DAS-based track disease detection and identification method, system and equipment, and medium

Abstract

The invention discloses a DAS-based track disease detection and identification method, system and device and a medium, belongs to track disease detection in the technical field of optical fiber sensing, and aims to solve the problem of low track disease detection and identification accuracy due to limited vibration displacement information in the prior art. According to the recognition method, three-mode signals including vibration displacement, vibration speed and vibration acceleration are obtained, the frequency spectrum characteristic and the time characteristic of each mode signal are calculated, the frequency spectrum characteristics and the time characteristics of the three-mode signals are fused, detection and recognition are carried out according to the obtained fused characteristics, and a classification result is obtained. According to the method, rich sensing information in the two modal signals of the speed and the acceleration is fully utilized, the spectrum features and the time features corresponding to the three-modal signals are fused and then are used for track disease recognition and classification, the track disease recognition and classification capacity is higher, and the accuracy of track disease recognition and classification is improved. And accurate detection of track diseases and fine identification of disease types can be realized.

Description

Technical Field

[0001] The present invention belongs to the field of optical fiber sensing technology, and relates to a track defect detection technology, and in particular to a track defect detection and identification method, system, equipment and medium based on DAS. Background Art

[0002] Railway transportation is the core force of my country's transportation. With the increase in traffic density, the load between railway tracks and train wheels gradually increases, resulting in track surface contact fatigue (such as fish scale injury) caused by wheel-rail interaction, wavy wear on the rail top (corrugation), and gaps between sleepers and ballast (whitening due to empty suspension) caused by severe corrugation. These track defects have become a major hidden danger to railway transportation safety. Existing track defect detection technologies include manual inspection, visual inspection, laser Doppler detection, and acoustic propagation fitting, but these methods are inefficient, costly, and easily affected by environmental factors such as light and climate. In addition, current detection methods can only be used during the maintenance window period, and cannot achieve all-weather online monitoring during train operation.

[0003] The distributed acoustic sensing (DAS) system based on the principle of phase-sensitive optical time-domain reflectometry (Φ-OTDR) uses widely laid communication optical cables to build a large-scale, high-density, low-cost sensing network, which has broad application prospects in the field of long-distance vibration event monitoring. When the DAS system is applied to railway track disease monitoring, the communication optical cables laid along the railway can be used as sensing optical fibers. When the train passes over the track, the interaction between the wheels and the track will generate vibrations and excite sound waves. After the optical cables along the railway sense these sound wave signals, they are collected, analyzed and processed by the DAS system. Since the vibration and sound wave characteristics generated by different track diseases when interacting with trains are different, by analyzing the differences in these vibration signals and combining intelligent signal processing and recognition algorithms, real-time monitoring of railway track diseases along the optical fiber can be achieved, thereby providing all-weather, full-line online monitoring methods for railway transportation safety.

[0004] The invention patent application with application number 202310249537.0 discloses a railway track defect identification method based on optical fiber distributed vibration detection, which includes the following steps: Step 1, emitting detection light pulses to the sensing optical fiber through the optical fiber distributed acoustic wave sensing system DAS, so as to realize real-time quantitative monitoring of vibration information along the sensing optical fiber, and obtain the original vibration signal; Step 2, constructing an initial data set for the original vibration signal obtained in Step 1, and marking the data samples of normal rail signals and three types of disease signals; Step 3, selecting M groups of normal rail signals and three types of disease signals from the initial multidimensional fusion feature vector data set obtained in Step 2 Data, construct a multidimensional fusion feature vector data set, and pre-weight the multidimensional fusion feature vector based on the average standard deviation ASD of the multidimensional fusion feature vectors of normal rail signals and three kinds of disease signals as the weight of each eigenvalue of the multidimensional fusion feature vector, and obtain a pre-weighted multidimensional fusion feature vector sample set; Step 4, input the pre-weighted multidimensional fusion feature vector sample set obtained in step 3 into the support vector machine SVM model for classification training, use the trained SVM model to realize the classification and recognition of normal rail signals and three kinds of disease signals, record the location and type of the disease and output and alarm, and realize real-time disease diagnosis and positioning output.

[0005] In addition, the invention patent with application number 2024105324443 also discloses an artificial intelligence-based equipment vibration fault diagnosis method, which includes the following steps: S1: real-time collection of equipment vibration signals; S2: preprocessing operations on the collected vibration signals; S3: extracting features from the preprocessed signals, and then merging the extracted features into a comprehensive feature vector; S4: building a deep neural network model; S5: when the model identifies an abnormal vibration pattern, immediately triggering the early warning mechanism; S6: after determining that the equipment has a fault, automatically generating a diagnostic report based on the analysis results of the deep learning model. By combining advanced artificial intelligence technology, especially convolutional neural network CNN and long short-term memory network LSTM, it can deeply analyze and understand the complex patterns of equipment vibration signals. This method not only significantly improves the accuracy of fault diagnosis, but also can accurately identify the fault type and its cause. Compared with the traditional experience-based fault detection method, it greatly enhances the reliability and effectiveness of fault diagnosis.

[0006] The above-mentioned invention patent application for the railway track defect identification method discloses the use of a DAS system to obtain the original vibration signal and realize the classification and identification of normal rail signals and three types of defect signals; the above-mentioned invention patent application for the equipment vibration fault diagnosis method discloses a composite model combining CNN and LSTM for identifying the pattern of vibration signals. However, since the optical fiber is generally laid in the cable trench or soil beside the track and is not directly in contact with the railway track, the vibration signal generated by the collected wheel-rail action is indirectly transmitted to the optical fiber through the roadbed and soil, which belongs to indirect measurement. During the underground propagation process, the vibration signal is inevitably affected by soil particles and soil structure, resulting in signal attenuation and filtering, reducing the coupling efficiency between wheel-rail vibration and optical fiber, and the useful component of the wheel-rail vibration signal perceived by the DAS system is further reduced relative to the original wheel-rail vibration signal. Therefore, the wheel-rail action information contained in the directly collected optical fiber vibration signal is limited, so it is difficult to explore the wheel-rail fault characteristics and accurately identify the fault, resulting in the final reduction in the accuracy of track disease detection and identification. In addition, when the DAS system obtains the original vibration signal, it mostly obtains parameters such as the number of signal pulses and vibration displacement, while the information contained in the vibration displacement signal is limited. When the subsequent model identifies and classifies the vibration displacement time-frequency diagram, the model's recognition and classification capabilities are limited, and the accuracy of track disease detection and identification is low. Summary of the invention

[0007] The purpose of the present invention is to solve the problem of low accuracy in track defect detection and identification due to limited vibration displacement information in the prior art, and to provide a track defect identification method, system, equipment and medium based on DAS. The vibration displacement is further differentiated into vibration velocity and vibration acceleration to form three modal signals of vibration displacement, vibration velocity and vibration acceleration. The spectrum energy distribution characteristics obtained after processing by variational mode decomposition (VMD) technology are fused with the time structure change characteristics output in the middle of the model and then input into the fully connected layer of the model for final identification and classification. The model can more easily learn richer information in the velocity time-frequency diagram and the acceleration time-frequency diagram, and the model has stronger identification and classification capabilities, so as to realize accurate detection of track defects and fine identification of defect types, and the accuracy of defect detection is high.

