A digital twin driven rail state migration diagnosis method

Abstract

The application discloses a kind of digital twin driven steel rail state migration diagnostic method, belong to rail transit fault diagnosis field, this method includes establishing the digital twin model of vehicle-track coupling system;Simulation is carried out using digital twin model, obtain several groups of vibration response signals;The time-frequency analysis of each group of vibration response signals and each group of real collected axlebox vibration signals is carried out by wavelet transform, obtain the time-frequency diagram of each group of vibration response signals and the time-frequency diagram of each group of axlebox vibration signals;The time-frequency diagram of each group of vibration response signals is regarded as the source domain data of sub-domain alignment migration network, and the time-frequency diagram of each group of axlebox vibration signals is regarded as the target domain data of sub-domain alignment migration network, to obtain the sub-domain alignment migration network of training completion;Predict using the sub-domain alignment migration network of training completion to measured axlebox vibration signal, complete the diagnostic identification of different steel rail state.The application solves the problem of insufficient steel rail state diagnosis precision under the condition of sample label scarcity.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of rail transit fault diagnosis, and particularly relates to a rail state migration diagnosis method driven by digital twinning. BACKGROUND

[0002] At present, the methods for diagnosing the state of railway tracks in China mainly include static diagnosis and dynamic diagnosis. The static track detection method refers to the detection of the geometric shape of the track and turnout under the action of no wheel load, mainly composed of a static detection trolley, a high-precision total station and software for processing data, and can detect the elevation, alignment, gauge, super-elevation and smoothness of the track. Although this type of diagnosis method can accurately diagnose the state of the rail, it must be carried out during the off-peak period of rail transit and consumes a large amount of human resources.

[0003] Dynamic detection refers to installing a detection device on a train to reflect the track state and track geometric shape through driving stability. At present, the special track inspection vehicle is the most widely used dynamic detection method, which installs detection equipment on the track inspection vehicle train or bogie to obtain the health status of the track structure. However, the development, maintenance and repair of the special track inspection vehicle require a high cost. More importantly, it is very difficult to develop a special schedule for the special track inspection vehicle for the railway department with high operation density. In addition, the track state changes with the load of the train and other different working conditions, and the special track inspection vehicle cannot complete the adaptability evaluation of different trains, so it is difficult to realize real-time track state diagnosis.

[0004] The train vibration response data collected by arranging sensors on the operating vehicle can reflect the state during the train operation, and the health of the track structure can be effectively reflected according to the train operation state. Moreover, the monitoring system based on the train vibration response characteristics has the advantages of simple arrangement and low cost. Therefore, the method for quickly identifying and positioning the high-speed railway track structure, and real-time monitoring based on the vibration response information of the operating vehicle can provide a new idea for track structure health detection. However, the traditional digital signal analysis technology has problems such as non-adaptive analysis, non-optimal time-frequency resolution and modal aliasing, which greatly affect the results.

[0005] With the development of deep learning technology, some researches have been carried out around the diagnosis of key components in the vehicle-track system based on neural networks. However, these deep learning-based methods need to manually collect vibration signals that can reflect the track state and make labels, which is often impractical and will consume a large amount of human resources. In addition, the collected vibration signals may have low quality, and the fault information is submerged, which will have a negative effect on the training of the network.

[0006] With the rapid development of intelligent manufacturing technology, digital twin emerges as the times require, which connects the physical world with the virtual world and provides a new way to obtain the physical world. Collecting simulation working condition data through the twin model and enriching the label scarce data set is an effective way to improve the reliability of fault diagnosis. Although the digital twin technology develops rapidly, the data simulated by the twin model and the actual data cannot completely maintain the same characteristics. How to overcome this difference and generalize the reliable diagnosis model from the twin model to the physical model is a key problem to ensure the accuracy of diagnosis.

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

[0008] In view of the above problems in the prior art, the digital twin driven rail state migration diagnosis method provided by the present application solves the problem of insufficient rail state diagnosis precision under the condition of sample label scarcity.

