is the classification probability that the i-th source domain sample belongs to the category represented by the k-th domain discriminator.

[0084] The expression of the domain discrimination loss of all the data is:

[0085]

[0086] Among them, n k is the total number of samples belonging to the category represented by the kth domain discriminator in the source domain data and the target domain data; r is the sample number; l c () is the cross entropy loss; F(x r ) is the rth sample x r Features extracted after passing through the feature extractor; D k (F(x r )) represents the feature F(x r ) Input the kth domain discriminator; d r For sample x r The domain label of x. If it is 0, then x r From the source domain, otherwise, x r From the target domain; I[y r = k] is used to determine the rth sample x r Whether it belongs to the category represented by the k-th domain discriminator, if so, it is 1, otherwise, it is 0; y r is the category label of the rth sample; is the rth sample x r The domain classification probability.

[0087] The expression of Wasserstein loss between the source domain data and the target domain data of the same category is:

[0088]

[0089] Among them, n s is the number of source domain samples in the source domain data; i is the source domain sample number in the source domain data; y is the domain discriminator number corresponding to the category label; n t is the number of target domain samples in the target domain data; j is the target domain sample number in the target domain data; P s,y＝k is the probability distribution of the source domain sample in the source domain data belonging to the category represented by the kth domain discriminator; P t,y＝k is the probability distribution of the target domain sample in the target domain data belonging to the category represented by the kth domain discriminator; P s,y≠k is the probability distribution of the source domain sample in the source domain data not belonging to the category represented by the kth domain discriminator; Π(P s,y＝k ,P t,y＝k ) is P s,y＝k and P t,y＝k The joint probability distribution of s,y≠k ,Pt,y＝k ) is P s,y≠k and P t,y＝k The joint probability distribution of When the i-th source domain sample When it belongs to the category represented by the k-th domain discriminator, Extracted features; For When it does not belong to the category represented by the k-th domain discriminator, Extracted features; is the jth target domain sample When it belongs to the category represented by the k-th domain discriminator, The features extracted from s,y＝k Appear in And the distribution P t,y＝k Appear in The joint probability distribution of s,y≠k Appear in And the distribution P t,y＝k Appear in The joint probability distribution of For the joint probability distribution Π(P s,y＝k ,P t,y＝k ) and The expectation of the distance, there exists γ1 that minimizes this expectation; for The infimum of For the joint probability distribution Π(P s,y≠k ,P t,y＝k ) and The expectation of the distance, there exists γ2 that minimizes this expectation; for The infimum of ; |||| is the norm symbol.

[0090] The expression of intra-class variance loss of the target domain data of the same category is:

[0091]

[0092] Among them, n t is the number of target domain samples in the target domain data; j is the target domain sample number in the target domain data; y is the domain discriminator number corresponding to the category label; is the jth target domain sample When it belongs to the category represented by the k-th domain discriminator, The features extracted from is the average value of all samples belonging to the category represented by the kth domain discriminator.

[0093] In this embodiment, during the network training process, the category labels of the target domain data are actually pseudo labels, not real categories.

[0094] In this embodiment, the classification loss is obtained by calculating the cross entropy loss between the source domain sample label and the source domain data classification probability; the discrimination loss is obtained by calculating the cross entropy loss between the domain labels of the source domain samples and the target domain samples in each domain discriminator and their domain classification probabilities; the Wasserstein loss is obtained by calculating the Wasserstein distance between the source domain and target domain data features of the same category, where the category of the target domain sample is determined by the pseudo label obtained by inputting the sample into the state diagnostic device; the intra-class variance loss is obtained by calculating the distance between the target domain data features of the same category and the centroid positions of these features.

[0095] The subdomain alignment migration network uses the Adam method to update parameters.

[0096] The training size of each batch in the subdomain alignment transfer network is 48 samples.

[0097] In this embodiment, the present invention uses the dynamic parameters of the vehicle-rail coupling system in the physical world to establish a digital twin model, and uses the model to simulate the axle box vibration response signal when the vehicle travels on rails in different states, expands the axle box signal data set with scarce labels, and finally realizes the diagnosis task for the rail state. The present invention uses the multi-layer convolution layer in the feature extractor to extract features from samples, adaptively learns fault features layer by layer, realizes feature learning of different rail states, and finally realizes the diagnosis task of the rail state. The present invention is equipped with an independent domain discriminator for each category of samples to ensure that the extracted source domain and target domain features will only be aligned according to their own categories. The present invention designs a Wasserstein loss function to shorten the source domain and target domain data features in the same category while maximizing the distance between data features of different categories. The present invention uses the intra-class variance loss function to shorten the distance between target domain data features in the same category to improve the accuracy of rail state diagnosis. The present invention simultaneously considers the time domain characteristics and frequency domain characteristics of the signal, and uses the wavelet time-frequency diagram of the simulation signal and the collected signal as samples to improve the effect of network feature learning. The present invention can solve the problem of poor initialization effect, help the gradient propagate to each layer of the network, accelerate the convergence of the model, and effectively alleviate the problem of model overfitting, avoid the collapse of the generated model, and prevent the generated samples from converging to the same point. The present invention uses the Adam method to optimize the subdomain alignment migration network. The method uses the first-order moment estimation and second-order moment estimation of the gradient to dynamically adjust the learning rate of each parameter. In the back propagation stage, the weights are updated to obtain the global optimal solution so that the loss function reaches the minimum value.