[0008] In order to achieve the above-mentioned purpose, the present invention specifically adopts the following technical solutions: A DAS-based track defect identification method comprises the following steps: Step S1, obtaining a tri-modal signal; Collect vibration displacement sample data, and obtain vibration velocity sample data and vibration acceleration sample data after two differential processes of the vibration displacement sample data; label the vibration displacement sample data, the vibration velocity sample data, and the vibration acceleration sample data to obtain label data; Step S2, obtaining spectrum and time characteristics; The vibration displacement data to be measured, the vibration velocity data to be measured, and the vibration acceleration data to be measured are respectively decomposed to obtain the energy distribution characteristics of the displacement spectrum, the energy distribution characteristics of the velocity spectrum, and the energy distribution characteristics of the acceleration spectrum; Processing the vibration displacement data to be measured, the vibration velocity data to be measured, and the vibration acceleration data to be measured to obtain the time structure change characteristics of the displacement, the time structure change characteristics of the velocity, and the time structure change characteristics of the acceleration; Step S3, fusion features; The energy distribution features of the displacement spectrum, the energy distribution features of the velocity spectrum, the energy distribution features of the acceleration spectrum, the time structure change features of the displacement, the time structure change features of the velocity and the time structure change features of the acceleration are fused to obtain fusion features; Step S4, identification and classification; The fused features are detected and identified to obtain the classification results.

[0009] Furthermore, in step S1, the specific method for obtaining the vibration velocity sample data and the vibration acceleration sample data is: Step S1-1: The vehicle generates vibration when passing through the track. The DAS system continuously accumulates the original signal trajectory on the time axis to form a space-time matrix. The spatial channel where the track disease is located is extracted from the space-time matrix to obtain a one-dimensional time series S with a total length of L sampling points, and the one-dimensional time series S is divided into M vibration displacement sample data with a length of N sampling points along the time axis. The range of the i-th sample in the time series S is:

[0010] Step S1-2, performing differential differentiation on the segmented vibration displacement sample data to obtain vibration velocity sample data; performing differential differentiation on the vibration velocity sample data to obtain vibration acceleration sample data; wherein the calculation formula for differential differentiation is:

[0011] in, represents the sampling starting point, Indicates the sliding offset, Indicates the sample number, represents the total number of samples, They represent the velocity and acceleration modes respectively. Indicates a point in time.

[0012] Furthermore, in step S2, the variational modal decomposition technique is used to perform signal decomposition on the vibration displacement sample data, the vibration velocity sample data, and the vibration acceleration sample data, and the spectrum energy distribution characteristics are calculated, specifically: Step S2-1, the multimodal signal composed of vibration displacement sample data, vibration velocity sample data, and vibration acceleration sample data is expressed as:

[0013] The variational mode decomposition technique is used to decompose the multimodal signal into IMF components, so that each mode function The sum of the estimated bandwidths is the smallest, expressed as:

[0014]

[0015]

[0016] Step S2-2, calculate the energy of each IMF component, the calculation formula is:

[0017] The calculated energy is normalized and the calculation formula is:

[0018] Step S2-3, construct and obtain the corresponding spectrum energy distribution characteristics, the calculation formula is:

[0019] in, Indicates time point The total number of represents the total number of samples, Indicates time point signal, Indicates the sample number, They represent displacement, velocity and acceleration modes respectively. Representing modality The signal is decomposed into Quantity, represents the number of decomposition levels, , Respectively The magnitude and phase of Representing modality The signal is decomposed into The time signal point of each component.

[0020] Furthermore, in step S2-1, the variational mode decomposition technique is used to decompose the different mode signals respectively, which specifically includes the steps of: Step S2-1-1, set the number of decomposition levels , for each , the Hilbert transform is used to calculate the corresponding analytical signal and obtain the one-sided spectrum. The calculation formula is:

[0021]

[0022] Step S2-1-2, for each , through the aliasing of the exponential term corresponding to its center, each The spectrum is modulated to the corresponding "baseband", and the calculation formula is:

[0023]

[0024] Step S2-1-3, square norm of the signal gradient of step S2-1-2 Solve and estimate The bandwidth of , thus constructing the constrained variational problem:

[0025]

[0026]

[0027]

[0028] Step S2-1-4, solve the variational problem of step S2-1-3 and construct the Lagrangian augmented function:

[0029] Step S2-1-5, using the alternating direction multiplier method to solve the unconstrained variational problem of step S2-1-4, , The optimal solution is obtained through iterative updating. The iterative formula is:

[0030]

[0031]

[0032] Step S2-1-6, Update, the update formula is:

[0033]

[0034] Step S2-1-7, given determination accuracy If the convergence condition is met, the iteration stops; otherwise, it returns to step S2-1-5 to continue the iteration; the convergence condition is:

[0035]

[0036] Step S2-1-8, the iterative completion of the obtained Perform inverse Fourier transform and get An IMF component with a certain bandwidth and fluctuating around the center frequency.

[0037] Furthermore, in step S2, step S3, and step S4, the track defect detection and recognition model is used to obtain the temporal structural change characteristics, perform feature fusion, and perform recognition and classification; The constructed track defect detection and recognition model includes three CNN-LSTM networks, one AFF module and two fully connected layers. When training the constructed track defect detection and recognition model, the specific training steps are as follows: First, the vibration displacement sample data collected by the DAS system is subjected to two differential processes to obtain the vibration velocity sample data and the vibration acceleration sample data; Then, the vibration displacement sample data, the vibration velocity sample data, and the vibration acceleration sample data are respectively input into the corresponding CNN-LSTM networks, and the three CNN-LSTM networks respectively output the vibration displacement sample time structure change characteristics, the vibration velocity sample time structure change characteristics, and the vibration acceleration sample time structure change characteristics; the vibration displacement sample data, the vibration velocity sample data, and the vibration acceleration sample data are respectively subjected to signal decomposition to obtain the vibration displacement sample spectrum energy distribution characteristics, the vibration velocity sample spectrum energy distribution characteristics, and the vibration acceleration sample spectrum energy distribution characteristics; Then, the energy distribution features of the spectrum of vibration displacement samples, the energy distribution features of the spectrum of vibration velocity samples, the energy distribution features of the spectrum of vibration acceleration samples, the time structure change features of vibration displacement samples, the time structure change features of vibration velocity samples, and the time structure change features of vibration acceleration samples are input into the AFF module for fusion to obtain fusion features; Finally, the fusion features are input into the first fully connected layer. The output of the first fully connected layer is used as the input of the second fully connected layer. The second fully connected layer outputs the classification result. The loss is calculated based on the classification result and the corresponding label data, and the parameters of the track disease detection and recognition model are updated.