[0009] In order to achieve the above-mentioned purpose, the technical scheme adopted by the present application is as follows: a digital twin driven rail state migration diagnosis method, comprising the following steps:

[0010] S1, a digital twin model of a vehicle-track coupling system is established through multi-body dynamics;

[0011] S2, the vibration response signals of the axle box when the vehicle passes through different rail state sections are simulated by using the digital twin model of the vehicle-track coupling system, and a plurality of groups of vibration response signals are obtained;

[0012] S3, the axle box vibration signals of the axle box when the vehicle passes through different rail state sections are collected, a plurality of groups of actually collected axle box vibration signals are obtained, and time-frequency analysis is respectively performed on each group of vibration response signals and each group of axle box vibration signals through wavelet transform, to obtain the time-frequency diagram of each group of vibration response signals and the time-frequency diagram of each group of axle box vibration signals;

[0013] S4, the time-frequency diagram of each group of vibration response signals is taken as the source domain data of the sub-domain alignment migration network, and the time-frequency diagram of each group of axle box vibration signals is taken as the target domain data of the sub-domain alignment migration network, the sub-domain alignment migration network is trained, and a trained sub-domain alignment migration network is obtained;

[0014] S5, acquire the time-frequency graph of the to-be-tested axle box vibration signal, and input the trained sub-domain alignment transfer network to obtain the real state label of the to-be-tested axle box vibration signal, and complete the diagnosis and identification of different rail states.

[0015] Further, before the time-frequency analysis of each group of axle box vibration signals by wavelet transform in step S3, the low-pass filtering is performed on each group of axle box vibration signals.

[0016] Further, the sub-domain alignment transfer network in step S4 comprises a feature extractor, a state diagnostic device and N field discriminators; the feature extractor comprises 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 full connection layer connected in sequence; batch normalization operation is performed after each pooling layer;

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

[0018] Each of the field discriminators comprises a gradient reversal layer, a fifth full connection layer, a sixth full connection layer and a seventh full connection layer connected in sequence;

[0019] The fourth full connection layer and the seventh full connection layer in the sub-domain alignment transfer network both use the SoftMax activation function, and the remaining network layers all use the ReLU activation function.

[0020] Further, the step S4 is specifically:

[0021] S401, the time-frequency graph of each group of vibration response signals is taken as the source domain data of the sub-domain alignment transfer network, and the time-frequency graph of each group of axle box vibration signals is taken as the target domain data of the sub-domain alignment transfer network;

[0022] S402, the features of the source domain data and the features of the target domain data are extracted by using the sub-domain alignment transfer network respectively;

[0023] S403, according to the features of the source domain data and the features of the target domain data, the classification probability of the source domain data, the pseudo label of the target domain data and the field classification probability of all data are obtained by using the sub-domain alignment transfer network respectively;

[0024] S404, the network parameter of the alignment transfer network is updated according to the classification probability of the source domain data, the pseudo label of the target domain data and the field classification probability of all data;

[0025] S405, it is judged whether the network parameter of the sub-domain alignment transfer network reaches convergence, if yes, the trained sub-domain alignment transfer network is obtained, otherwise, the step S401 is returned.

[0026] Further, the expression of updating the network parameters of the sub-domain alignment transfer network in step S404 is:

[0027]

[0028] wherein, is the network parameter of the updated feature extractor; is the network parameter of the updated state diagnostic; is the network parameter of the updated state diagnostic; corresponding to minimizing F and C ; is the loss function of the sub-domain alignment transfer network; θ F is the network parameter of the feature extractor; θ C is the network parameter of the state diagnostic; θ D,k is the network parameter of the kth domain discriminator; is the set containing the network parameters of all domain discriminators; k is the number of domain discriminators; |C s | is the number of domain discriminators; is the updated set containing the network parameters of all domain discriminators; corresponding to maximizing and F 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 of all data; L w is the Wasserstein loss between the same class source domain data and the target domain data; L v is the intra-class variance loss of the same class target domain data.