[0098] Example 2

[0099] The vibration signals of the axle box are collected from Xi'an Metro Line 5 when the vehicle passes through the rail corrugation or rail weld section. The sampling frequency is set to 25kHz in the experiment.

[0100] like Figure 3 As shown, when the vehicle passes through a section with corrugated rails, the axle box vibration signal simulated by the digital twin model is highly similar to the real axle box vibration signal collected in both time domain and frequency domain, indicating that the digital twin model can accurately simulate the vibration state of the axle box when corrugated rails occur.

[0101] like Figure 4 As shown, when the vehicle passes through the rail weld section, the axle box vibration signal simulated by the digital twin model is highly similar to the real axle box vibration signal collected in both time domain and frequency domain, indicating that the digital twin model can accurately simulate the vibration state of the axle box when welds appear on the rails.

[0102] like Figure 5 As shown, the subdomain alignment transfer network can accurately diagnose the rail condition, and the accuracy rate is above 90% in the diagnosis tasks of various rail conditions, which confirms the effectiveness of the method.

[0103] like Figure 6 As shown in the figure, the t-SNE technology is used to visualize the source domain data features and the target domain data features. The features belonging to the same rail state are clustered together, and the features belonging to different rail states have good distinction between each other, indicating that the extracted features are similar between the source domain and the target domain, and are conducive to the diagnosis task of the rail state.

[0104] like Figure 7 As shown in the figure, the accuracy of rail condition diagnosis gradually increases and tends to be stable by gradually increasing the number of source domain samples used for training. When the number of source domain samples is small, the rail condition diagnosis effect is still considerable, which shows that the proposed method can achieve better condition diagnosis effect in the absence of training samples, thereby greatly reducing the time required for signal simulation.

Claims

1. A rail state migration diagnosis method driven by digital twin, characterized in that: The following steps are involved: S1. Establish a digital twin model of the vehicle-track coupling system through multi-body dynamics; S2. Using the digital twin model of the vehicle-rail coupling system to simulate the vibration response signal of the axle box when the vehicle passes through sections with different rail states, several groups of vibration response signals are obtained; S3, collecting axle box vibration signals of axle boxes when a vehicle passes through sections with different rail states, obtaining several groups of axle box vibration signals that are actually collected, and performing time-frequency analysis on each group of vibration response signals and each group of axle box vibration signals by wavelet transform, respectively, to obtain time-frequency diagrams of each group of vibration response signals and each group of axle box vibration signals; S4, using the time-frequency diagram of each group of vibration response signals as the source domain data of the subdomain alignment transfer network, and using the time-frequency diagram of each group of axle box vibration signals as the target domain data of the subdomain alignment transfer network, training the subdomain alignment transfer network, and obtaining a trained subdomain alignment transfer network; S5. Obtain the time-frequency diagram of the axle box vibration signal to be tested, and input the trained subdomain alignment transfer network to obtain the true state label of the axle box vibration signal to be tested, and complete the diagnosis and identification of different rail states.

2. The rail state migration diagnosis method driven by digital twin according to claim 1 is characterized in that: Before performing time-frequency analysis on each group of axle box vibration signals by wavelet transform in step S3, each group of axle box vibration signals is firstly low-pass filtered.

3. The rail state migration diagnosis method driven by digital twin according to claim 1 is characterized in that: The subdomain alignment migration network in step S4 includes a feature extractor, a state diagnostic device and N domain discriminators; the feature extractor includes a first convolutional layer, a first pooling layer, a second convolutional layer, a second pooling layer, a third convolutional layer, a flattening layer and a first fully connected layer connected in sequence; a batch normalization operation is performed after each pooling layer; The state diagnostic device comprises a second fully connected layer, a third fully connected layer and a fourth fully connected layer connected in sequence; Each of the domain discriminators includes a gradient reversal layer, a fifth fully connected layer, a sixth fully connected layer and a seventh fully connected layer connected in sequence; The fourth fully connected layer and the seventh fully connected layer in the subdomain alignment migration network both use the SoftMax activation function, and the remaining network layers all use the ReLU activation function.

4. The rail state migration diagnosis method driven by digital twin according to claim 1, characterized in that: The step S4 is specifically as follows: S401, using the time-frequency diagram of each group of vibration response signals as the source domain data of the subdomain alignment migration network, and using the time-frequency diagram of each group of axle box vibration signals as the target domain data of the subdomain alignment migration network; S402, using a subdomain alignment transfer network to extract features of source domain data and features of target domain data respectively; S403, according to the characteristics of the source domain data and the characteristics of the target domain data, the classification probability of the source domain data, the pseudo label of the target domain data and the domain classification probability of all data are obtained by using the subdomain alignment migration network; S404, aligning the migration network to update network parameters according to the classification probability of the source domain data, the pseudo-label of the target domain data, and the domain classification probability of all data; S405, judging whether the network parameters of the subdomain alignment transfer network have reached convergence, if so, obtaining a trained subdomain alignment transfer network, otherwise, returning to step S401.