[0038] Furthermore, in step S2, a CNN-LSTM network is used to extract the temporal structural change characteristics of the sample data, including a CNN subnetwork and an LSTM layer. The CNN subnetwork includes a convolution layer Conv1, a pooling layer MaxPool1, ReLU1, a convolution layer Conv2, a pooling layer MaxPool2, ReLU2, a convolution layer Conv3, a pooling layer MaxPool3, and ReLU3.

[0039] Furthermore, the AFF module includes two convolutional layers. The specific method of the AFF module for feature fusion is: The Xavier initialization method is used to initialize the structural parameters of the convolutional layer of the AFF module; Parallel splicing of the spectrum energy distribution characteristics corresponding to the three-modal signals to obtain the spectrum energy characteristics ; Parallel splicing of the time structure change characteristics corresponding to the three modal signals to obtain the time structure characteristics ;

[0040] The spectrum energy characteristics , Time structure characteristics Perform convolution operation to obtain spectrum energy characteristics , Time structure characteristics ;

[0041] Spectral energy characteristics and temporal structural characteristics After adding, local attention features and global attention features are obtained respectively;

[0042] The local attention feature and the global attention feature are added together and the result is obtained through the Sigmoid function. , and according to the results Get fusion features .

[0043] Furthermore, in step S4, the fused features are passed through two fully connected layers and the final classification result is obtained using the SoftMax method, and the temporal structure change feature extraction module, feature fusion module and recognition are trained offline based on the same objective function to obtain the optimal model and the feature fusion recognition model.

[0044] The specific process of designing the track defect detection and identification model includes three steps: network structure design and parameter setting, network initialization and network training, and parameter update and optimization. The specific methods are as follows: Step S4-1, network structure design and parameter setting; After the temporal structure change characteristics and spectral energy distribution characteristics of the three modal signals are extracted, they are simultaneously input into the attention module for fusion. The output of the AFF module passes through two fully connected layers in turn and the final classification result is obtained using the SoftMax method.

[0045] Step S4-2: Network initialization and network training; After the parameters of the rail defect detection and recognition model are initialized, the predicted category probability distribution is obtained through network forward propagation, and the cross entropy loss function is used to calculate the loss value between the category probability output by the fully connected layer and the true probability. The loss value is used for back propagation to calculate the gradient of each learning parameter. Finally, the model parameters are updated according to the principle of gradient descent according to the specified learning rate. , the model parameters Including matrix weights and bias . Take the first iterative learning process of the model as an example to illustrate:

[0046] Initialize the fusion network parameters based on the attention module. Good initialization parameters make the model easier to learn and converge quickly. The present invention uses the Xavier initialization method to initialize the network structure parameters. In order to ensure that the variance of each layer is consistent during forward propagation and backward propagation, the distribution range of the parameter random initialization is the number of input parameters passing through the layer. , output parameter number The uniform distribution within the obtained distribution range, the distribution range formula is:

[0047]

[0048] Step S4-3: parameter update and optimization; The back propagation error is calculated according to the set objective function and the obtained predicted label, and the error is used to update and tune the parameters of the constructed fusion network based on the attention module.

[0049] Furthermore, in step S4-3, the specific steps are as follows: Step S4-3-1 Fully connected layer FC receives fusion features , output predicted label , the calculation formula is:

[0050]

[0051] in, , , Represent the weight matrix, bias vector and activation function respectively, Represents the predicted label.

[0052] Step S4-3-2 uses the cross information entropy loss function to calculate the loss value between the predicted label and the true label: According to the cross information entropy loss function, the distance between the predicted event label and the true label is calculated, and the loss value The calculation formula is as follows:

[0053]

[0054] in: M represents the total number of samples, represents the true label.

[0055] Step S4-3-3 uses the loss value to reversely calculate the parameter gradient of the fusion network model based on the attention mechanism, and uses the parameter gradient to update the fusion network deep learning model based on the attention mechanism; the present invention adopts the Adam algorithm for update optimization, and the calculation steps are as follows:

[0056]

[0057]

[0058] in: is the target gradient for calculation, , They are the first-order and second-order moment estimates of the gradient, , are the exponential decay rates of the first and second order moments of the gradient, respectively.

[0059] Let the tth iteration, , , , then the network parameter update formula is:

[0060]

[0061] in: is the learning rate, For very small numbers, prevent division by zero.

[0062] A track defect identification system based on DAS, comprising: The tri-modal signal acquisition module is used to perform two differential processes on the collected vibration displacement data to be measured to obtain the vibration velocity data to be measured and the vibration acceleration data to be measured; The spectrum and time feature acquisition module is used to perform signal decomposition on the vibration displacement data to be measured, the vibration velocity data to be measured, and the vibration acceleration data to be measured, respectively, to obtain the energy distribution characteristics of the displacement spectrum, the energy distribution characteristics of the velocity spectrum, and the energy distribution characteristics of the acceleration spectrum; Processing the vibration displacement data to be measured, the vibration velocity data to be measured, and the vibration acceleration data to be measured to obtain the time structure change characteristics of the displacement, the time structure change characteristics of the velocity, and the time structure change characteristics of the acceleration; A feature fusion module is used to fuse the displacement spectrum energy distribution features, the velocity spectrum energy distribution features, the acceleration spectrum energy distribution features, the displacement time structure change features, the velocity time structure change features and the acceleration time structure change features to obtain a fusion feature; The recognition and classification module is used to detect and recognize the fusion features of the signal to be tested and obtain the classification result.

[0063] A computer device comprises a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the above method.

[0064] A computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the processor is caused to execute the steps of the above method.

[0065] The beneficial effects of the present invention are as follows: 1. In the present invention, for the first time, a multimodal vibration information extraction method is proposed for the application of DAS in train track defect detection, and it is applied to the early detection of track defects and the high-precision identification of defect types. Compared with the existing track defect detection technology, DAS has the advantages of low cost, long distance, high precision, and all-weather. At the same time, according to research, compared with vibration displacement, vibration velocity and vibration acceleration have richer frequency domain information. Therefore, this application will convert only a single vibration displacement signal into the simultaneous use of displacement, velocity, and acceleration three modal signals, effectively enriching the dimension of the recognition model perception information. 2. In the present invention, VMD is used to replace traditional signal decomposition methods such as spectrum equal division and empirical mode decomposition (EMD) to avoid problems such as mode aliasing and endpoint effect, provide richer and more accurate frequency domain information for the analysis of vibration signals, and effectively improve the frequency domain information representation capability of each vibration mode signal in the low frequency band; the CNN-LSTM combined network is used to extract the deep features of vibration signals of different modes, and effectively represent the multi-scale time information of different disease signals. The network extracts the local structural features of the signal through CNN, and uses LSTM to capture the interdependence of the signal on a long time scale, and digs out the temporal relationship between short-time signals, which significantly enhances the feature representation capability of the disease signal on a time scale.