[0029] Further, the expression of the classification loss of the source domain data is:

[0030]

[0031] wherein, n s is the number of source domain samples in the source domain data; i is the number of source domain samples in the source domain data; l c () is the cross-entropy loss; is the classification probability of each class of the ith source domain sample; is the feature extracted by the feature extractor from the ith source domain sample; is the ith source domain sample; is the class of the ith source domain sample; is the set of classification probabilities of the ith source domain sample; ​is a judgment function for judging whether the class represented by the kth domain discriminator is the correct classification of the ith source domain sample, and is 1 if yes, otherwise 0; log() is a logarithmic symbol; is the classification probability of the ith source domain sample belonging to the class represented by the kth domain discriminator.

[0032] Further, the expression of the domain discriminator loss of all data is:

[0033]

[0034] wherein, n k is the total number of samples belonging to the class represented by the kth domain discriminator in the source domain data and the target domain data; r is the sample number; l c is a cross-entropy loss; F(x r ) is the feature extracted after the feature extractor; D r (F(x k )) represents inputting the feature F(x r ) into the kth domain discriminator; d r is the domain label of the sample x r , if 0, then x r comes from the source domain, otherwise, x r comes from the target domain; I[y r =k] is a judgment function for judging whether the rth sample x r belongs to the class represented by the kth domain discriminator, and is 1 if yes, otherwise 0; y r is the class label of the rth sample; r is the domain classification probability of the rth sample x r . s

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

[0036]

[0037] wherein, n t 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 class label; n s,y＝k 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 t,y＝k is the probability distribution of the source domain sample in the source domain data belonging to the class 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 k-th 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 Π(P 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 the time 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 P s,y＝k Appears in And the distribution P t,y＝k Appears in The joint probability distribution of P s,y≠k Appears in And the distribution P t,y＝k Appears in The joint probability distribution of For the joint probability distribution Π(P s,y＝k ,P t,y＝k ) all the results obtained under 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 ) all the results obtained under and The expectation of the distance, there exists γ2 that minimizes this expectation; for The lower bound 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 feature extracted from the jth target domain sample belongs to the kth domain discriminator represented category; is the average value of all samples belonging to the kth domain discriminator represented category.

[0041] The beneficial effects of the present application are: the present application establishes a digital twin model using the dynamic parameters of the vehicle-rail coupling system in the physical world, simulates the axle box vibration response signal when the vehicle travels on rails in different states using the model, and expands the label-scarce axle box signal dataset; the multi-layer convolution layer in the feature extractor is used to extract features from the samples, and the fault features are learned adaptively layer by layer to realize feature learning of different rail states; an independent domain discriminator is provided 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 source domain and target domain data features in the same category while trying to increase the distance between the data features in different categories; the intra-class variance loss function is used to shorten the distance between the target domain data features in the same category to improve the accuracy of rail state diagnosis; both the time domain features and the frequency domain features of the signal are considered, and the wavelet time-frequency diagram of the simulation signal and the collected signal is used as a sample to improve the effect of network feature learning; the batch processing method used in the present application can solve the problem of poor initialization effect, help the gradient propagate to each layer of the network, accelerate the convergence of the model, effectively reduce the problem of model overfitting, avoid the collapse of the generated model, and prevent the convergence of the generated samples to the same point. BRIEF DESCRIPTION OF DRAWINGS

[0042] Figure 1 is the flow chart of the method of the present application.