5. The rail state migration diagnosis method driven by digital twin according to claim 4 is characterized in that: The expression for updating the network parameters of the subdomain alignment migration network in step S404 is: in, is the network parameter of the updated feature extractor; The network parameters of the updated status diagnostic device; To make Minimize the corresponding θ F and θ C ; is the loss function of the subdomain alignment transfer network; θ F is the network parameter of the feature extractor; θ C is the network parameter of the state diagnostics; θ D,k is the network parameter of the k-th domain discriminator; is a set of network parameters of each domain discriminator; k is the domain discriminator number; |C s | is the number of domain discriminators; is the updated set of network parameters of the discriminator in each field; To make Maximize the corresponding θ F and L c is the classification loss of the source domain data; λ is the importance weight between the domain alignment target and the classification target; L d is the domain discrimination loss for all data; L w is the Wasserstein loss between source domain data and target domain data of the same category; L v It is the intra-class variance loss of the target domain data of the same category.

6. The rail state migration diagnosis method driven by digital twin according to claim 5, characterized in that: The expression of the classification loss of the source domain data is: Among them, n s is the number of source domain samples in the source domain data; i is the source domain sample number in the source domain data; l c ( ) is the cross entropy loss; is the classification probability of each category of the i-th source domain sample; is the feature extracted by the i-th source domain sample after passing through the feature extractor; is the i-th source domain sample; is the category of the i-th source domain sample; is the classification probability set of the i-th source domain sample; It is a judgment function to determine whether the category represented by the k-th domain discriminator is the correct classification of the i-th source domain sample. If so, it is 1, otherwise, it is 0; log() is the logarithmic sign; is the classification probability that the i-th source domain sample belongs to the category represented by the k-th domain discriminator.

7. The rail state migration diagnosis method driven by digital twin according to claim 5, characterized in that: The expression of the domain discrimination loss of all the data is: Among them, n k is the total number of samples belonging to the category represented by the kth domain discriminator in the source domain data and the target domain data; r is the sample number; l c () is the cross entropy loss; F(x r ) is the rth sample x r Features extracted after passing through the feature extractor; D k (F(x r )) represents the feature F(x r ) Input the kth domain discriminator; d r For sample x r The domain label of x. If it is 0, then x r From the source domain, otherwise, x r From the target domain; I[y r = k] is used to determine the rth sample x r Whether it belongs to the category represented by the k-th domain discriminator, if so, it is 1, otherwise, it is 0; y r is the category label of the rth sample; is the rth sample x r The domain classification probability.

8. The rail state migration diagnosis method driven by digital twin according to claim 5, characterized in that: The expression of Wasserstein loss between the source domain data and the target domain data of the same category is: Among them, n s is the number of source domain samples in the source domain data; i is the source domain sample number in the source domain data; y is the domain discriminator number corresponding to the category label; n t is the number of target domain samples in the target domain data; j is the target domain sample number in the target domain data; P s,y＝k is the probability distribution of the source domain sample in the source domain data belonging to the category represented by the kth domain discriminator; P t,y＝k is the probability distribution of the target domain sample in the target domain data belonging to the category represented by the kth domain discriminator; P s,y≠k is the probability distribution of the source domain sample in the source domain data not belonging to the category represented by the kth domain discriminator; Π(P s,y＝k ,P t,y＝k ) is P s,y＝k and P t,y＝k The joint probability distribution of s,y≠k ,P t,y＝k ) is P s,y≠k and P t,y＝k The joint probability distribution of When the i-th source domain sample When it belongs to the category represented by the k-th domain discriminator, Extracted features; For When it does not belong to the category represented by the k-th domain discriminator, Extracted features; is the jth target domain sample When it belongs to the category represented by the k-th domain discriminator, The features extracted from s,y＝k Appear in And the distribution P t,y＝k Appear in The joint probability distribution of s,y≠k Appear in And the distribution P t,y＝k Appear in The joint probability distribution of For the joint probability distribution Π(P s,y＝k ,P t,y＝k ) and The expectation of the distance, there exists γ1 that minimizes this expectation; for The infimum of For the joint probability distribution Π(P s,y≠k ,P t,y＝k ) and The expectation of the distance, there exists γ2 that minimizes this expectation; for The infimum of ; || || is the norm symbol.

9. The rail state migration diagnosis method driven by digital twin according to claim 5, characterized in that: The expression of intra-class variance loss of the target domain data of the same category is: Among them, n t is the number of target domain samples in the target domain data; j is the target domain sample number in the target domain data; y is the domain discriminator number corresponding to the category label; is the jth target domain sample When it belongs to the category represented by the k-th domain discriminator, The features extracted from is the average value of all samples belonging to the category represented by the kth domain discriminator.

Images