[0066] 3. In the present invention, the attention module is used to fuse the displacement spectrum energy distribution characteristics, velocity spectrum energy distribution characteristics, acceleration spectrum energy distribution characteristics, displacement time structure change characteristics, velocity time structure change characteristics and acceleration time structure change characteristics, focusing on key features, improving the model's ability to understand signal characteristics, and realizing accurate detection of disease signals and fine identification of disease types. BRIEF DESCRIPTION OF THE DRAWINGS

[0067] Figure 1 A flow chart of a track defect detection and identification method based on DAS multi-modal vibration information extraction provided by the present invention; Figure 2 The DAS system structure and working principle of the present invention; Figure 3 Schematic diagram of three modal signals in the present invention; Figure 4 It is a schematic diagram of the spectrum of the signal VMD decomposition in the present invention; Figure 5 It is a schematic diagram of the spectrum energy distribution characteristics of different disease types in the present invention; Figure 6 It is the CNN-LSTM network structure in the present invention; Figure 7 The structure of the AFF module in the present invention; Figure 8 The test set confusion matrix of the recognition model of the present invention in the track disease signal data set; Fig. 9 This is a structural diagram of a single-mode input signal recognition model in the present invention; Fig.10 The accuracy of different models of the present invention is compared in the test set of the track disease signal dataset. DETAILED DESCRIPTION

[0068] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments.

[0069] Therefore, based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making any creative work shall fall within the scope of protection of the present invention.

[0070] Example 1 This embodiment provides a DAS-based track defect identification method, which uses DAS multi-modal vibration information extraction. The DAS system uses the communication optical fiber buried near the train track as the sensing medium, without the need for additional laying. The entire signal processing process is as follows: Figure 1 As shown, the specific steps include: Step S1, obtaining a tri-modal signal; Collecting vibration displacement data to be measured, performing two differential processes on the collected vibration displacement data to be measured, and obtaining vibration velocity data to be measured and vibration acceleration data to be measured; The DAS system and the communication optical fiber laid along the track are used to collect effective signals of wheel-rail interaction when the train passes. Three types of defects (including corrugation, whitening due to empty suspension, and fish scale damage) and one normal type are selected as event labels. The one-dimensional time series signals at the typical defect space points are segmented by sliding, and after two differential derivatives, three modal signal data sets of vibration displacement, vibration velocity, and vibration acceleration, as well as the corresponding label data, are constructed.

[0071] Step S2, obtaining spectrum and time characteristics; The vibration displacement data to be measured, the vibration velocity data to be measured, and the vibration acceleration data to be measured are respectively decomposed to obtain the stationary IMF components of different scales, and the energy distribution characteristics of the displacement spectrum, the energy distribution characteristics of the velocity spectrum, and the energy distribution characteristics of the acceleration spectrum are calculated; The vibration displacement data to be measured, the vibration velocity data to be measured, and the vibration acceleration data to be measured are processed, and the displacement time structure change characteristics, the velocity time structure change characteristics, and the acceleration time structure change characteristics are obtained through three parallel CNN-LSTM networks; Step S3, fusion features; The energy distribution features of the displacement spectrum, the energy distribution features of the velocity spectrum, the energy distribution features of the acceleration spectrum, the time structure change features of the displacement, the time structure change features of the velocity and the time structure change features of the acceleration are fused to obtain fusion features; Step S4, identification and classification; The fused features are detected and identified to obtain the classification results.

[0072] Example 2 This embodiment further optimizes the present invention based on the embodiment 1.

[0073] The DAS system used for signal acquisition has the following structure: Figure 2 As shown, it includes a narrow bandwidth laser, an acousto-optic coupler, an erbium-doped fiber amplifier, a circulator, a sensing fiber, a phase demodulation device, a data collector, and a signal processing host; the narrow bandwidth laser generates continuous light, the acousto-optic coupler modulates the continuous light into a pulse light with a specific period and pulse width, the erbium-doped fiber amplifier amplifies the pulse light, the amplified pulse light enters the sensing fiber through the circulator, the pulse light propagates in the sensing fiber and generates backward Rayleigh scattered light, the backward Rayleigh scattered light is linearly modulated by the coupler in the phase demodulation device and then converted by the data collector into the original signal trajectory distributed along the space.

[0074] The specific method of collecting signals is: Step S1-1: The vehicle generates vibration when passing through the track. The DAS system continuously accumulates the original signal trajectory on the time axis to form a space-time matrix. The spatial channel where the track disease is located is extracted from the space-time matrix to obtain a one-dimensional time series S with a total length of L sampling points, and the one-dimensional time series S is divided into M vibration displacement sample data with a length of N sampling points along the time axis. The range of the i-th sample in the time series S is:

[0075] Step S1-2, performing differential differentiation on the segmented vibration displacement sample data to obtain vibration velocity sample data; performing differential differentiation on the vibration velocity sample data to obtain vibration acceleration sample data; wherein the calculation formula for differential differentiation is:

[0076] in, represents the sampling starting point, Indicates the sliding offset, Indicates the sample number, represents the total number of samples, They represent the velocity and acceleration modes respectively. In this embodiment, the sampling frequency Set to 1kHz, sample length N to 5000, slide offset Set to 1000.

[0077] Typical signal sample diagrams of the three modes are shown in Figure 3 As shown, the four columns correspond to four rail disease labels (normal, corrugation, whitening due to empty suspension, and fish scale injury), and the four rows correspond to the original signal time domain waveform, displacement modal time-frequency diagram, velocity modal time-frequency diagram, and acceleration modal time-frequency diagram. As shown in the figure, the main frequency components of the displacement modal signal are concentrated below 100Hz, and it is difficult to see the difference between the four types of samples. There are obvious differences in the spectral distribution and energy intensity of the velocity mode and acceleration mode of different types of signals. The low frequency (below 200Hz) of the normal type velocity modal signal is stronger than the high frequency (300-500Hz), and the spectrum of the acceleration modal signal is evenly distributed and concentrated in 1-3s; the spectrum of the velocity modal signal of corrugation and fish scale injury types is concentrated below 200Hz, and the spectrum of the acceleration modal signal is distributed in 0-200Hz and 400-500Hz, and the energy of the fish scale injury type signal is significantly stronger than that of the corrugation type; the frequency components of the velocity mode and acceleration mode of the empty suspension type are evenly distributed in the entire frequency range (0-500Hz). The simultaneous application of signals from multiple modes to the identification of track defect types effectively enriches the perception capability of the identification model and improves its resolution.

[0078] Typical rail disease labels include three types of diseases (corrugation, whitening of empty rails, and fish scale damage) and one normal type. The data in the typical event dataset are divided into training set and test set in a ratio of 7:3.

[0079] Example 3 This embodiment further optimizes the present invention based on Embodiment 2.