[0043] Figure 2 is the structure diagram of the sub-domain alignment transfer network of the present application;

[0044] Figure 3 is the time-frequency domain comparison diagram of the simulation signal and the collected signal in the rail corrugation state in embodiment 2 of the present application;

[0045] Figure 4 is the time-frequency domain comparison diagram of the simulation signal and the collected signal in the inter-rail weld state in embodiment 2 of the present application;

[0046] Figure 5 is the confusion matrix diagram of the rail state diagnosis result in embodiment 2 of the present application;

[0047] Figure 6 ​A visualization diagram for extracting features in Embodiment 2 of the present application;

[0048] Figure 7 A diagram showing the change of diagnostic effect with the number of source domain samples in Embodiment 2 of the present application. DETAILED DESCRIPTION

[0049] The specific embodiments of the present application are described below to facilitate the understanding of the present application for those skilled in the art, but it should be clear that the present application is not limited to the scope of the specific embodiments, and for those skilled in the art, any changes within the spirit and scope of the present application as defined and determined by the appended claims are obvious, and all inventions utilizing the concept of the present application are within the scope of protection.

[0050] Embodiment 1

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

[0052] S1, establishing a digital twin model of a vehicle-track coupling system through multi-body dynamics;

[0053] S2, simulating the vibration response signals of the axle box when the vehicle passes through different rail state sections using the digital twin model of the vehicle-track coupling system, to obtain a plurality of sets of vibration response signals;

[0054] S3, collecting the axle box vibration signals of the axle box when the vehicle passes through different rail state sections to obtain a plurality of sets of actually collected axle box vibration signals, and performing time-frequency analysis on each set of vibration response signals and each set of axle box vibration signals through wavelet transform to obtain the time-frequency diagram of each set of vibration response signals and the time-frequency diagram of each set of axle box vibration signals;

[0055] S4, taking the time-frequency diagram of each set of vibration response signals as the source domain data of the sub-domain alignment migration network, and taking the time-frequency diagram of each set of axle box vibration signals as the target domain data of the sub-domain alignment migration network, training the sub-domain alignment migration network to obtain the trained sub-domain alignment migration network;

[0056] S5, obtaining the time-frequency diagram of the to-be-tested axle box vibration signal and inputting it into the trained sub-domain alignment migration network to obtain the real state label of the to-be-tested axle box vibration signal, and completing the diagnosis and identification of different rail states.

[0057] Before the time-frequency analysis of each set of axle box vibration signals through wavelet transform in step S3, each set of axle box vibration signals is first subjected to low-pass filtering.

[0058] In the embodiment, the digital twin model comprises 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 a rail and a roadbed. Mathematical models of rail corrugation and rail joint weld 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 part of the noise.

[0060] The sub-domain alignment transfer network in the step S4 comprises a feature extractor, a state diagnostic device and N field discriminators;

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

[0062] The state diagnostic device is configured to obtain a classification probability of the source domain data according to the features of the source domain data, and obtain pseudo-labels of the target domain data according to the features of the target domain data;

[0063] Each of the field discriminators is configured to obtain a respective field classification probability according to the features of the source domain data and the target domain data;

[0064] Wherein, N is the number of rail states.

[0065] The sub-domain alignment transfer network in the step S4 comprises a feature extractor, a state diagnostic device and N field discriminators; the feature extractor comprises 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; batch normalization operations are performed after each of the pooling layers;

[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 field discriminators comprises 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 sub-domain alignment transfer network both use a SoftMax activation function, and the remaining network layers all use a ReLU activation function.

[0069] In the embodiment, during the training process, the sub-domain alignment transfer network comprises 1 feature extractor, 1 state diagnostic device and N field discriminators, as shown in Figure 2The input signal wavelet time-frequency diagram size in the feature extractor is 28*28*3, the convolution kernel size of each layer of the convolution layer is set to 5*5, and the number of convolution kernels is set to 16, 16, and 32 respectively. A pooling layer is arranged after the first two layers of the convolution layer, and the convolution kernel size in the pooling layer is set to 3*3; in the state diagnostic device, the number of neurons of each layer of the full connection layer is 64, 32, and 3 respectively; in each field discriminator, the number of neurons of each layer of the full connection layer is 64, 32, and 2 respectively.

[0070] In the training process, the wavelet time-frequency diagram of the simulation 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 feature extractor is used to extract the source domain data features and the target domain data features. 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 corresponding pseudo label of the target domain data, so as to obtain the category of the target domain sample.

[0071] In the structure of the sub-domain alignment transfer network, the output layers of the state diagnostic device and the field discriminator use SoftMax activation function, and the rest use ReLU activation function, so as to solve the gradient disappearance and accelerate the convergence speed.

[0072] The step S4 is specifically:

[0073] S401, the time-frequency diagram of each group of vibration response signals is used as the source domain data of the sub-domain alignment transfer network, and the time-frequency diagram of each group of shaft box vibration signals is used as the target domain data of the sub-domain alignment transfer network;

[0074] S402, the feature extractor is used to extract the features of the source domain data and the features of the target domain data respectively;

[0075] S403, according to the features of the source domain data and the features of the target domain data, the classification probability of the source domain data, the pseudo label of the target domain data and the field classification probability of all data are obtained by using the sub-domain alignment transfer network;

[0076] S404, according to the classification probability of the source domain data, the pseudo label of the target domain data and the field classification probability of all data, the network parameter of the alignment transfer network is updated;

[0077] S405, whether the network parameter of the sub-domain alignment transfer network reaches convergence is judged, if yes, the trained sub-domain alignment transfer network is obtained, otherwise, the step S401 is returned.

[0078] The expression for updating the network parameter of the sub-domain alignment transfer network in the step S404 is:

[0079]

[0080] in, are the network parameters of the updated feature extractor; The network parameters of the updated status diagnostics; 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 classification loss of the source domain data is expressed as:

[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 feature extractor after the i-th source domain sample passes 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 that determines 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 discriminant loss of all data is:

[0085]

[0086] Wherein, 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 feature extracted by the feature extractor after the rth sample x r ; D k (F(x r )) represents inputting the feature F(x r ) into the kth domain discriminator; d r is the domain label of the sample x r , if 0, x r comes from the source domain, otherwise, x r comes from the target domain; I[y r =k] is to judge whether the rth sample x r belongs to the category represented by the kth domain discriminator, if yes, 1, otherwise, 0; y r is the category label of the rth sample; is the domain classification probability of the rth sample x r .

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

[0088]

[0089] Wherein, 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 the joint probability distribution of P s,y＝k and P t,y＝k ; Π(P 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 the time 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 P s,y＝k Appears in And the distribution P t,y＝k Appears in The joint probability distribution of P s,y≠k Appears in And the distribution P t,y＝k Appears in The joint probability distribution of For the joint probability distribution Π(P s,y＝k ,P t,y＝k ) all the results obtained under 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 ) all the results obtained under and The expectation of the distance, there exists γ2 that minimizes this expectation; for The lower bound of ; |||| is the norm symbol.

[0090] The expression of the 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; When 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 k-th domain discriminator.

[0093] In the embodiment, the class label of the target domain data is essentially a pseudo label, and is not a real class in the network training process.

[0094] In the 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 label of the source domain sample and the target domain sample in each domain discriminator and their domain classification probability; the Wasserstein loss is obtained by calculating the Wasserstein distance between the features of the same class source domain and target domain data, wherein the class of the target domain sample is determined by the pseudo label obtained by inputting the sample into the state diagnosticator; and the intra-class variance loss is obtained by calculating the distance between the features of the same class target domain data and the centroid position in the features.

[0095] The sub-domain alignment transfer network adopts the Adam method for parameter updating.