[0080] For the above multimodal data set, step S2 uses variational mode decomposition (VMD) technology to decompose the vibration displacement sample data, vibration velocity sample data, and vibration acceleration sample data into K stationary IMF components of different scales, and calculates the high-resolution spectrum energy distribution characteristics of the signal. The specific steps are: Step S2-1, the multimodal signal composed of the vibration displacement data to be measured, the vibration velocity data to be measured, and the vibration acceleration data to be measured is expressed as:

[0081] The variational mode decomposition technique is used to decompose the multimodal signal into K IMF components, so that each mode function The sum of the estimated bandwidths is the smallest, expressed as:

[0082]

[0083]

[0084] Step S2-2, calculate the energy of each IMF component, the calculation formula is:

[0085] The calculated energy is normalized and the calculation formula is:

[0086] Step S2-3, construct and obtain the corresponding spectrum energy distribution characteristics, the calculation formula is:

[0087] in, Indicates time point The total number of Indicates the total number of vibration displacement data to be measured, Indicates time point signal, Indicates the data sequence number. They represent displacement, velocity and acceleration modes respectively. Representing modality The signal is decomposed into Quantity, represents the number of decomposition levels, , Respectively The magnitude and phase of Representing modality The signal is decomposed into The time signal point of each component.

[0088] The core of this embodiment is to use VMD to decompose the signal into The basic idea is to construct and solve the variational problem, which solves the problems of mode overlap and endpoint effect in signal decomposition. The variational mode decomposition technology is used to decompose different mode signals separately, which includes the following steps:

[0089] Step S2-1-1, set the number of decomposition levels , for each , the Hilbert transform is used to calculate the corresponding analytical signal and obtain the one-sided spectrum. The calculation formula is:

[0090]

[0091] Step S2-1-2, for each , through the aliasing of the exponential term corresponding to its center, each The spectrum is modulated to the corresponding "baseband", and the calculation formula is:

[0092]

[0093] Step S2-1-3, square norm of the signal gradient of step S2-1-2 Solve and estimate The bandwidth of , thus constructing the constrained variational problem:

[0094]

[0095]

[0096]

[0097] Step S2-1-4, solve the variational problem of step S2-1-3 and construct the Lagrangian augmented function:

[0098] Step S2-1-5, using the alternating direction multiplier method to solve the unconstrained variational problem of step S2-1-4, , The optimal solution is obtained through iterative updating. The iterative formula is:

[0099]

[0100]

[0101] Step S2-1-6, Update, the update formula is:

[0102]

[0103] Step S2-1-7, given determination accuracy If the convergence condition is met, the iteration stops; otherwise, it returns to step S2-1-5 to continue the iteration; the convergence condition is:

[0104]

[0105] Step S2-1-8, the iterative completion of the obtained Perform inverse Fourier transform and get An IMF component with a certain bandwidth and fluctuating around the center frequency;

[0106] in, represents the impulse function, represents an imaginary number, Indicates a point in time, represents the center frequency corresponding to the IMF component, represents the K IMF components obtained by decomposition, represents the center frequency of the IMF component, represents the quadratic penalty term, , All represent Lagrange multipliers, It means to find the derivative of the function with respect to time t. represents a multimodal signal, Indicates frequency, , , Respectively represent the components ,Signal , Lagrange multipliers The Fourier transform of represents the number of iterations, represents the noise margin parameter, Indicates the judgment accuracy.

[0107] The number of VMD decompositions is set to , the spectrum diagram of each IMF component after decomposition is as follows Figure 4 As shown in Figure 1, the lower frequency band (0-200Hz) containing the main frequency components of wheel-rail interaction is decomposed into several stationary IMF components of different scales, and the signal energy is concentrated in components IMF6-IMF12. The energy characteristics of signals of different disease types are shown in Figure 1. Figure 5 As shown in the figure, the signal energy of the normal type and the empty hanging whitening type is mainly distributed in IMF9-IMF12. The energy characteristic values ​​between each IMF are similar, but the normal type is stronger than the empty hanging whitening type; the signal energy of the corrugated type is concentrated in IMF12, and the signal energy of the fish scale type is concentrated in IMF11 and IMF12. Compared with traditional signal decomposition methods such as spectrum equal division and EMD, VMD avoids problems such as mode aliasing and endpoint effect, and effectively mines out weak disease information hidden in the complex structure of the signal. The use of spectrum energy distribution characteristics for characterization and quantification provides richer and more accurate frequency domain information for the analysis of vibration signals, and effectively improves the frequency domain information characterization capability of each vibration modal signal in the low frequency band.

[0108] Example 4 This embodiment further optimizes the present invention based on Embodiment 2.

[0109] In the present invention, steps S2, S3 and S4 all use the track defect detection and recognition model to obtain the temporal structural change characteristics, perform feature fusion and recognition and classification; The constructed track defect detection and recognition model includes three CNN-LSTM networks, one AFF module and two fully connected layers. When training the constructed track defect detection and recognition model, the specific training steps are as follows: First, the vibration displacement sample data collected by the DAS system is subjected to two differential processes to obtain the vibration velocity sample data and the vibration acceleration sample data; Then, the vibration displacement sample data, the vibration velocity sample data, and the vibration acceleration sample data are respectively input into the corresponding CNN-LSTM networks, and the three CNN-LSTM networks respectively output the vibration displacement sample time structure change characteristics, the vibration velocity sample time structure change characteristics, and the vibration acceleration sample time structure change characteristics; the vibration displacement sample data, the vibration velocity sample data, and the vibration acceleration sample data are respectively subjected to signal decomposition to obtain the vibration displacement sample spectrum energy distribution characteristics, the vibration velocity sample spectrum energy distribution characteristics, and the vibration acceleration sample spectrum energy distribution characteristics; Then, the energy distribution features of the spectrum of vibration displacement samples, the energy distribution features of the spectrum of vibration velocity samples, the energy distribution features of the spectrum of vibration acceleration samples, the time structure change features of vibration displacement samples, the time structure change features of vibration velocity samples, and the time structure change features of vibration acceleration samples are input into the AFF module for fusion to obtain fusion features; Finally, the fusion features are input into the first fully connected layer. The output of the first fully connected layer is used as the input of the second fully connected layer. The second fully connected layer outputs the classification result. The loss is calculated based on the classification result and the corresponding label data, and the parameters of the track disease detection and recognition model are updated.

[0110] Example 5 This embodiment further optimizes the present invention based on Embodiment 4.

[0111] In the present invention, three parallel CNN-LSTM networks are constructed, and the signals of the three modes of vibration displacement, vibration velocity and vibration acceleration are respectively used as the input of the three CNN-LSTM networks, that is, a signal of one mode is input into one CNN-LSTM network, and the three CNN-LSTM networks with the same structure are sequentially passed through, sharing the network loss, and the CNN-LSTM network outputs the time structure change characteristics of the modal signal .

[0112] The specific structure of the CNN-LSTM network is as follows Figure 6As shown in the figure, it includes a CNN subnetwork and an LSTM layer. The CNN subnetwork includes a convolution layer Conv1, a pooling layer MaxPool1, ReLU1, a convolution layer Conv2, a pooling layer MaxPool2, ReLU2, a convolution layer Conv3, a pooling layer MaxPool3, and ReLU3. The specific network structure parameters of the CNN-LSTM network are: 1*3 / 1*1 / 1, 1*10, ReLU, 1*3 / 1*1 / 1, 1*5, ReLU, 1*3 / 1*1 / 1, 1*5, ReLU, 64.

[0113] Example 6 This embodiment further optimizes the present invention based on Embodiment 3, Embodiment 4 and Embodiment 5.