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

[0097] In the embodiment, the present application uses the dynamic parameters of the vehicle-rail coupling system in the physical world to establish a digital twin model, 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 dataset which is short of labels, and finally realizes the diagnosis task for the rail state. The present application uses the multi-layer convolutional layer in the feature extractor to extract features from the 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 application provides an independent domain discriminator for each class of sample, which ensures that the extracted source domain and target domain features will only be aligned according to their own class. The present application designs a Wasserstein loss function to shorten the features of the source domain and the target domain in the same class while trying to increase the distance between the data features of different classes. The present application uses an intra-class variance loss function to shorten the distance between the target domain data features in the same class to improve the accuracy of the rail state diagnosis. The present application simultaneously considers the time domain features and frequency domain features of the signal, and uses the wavelet time-frequency diagram of the simulation signal and the collected signal as samples to improve the feature learning effect of the network. The present application can solve the problem of poor initialization effect, help the gradient to propagate to each layer of the network, accelerate the convergence of the model, effectively alleviate the problem of model overfitting, avoid the collapse of the generated model, and prevent the convergence of the generated samples to the same point. The present application optimizes the sub-domain alignment transfer network using the Adam method, which dynamically adjusts the learning rate of each parameter using the first moment estimation and the second moment estimation of the gradient. In the backpropagation stage, the global optimal solution is obtained by updating the weight value, so that the loss function reaches the minimum value.

[0098] Example 2

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

[0100] As shown in Figure 3 , when the vehicle passed through the rail corrugation section, the similarity of the axle box vibration signal simulated by the digital twin model and the collected real axle box vibration signal in the time domain and frequency domain was very high, indicating that the digital twin model could accurately simulate the vibration state of the axle box when the rail appeared corrugation.

[0101] As shown in Figure 4 , when the vehicle passed through the rail weld section, the similarity of the axle box vibration signal simulated by the digital twin model and the collected real axle box vibration signal in the time domain and frequency domain was very high, indicating that the digital twin model could accurately simulate the vibration state of the axle box when the rail appeared weld.

[0102] As shown in Figure 5 , the sub-domain alignment transfer network can accurately diagnose the rail state, with an accuracy rate of more than 90% in various rail state diagnosis tasks, confirming the effectiveness of the method.

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

[0104] As shown in Figure 7 , gradually increasing the number of source domain samples for training, the rail state diagnosis accuracy gradually rises and tends to be stable, and the rail state diagnosis effect is still considerable under the condition of small source domain sample size, indicating that the proposed method can achieve good state diagnosis effect under the condition of lack 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, simulate the vibration response signals of the axle box when the vehicle passes through sections with different rail conditions, and obtain several sets of vibration response signals; S3. Collecting axle box vibration signals of axle boxes when a vehicle passes through sections with different rail states to obtain several groups of truly collected axle box vibration signals, and performing time-frequency analysis on each group of vibration response signals and each group of axle box vibration signals by wavelet transform to obtain time-frequency graphs of each group of vibration response signals and each group of axle box vibration signals; S4. Using the time-frequency graph of each group of vibration response signals as source domain data for the subdomain alignment transfer network, and using the time-frequency graph of each group of axle box vibration signals as target domain data for the subdomain alignment transfer network, the subdomain alignment transfer network is trained to obtain a trained subdomain alignment transfer network. The subdomain alignment migration network in step S4 includes a feature extractor, a state diagnostic device, and N domain discriminators; The feature extractor is used to extract features of source domain data and target domain data; 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; Each of the domain discriminators is used to obtain respective domain classification probabilities based on the features of the source domain data and the target domain data; Where N is the number of rail states; The step S4 is specifically as follows: S401, using the time-frequency graph of each group of vibration response signals as source domain data for the subdomain alignment transfer network, and using the time-frequency graph of each group of axle box vibration signals as target domain data for the subdomain alignment transfer network; S402, using the subdomain alignment transfer network to extract features of the source domain data and features of the target domain data respectively; S403: Based on the characteristics of the source domain data and the target domain data, the classification probability of the source domain data, the pseudo labels of the target domain data, and the domain classification probability of all data are obtained using the subdomain alignment transfer network; S404: Update the network parameters of the subdomain alignment migration network based on the classification probability of the source domain data, the pseudo-labels of the target domain data, and the domain classification probability of all data; S405: Determine whether the network parameters of the subdomain alignment transfer network have reached convergence. If so, obtain the trained subdomain alignment transfer network. Otherwise, return to step S401. S5. Obtain the time-frequency diagram of the axle box vibration signal to be tested and input it into 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, 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, characterized in that: The subdomain alignment transfer 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 includes 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 and seventh fully connected layers in the subdomain alignment transfer 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 expression for updating the network parameters of the subdomain alignment migration network in step S404 is: in, are the network parameters of the updated feature extractor; The network parameters of the updated status diagnostic device; To make Minimize the corresponding and ; Loss function for subdomain alignment transfer network; is the network parameter of the feature extractor; Network parameters for the status diagnostics; For the The network parameters of the domain discriminator; is a set of network parameters of discriminators in various fields; Number the domain discriminator; is the number of domain discriminators; is the updated set of network parameters of the discriminator in each field; To make Maximize the corresponding and ; is the classification loss of source domain data; Importance weight between domain alignment target and classification target; Domain discrimination loss for all data; It is the Wasserstein loss between source domain data and target domain data of the same category; It is the intra-class variance loss of the target domain data of the same category.