[0114] Based on the spectral energy distribution features extracted in Example 3 and the temporal structure change features extracted in Example 5, the temporal structure change features and spectral energy distribution features of the three modal signals of displacement, velocity and acceleration are fused based on the attention (AFF) module. The specific steps are as follows: The spectral energy distribution characteristics of the three modal signals are parallel spliced ​​and recorded as

[0115] The temporal structural change characteristics of the three modal signals are parallelly spliced ​​and recorded as

[0116] in They correspond to displacement, velocity, and acceleration modes respectively. and Perform a corresponding convolution operation respectively to make the two have the same dimension. Suppose the spectrum energy distribution characteristics after the convolution operation are , the temporal structural change characteristics are , the attention module AFF module is as follows Figure 7 As shown, and After adding, we obtain the local attention features and the global attention features respectively; let the input feature be , then the corresponding local attention feature is calculated as

[0117]

[0118] The calculation formula of the global attention feature is

[0119] in, is the local attention feature, is the global attention feature, is a convolution operation with a convolution size of 1×1. represents the BatchNorm layer, represents the ReLU activation function, represents the average pooling operation. Therefore, based on and Obtain local attention features and global attention features, add them together and pass them through the Sigmoid function for:

[0120]

[0121] The fusion feature after the attention fusion module is recorded as , the result is as follows:

[0122]

[0123] The network model parameters of the AFF module are set to: 1*5 / 1*1 / 0, 1*5 / 1*1 / 0 (convolution kernel / stride / padding).

[0124] Example 7 This embodiment further optimizes the present invention based on Embodiment 6.

[0125] In step S4, the fused features are passed through two fully connected layers and the final classification result is obtained using the SoftMax method. Based on the same objective function, the temporal structure change feature extraction module, feature fusion module and recognition are trained offline to obtain the optimal model and the feature fusion recognition model.

[0126] The specific process of designing the rail defect detection and identification model includes three steps: network structure design and parameter setting, network initialization and network training, and parameter update and optimization. The specific methods are as follows: Step S4-1, network structure design and parameter setting; After extracting the temporal structure change characteristics and spectral energy distribution characteristics of the three modal signals, they are simultaneously input into the attention module for fusion. The output of the AFF module passes through two fully connected layers in turn and uses the SoftMax method to obtain the final classification result. Among them, the parameters of the fully connected layer are set to: Linear / 32*64 / 1*128, Linear / 1*128 / 1*4 (activation function / input dimension / output dimension).

[0127] Step S4-2, network initialization and network training; After the parameters of the rail defect detection and recognition model are initialized, the predicted category probability distribution is obtained through network forward propagation, and the cross entropy loss function is used to calculate the loss value between the category probability output by the fully connected layer and the true probability. The loss value is used for back propagation to calculate the gradient of each learning parameter. Finally, the model parameters are updated according to the principle of gradient descent according to the specified learning rate. , the model parameters Including matrix weights and bias . Take the first iterative learning process of the model as an example to illustrate:

[0128] Initialize the fusion network parameters based on the attention module. Good initialization parameters make the model easier to learn and converge quickly. The present invention uses the Xavier initialization method to initialize the network structure parameters. In order to ensure that the variance of each layer is consistent during forward propagation and backward propagation, the distribution range of the parameter random initialization is the number of input parameters passing through the layer. , output parameter number The uniform distribution within the obtained distribution range, the distribution range formula is:

[0129]

[0130] Step S4-3, parameter update and tuning; The back propagation error is calculated according to the set objective function and the obtained predicted label, and the error is used to update and tune the parameters of the constructed fusion network based on the attention module.

[0131] Furthermore, in step S4-3, the specific steps are as follows: Step S4-3-1 Fully connected layer FC receives fusion features , output predicted label , the calculation formula is:

[0132]

[0133] in, , , Represent the weight matrix, bias vector and activation function respectively, Represents the predicted label.

[0134] Step S4-3-2 uses the cross information entropy loss function to calculate the loss value between the predicted label and the true label: According to the cross information entropy loss function, the distance between the predicted event label and the true label is calculated, and the loss value The calculation formula is as follows:

[0135]

[0136] in: M represents the total number of samples, represents the true label.

[0137] Step S4-3-3 uses the loss value to reversely calculate the parameter gradient of the fusion network model based on the attention mechanism, and uses the parameter gradient to update the fusion network deep learning model based on the attention mechanism; the present invention adopts the Adam algorithm for update optimization, and the calculation steps are as follows:

[0138]

[0139]

[0140] in: is the target gradient for calculation, , They are the first-order and second-order moment estimates of the gradient, , are the exponential decay rates of the first and second order moments of the gradient, respectively.

[0141] Let the tth iteration, , , , then the network parameter update formula is:

[0142]

[0143] in: is the learning rate, For very small numbers, prevent division by zero.

[0144] In the feature fusion network model based on the AFF module, the model parameters are first used Update the network model. Then, determine whether the network model has converged based on the change in the training loss value. Once the convergence state is reached, the training process stops; otherwise, jump to step S4-3-2 and continue to iterate until the set maximum number of iterations is reached. When the loss function value is less than a certain threshold or the iteration exceeds the preset threshold, the model is considered to have converged and the above iteration process is stopped. Finally, the model with the best result is selected as the final event recognition model.

[0145] Example 8 This embodiment is based on Embodiment 7 and performs classification and recognition tests on the trained model.

[0146] Based on the previous data set, the model is trained and tested. The confusion matrix of the test set under the track disease data set is as follows Figure 8To evaluate the performance of this model, we compared its recognition effect with several other models based on different modal signals and feature extraction methods: 1) Compare model A with three models that use only a single modal signal input (model B-1: displacement mode; model B-2: velocity mode; model B-3: acceleration mode). The structure of the single modal signal input model is as follows: Fig. 9 As shown, the method for extracting the spectrum energy distribution feature and the time structure change feature is the same as that of Example 3 and Example 5. Fig.10 (a) is the recognition accuracy distribution histogram of different input signal modes. The results show that compared with the single-modal signal model, the multimodal signal of model A improves the recognition rate by 0.57%-3.27%, indicating that the introduction of DAS multimodal signals enriches the dimension of network perception information and effectively improves the recognition ability of the model.

[0147] 2) Remove the LSTM layer of the deep network in Example 3 and use only the CNN layer (such as Figure 6 The dotted box shows the time structure change characteristics, which is called model C. The spectrum energy distribution feature extraction method is the same as that in Example 3. Fig.10 (b) is the distribution histogram of recognition accuracy of different time structure change feature extraction methods. The results show that the CNN-LSTM network can mine the global temporal relationship of the signal that the CNN network may ignore, thereby improving the distinguishability of the model.

[0148] 3) Model A was compared with different spectrum energy distribution feature extraction models (Model D-1: no signal multi-scale decomposition and spectrum energy distribution feature extraction were performed, and only the time structure change features of the three modal signals were used for identification and classification; Model D-2: no signal multi-scale decomposition was performed, the signal spectrum was evenly divided into K frequency bands and the energy of each frequency band was calculated to form the spectrum energy distribution feature; Model D-3: the signal decomposition method was replaced by EMD from VMD, and the spectrum energy distribution features were calculated using the first 10 IMFs, and the calculation steps were consistent with step S3-2). The time structure change feature extraction method and CNN-LSTM network parameters were consistent with those of Example 4 and Example 5. Fig.10 (c) is the histogram of the recognition accuracy of different spectral energy distribution feature extraction models. VMD decomposes complex signals into stationary components of different scales, and refines the lower frequency bands containing the main frequency components of the wheel-rail interaction, thereby revealing the weak disease information hidden between different components. At the same time, it avoids the potential problems of modal aliasing and endpoint effects of traditional decomposition methods such as spectrum equal division and EMD.

[0149] In summary, the track defect identification method proposed in this embodiment has advantages in enriching the network perception information dimension and improving the feature representation ability of the signal in the time scale and frequency domain information. The recognition accuracy rate reaches 98.17%, which proves the efficiency and reliability of the method of this embodiment in the field of track defect detection.

[0150] Example 9 This embodiment provides a track defect identification system based on DAS, including: The tri-modal signal acquisition module is used to perform two differential processes on the collected vibration displacement data to be measured to obtain the vibration velocity data to be measured and the vibration acceleration data to be measured; The spectrum and time feature acquisition module is used to perform signal decomposition on the vibration displacement data to be measured, the vibration velocity data to be measured, and the vibration acceleration data to be measured, respectively, to obtain the energy distribution characteristics of the displacement spectrum, the energy distribution characteristics of the velocity spectrum, and the energy distribution characteristics of the acceleration spectrum; Processing the vibration displacement data to be measured, the vibration velocity data to be measured, and the vibration acceleration data to be measured to obtain the time structure change characteristics of the displacement, the time structure change characteristics of the velocity, and the time structure change characteristics of the acceleration; A feature fusion module is used to fuse the displacement spectrum energy distribution features, the velocity spectrum energy distribution features, the acceleration spectrum energy distribution features, the displacement time structure change features, the velocity time structure change features and the acceleration time structure change features to obtain a fusion feature; The recognition and classification module is used to detect and identify the fused features and obtain the classification results.

[0151] Example 10 A computer device comprises a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of a track defect identification method based on DAS.

[0152] The computer device may be a desktop computer, a notebook, a PDA, a cloud server, etc. The computer device may interact with the user through a keyboard, a mouse, a remote control, a touch pad, or a voice control device.

[0153] The memory includes at least one type of readable storage medium, and the readable storage medium includes flash memory, hard disk, multimedia card, card-type memory (for example, SD or D interface display memory, etc.), random access memory (RAM), static random access memory (SRAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), programmable read-only memory (PROM), magnetic memory, disk, optical disk, etc. In some embodiments, the memory may be an internal storage unit of the computer device, such as a hard disk or memory of the computer device. In other embodiments, the memory may also be an external storage device of the computer device, such as a plug-in hard disk, a smart memory card (Smart Media Card, SMC), a secure digital (Secure Digital, SD) card, a flash card, etc. equipped on the computer device. Of course, the memory may also include both the internal storage unit of the computer device and its external storage device. In this embodiment, the memory is often used to store the operating system and various application software installed on the computer device, such as the program code of the DAS-based track disease identification method, etc. In addition, the memory may also be used to temporarily store various data that have been output or are to be output.

[0154] In some embodiments, the processor may be a central processing unit (CPU), a controller, a microcontroller, a microprocessor, or other data processing chip. The processor is generally used to control the overall operation of the computer device. In this embodiment, the processor is used to run the program code stored in the memory or process data, such as running the program code of the track defect identification method based on the DAS number.

[0155] Embodiment 11 A computer-readable storage medium stores a computer program. When the computer program is executed by a processor, the processor executes the steps of a DAS-based track defect identification method.

[0156] The computer-readable storage medium stores an interface display program, and the interface display program can be executed by at least one processor so that the at least one processor performs the steps of the DAS-based track defect identification method as described above.

[0157] Through the description of the above implementation methods, those skilled in the art can clearly understand that the above-mentioned embodiment method can be implemented by means of software plus a necessary general hardware platform, and of course it can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present application, or the part that contributes to the prior art, can be embodied in the form of a software product, which is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk), and includes a number of instructions for a terminal device (which can be a mobile phone, computer, server or network device, etc.) to execute the DAS-based track defect identification method described in the embodiment of the present application.

Claims

1. A DAS-based track defect detection and identification method, characterized in that: The following steps are involved: Step S1, obtaining a tri-modal signal; Collecting vibration displacement data to be measured, performing two differential processes on the collected vibration displacement data to be measured, and obtaining vibration velocity data to be measured and vibration acceleration data to be measured; Step S2, obtaining spectrum and time characteristics; The vibration displacement data to be measured, the vibration velocity data to be measured, and the vibration acceleration data to be measured are respectively decomposed to obtain the energy distribution characteristics of the displacement spectrum, the energy distribution characteristics of the velocity spectrum, and the energy distribution characteristics of the acceleration spectrum; Processing the vibration displacement data to be measured, the vibration velocity data to be measured, and the vibration acceleration data to be measured to obtain the time structure change characteristics of the displacement, the time structure change characteristics of the velocity, and the time structure change characteristics of the acceleration; Step S3, fusion features; The energy distribution features of the displacement spectrum, the energy distribution features of the velocity spectrum, the energy distribution features of the acceleration spectrum, the time structure change features of the displacement, the time structure change features of the velocity and the time structure change features of the acceleration are fused to obtain fusion features; Step S4, identification and classification; The fused features are detected and identified to obtain the classification results.

2. A DAS-based track defect detection and identification method as claimed in claim 1, characterized in that: In step S1, the specific method of obtaining the tri-modal signal is: Step S1-1: The vehicle generates vibration when passing through the track. The DAS system continuously accumulates the original signal trajectory on the time axis to form a space-time matrix. The spatial channel where the track disease is located is extracted from the space-time matrix to obtain a one-dimensional time series S with a total length of L sampling points, and the one-dimensional time series S is divided into M vibration displacement data with a length of N sampling points along the time axis. The range of the i-th vibration displacement data to be measured in the time series S is: Step S1-2, performing differential differentiation on the segmented vibration displacement data to be measured to obtain vibration velocity data to be measured; performing differential differentiation on the vibration velocity data to be measured to obtain vibration acceleration data to be measured; wherein the calculation formula for differential differentiation is: in, represents the sampling starting point, Indicates the sliding offset, Indicates the data sequence number. Represents the total number of vibration displacement data to be measured, They represent the velocity and acceleration modes respectively. Indicates a point in time.