5. The rail state migration diagnosis method driven by digital twin according to claim 4 is characterized in that: The classification loss of the source domain data is expressed as: in, is the number of source domain samples in the source domain data; is the source domain sample number in the source domain data; is the cross entropy loss; For the The classification probability of each category of source domain samples; For the The features extracted by the source domain samples after passing through the feature extractor; For the source domain samples; For the The category of source domain samples; For the The classification probability set of source domain samples; To judge the Is the category represented by the domain discriminator the The judgment function of the correct classification of the source domain samples is 1 if yes, otherwise 0; is the logarithmic symbol; For the The source domain samples belong to The classification probability of the category represented by the domain discriminator.

6. The rail state migration diagnosis method driven by digital twin according to claim 4 is characterized in that: The expression of the domain discrimination loss of all the data is: in, The first The total number of samples of the category represented by each domain discriminator; Number the sample; is the cross entropy loss; For the samples Features extracted after passing through the feature extractor; Indicates that the feature Enter A domain discriminator; For samples If the field label is 0, then From the source domain, otherwise, from the target domain; To judge the samples Is it the first The category represented by the domain discriminator is 1 if yes, otherwise 0; For the The category labels of samples; For the samples The domain classification probability.

7. The rail state migration diagnosis method driven by digital twin according to claim 4 is characterized in that: The expression of Wasserstein loss between the source domain data and the target domain data of the same category is: in, is the number of source domain samples in the source domain data; is the source domain sample number in the source domain data; The domain discriminator number corresponding to the category label; is the number of target domain samples in the target domain data; is the target domain sample number in the target domain data; The source domain sample in the source domain data belongs to The probability distribution of the categories represented by each domain discriminator; The target domain sample in the target domain data belongs to The probability distribution of the categories represented by each domain discriminator; The source domain samples in the source domain data do not belong to The probability distribution of the categories represented by each domain discriminator; for and The joint probability distribution of for and The joint probability distribution of For the current source domain samples Belong to When the category represented by the domain discriminator is Extracted features; For the time Does not belong to When the category represented by the domain discriminator is Extracted features; For the target domain samples Belong to When the category represented by the domain discriminator is The features extracted from To make Appears in And distribution Appears in The joint probability distribution of For a certain Appears in And distribution Appears in The joint probability distribution of For the joint probability distribution All the following and The expectation of distance exists Minimize this expectation; for The infimum of For the joint probability distribution All the following and The expectation of distance exists Minimize this expectation; for The infimum of is the norm symbol.

8. The rail state migration diagnosis method driven by digital twin according to claim 4 is characterized in that: The expression of the intra-class variance loss of the target domain data of the same category is: in, is the number of target domain samples in the target domain data; is the target domain sample number in the target domain data; The domain discriminator number corresponding to the category label; For the current target domain samples Belong to When the category represented by the domain discriminator is The features extracted from For all belonging to The average value of samples of the category represented by the domain discriminator.

Images