3. A DAS-based track defect detection and identification method as claimed in claim 1, characterized in that: In step S2, the variational modal decomposition technology is used to perform signal decomposition on the vibration displacement data to be measured, the vibration velocity data to be measured, and the vibration acceleration data to be measured, specifically: Step S2-1, the multimodal signal composed of the vibration displacement data to be measured, the vibration velocity data to be measured, and the vibration acceleration data to be measured is expressed as: The variational mode decomposition technique is used to decompose the multimodal signal into IMF components, so that each mode function The sum of the estimated bandwidths is the smallest, expressed as: Step S2-2, calculate the energy of each IMF component, the calculation formula is: The calculated energy is normalized and the calculation formula is: Step S2-3, construct and obtain the corresponding spectrum energy distribution characteristics, the calculation formula is: in, Indicates time point The total number of Indicates the total number of vibration displacement data to be measured, Indicates time point signal, Indicates the data sequence number. They represent displacement, velocity and acceleration modes respectively. Representing modality The signal is decomposed into Quantity, represents the number of decomposition levels, , Respectively The magnitude and phase of Representing modality The signal is decomposed into The time signal point of each component.

4. A DAS-based track defect identification method as claimed in claim 3, characterized in that: In step S2-1, the variational mode decomposition technique is used to decompose different mode signals respectively, which specifically includes the following steps: Step S2-1-1, set the number of decomposition levels , for each , the Hilbert transform is used to calculate the corresponding analytical signal and obtain the one-sided spectrum. The calculation formula is: Step S2-1-2, for each , through the aliasing of the exponential term corresponding to its center, each The spectrum is modulated to the corresponding "baseband", and the calculation formula is: Step S2-1-3, square norm of the signal gradient of step S2-1-2 Solve and estimate The bandwidth of , thus constructing the constrained variational problem: Step S2-1-4, solve the variational problem of step S2-1-3 and construct the Lagrangian augmented function: Step S2-1-5, using the alternating direction multiplier method to solve the unconstrained variational problem of step S2-1-4, , The optimal solution is obtained through iterative updating. The iterative formula is: Step S2-1-6, Update, the update formula is: Step S2-1-7, given determination accuracy If the convergence condition is met, the iteration stops; otherwise, it returns to step S2-1-5 to continue the iteration; the convergence condition is: Step S2-1-8, the iterative completion of the obtained Perform inverse Fourier transform and get An IMF component with a certain bandwidth and fluctuating around the center frequency; in, represents the impulse function, represents an imaginary number, Indicates a point in time, represents the center frequency corresponding to the IMF component, Decomposed IMF components, represents the center frequency of the IMF component, represents the quadratic penalty term, , All represent Lagrange multipliers, It means to find the derivative of the function with respect to time t. represents a multimodal signal, Indicates frequency, , , Respectively represent the components ,Signal , Lagrange multipliers The Fourier transform of represents the number of iterations, represents the noise margin parameter, Indicates the judgment accuracy.

5. The DAS-based track defect identification method according to claim 1, characterized in that: In step S2, step S3, and step S4, the track defect detection and recognition model is used to obtain the temporal structural change characteristics, perform feature fusion, and perform recognition and classification; The constructed track defect detection and recognition model includes three CNN-LSTM networks, one AFF module and two fully connected layers. When training the constructed track defect detection and recognition model, the specific training steps are as follows: First, the vibration displacement sample data collected by the DAS system is subjected to two differential processes to obtain the vibration velocity sample data and the vibration acceleration sample data; Then, the vibration displacement sample data, the vibration velocity sample data, and the vibration acceleration sample data are respectively input into the corresponding CNN-LSTM networks, and the three CNN-LSTM networks respectively output the vibration displacement sample time structure change characteristics, the vibration velocity sample time structure change characteristics, and the vibration acceleration sample time structure change characteristics; the vibration displacement sample data, the vibration velocity sample data, and the vibration acceleration sample data are respectively subjected to signal decomposition to obtain the vibration displacement sample spectrum energy distribution characteristics, the vibration velocity sample spectrum energy distribution characteristics, and the vibration acceleration sample spectrum energy distribution characteristics; Then, the energy distribution features of the spectrum of vibration displacement samples, the energy distribution features of the spectrum of vibration velocity samples, the energy distribution features of the spectrum of vibration acceleration samples, the time structure change features of vibration displacement samples, the time structure change features of vibration velocity samples, and the time structure change features of vibration acceleration samples are input into the AFF module for fusion to obtain fusion features; Finally, the fusion features are input into the first fully connected layer. The output of the first fully connected layer is used as the input of the second fully connected layer. The second fully connected layer outputs the classification result. The loss is calculated based on the classification result and the corresponding label data, and the parameters of the track disease detection and recognition model are updated.

6. A DAS-based track defect identification method as claimed in claim 5, characterized in that: The CNN-LSTM network includes a CNN subnetwork and an LSTM layer that are set in sequence. The CNN subnetwork includes a convolution layer Conv1, a pooling layer MaxPool1, ReLU1, a convolution layer Conv2, a pooling layer MaxPool2, ReLU2, a convolution layer Conv3, a pooling layer MaxPool3, and ReLU3.

7. A DAS-based track defect identification method as claimed in claim 5, characterized in that: The AFF module includes two convolutional layers. The specific method of feature fusion in the AFF module is as follows: The Xavier initialization method is used to initialize the structural parameters of the convolutional layer of the AFF module; Parallel splicing of the spectrum energy distribution characteristics corresponding to the three-modal signals to obtain the spectrum energy characteristics ; Parallel splicing of the time structure change features corresponding to the three modal signals to obtain the time structure features ; The spectrum energy characteristics , Time structure characteristics Perform convolution operation to obtain spectrum energy characteristics , Time structure characteristics ; Spectral energy characteristics and temporal structural characteristics After adding, local attention features and global attention features are obtained respectively; The local attention feature and the global attention feature are added together and the result is obtained through the Sigmoid function. , and according to the results Get fusion features .

8. A DAS-based track defect detection and identification system, characterized in that: include: The tri-modal signal acquisition module is used to perform two differential processes on the collected vibration displacement data to be measured to obtain the vibration velocity data to be measured and the vibration acceleration data to be measured; The spectrum and time feature acquisition module is used to perform signal decomposition on the vibration displacement data to be measured, the vibration velocity data to be measured, and the vibration acceleration data to be measured, respectively, to obtain the energy distribution characteristics of the displacement spectrum, the energy distribution characteristics of the velocity spectrum, and the energy distribution characteristics of the acceleration spectrum; Processing the vibration displacement data to be measured, the vibration velocity data to be measured, and the vibration acceleration data to be measured to obtain the time structure change characteristics of the displacement, the time structure change characteristics of the velocity, and the time structure change characteristics of the acceleration; A feature fusion module is used to fuse the displacement spectrum energy distribution features, the velocity spectrum energy distribution features, the acceleration spectrum energy distribution features, the displacement time structure change features, the velocity time structure change features and the acceleration time structure change features to obtain a fusion feature; The recognition and classification module is used to detect and identify the fused features and obtain the classification results.

9. A computer device, characterized in that: The method comprises a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the method according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that: A computer program is stored, and when the computer program is executed by a processor, the processor is caused to perform the steps of the method according to any one of claims 1 to 7.

Images