DAS-based methods, systems, equipment, and media for detecting and identifying track defects.

Abstract

This invention discloses a method, system, device, and medium for track defect detection and identification based on DAS (Digital Optical Array Sensor), belonging to the field of track defect detection in the field of fiber optic sensing technology. Its purpose is to solve the problem of low accuracy in track defect detection and identification in existing technologies due to limited vibration displacement information. This identification method acquires three-mode signals including vibration displacement, vibration velocity, and vibration acceleration, calculates the spectral and temporal characteristics of each mode signal, and fuses the spectral and temporal characteristics of the three modes. Detection and identification are then performed based on the fused characteristics to obtain a classification result. This method fully utilizes the rich sensing information in the velocity and acceleration modes, and fuses the corresponding spectral and temporal characteristics of the three modes for track defect identification and classification. It has stronger track defect identification and classification capabilities, contributing to accurate detection of track defects and precise identification of defect types.

Description

Technical Field

[0001] This invention belongs to the field of fiber optic sensing technology, and relates to a track defect detection technology, and more particularly to a track defect detection and identification method, system, equipment and medium based on DAS. Background Technology

[0002] Railway transportation is a core component of my country's transportation system. With increasing traffic density, the load between railway tracks and train wheels is gradually increasing, leading to track defects such as track surface contact fatigue (e.g., fish-scale scratches), wavy wear on the rail top (corrugation), and gaps between sleepers and ballast (whitening of the ballast) caused by severe corrugation. These defects have become major hidden dangers 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 lighting and weather. Furthermore, current detection methods can only be used during maintenance windows and cannot achieve 24 / 7 online monitoring during train operation.

[0003] Distributed acoustic sensing (DAS) systems based on the phase-sensitive optical time-domain reflectometry (Φ-OTDR) principle utilize widely deployed optical fiber cables to construct a large-scale, high-density, and low-cost sensing network, showing broad application prospects in the field of long-distance vibration event monitoring. When applying DAS systems to railway track defect monitoring, the existing optical fiber cables along the railway line can be used as sensing fibers. When a train passes over the track, the interaction between the wheels and the track generates vibrations and excites sound waves. The optical fiber cables along the railway line sense these sound wave signals and collect, analyze, and process them through the DAS system. Since the vibration and sound wave characteristics generated when different track defects interact with trains vary, by analyzing these differences in vibration signals and combining them with intelligent signal processing and recognition algorithms, real-time monitoring of railway track defects along the optical fiber can be achieved, thus providing an all-weather, full-line online monitoring method for railway transportation safety.

[0004] Patent application number 202310249537.0 discloses a method for identifying railway track defects based on fiber optic distributed vibration detection, which includes the following steps: Step 1, transmitting detection light pulses to the sensing fiber through a fiber optic distributed acoustic wave sensing system (DAS) to achieve real-time quantitative monitoring of vibration information along the sensing fiber and acquire the raw vibration signal; Step 2, constructing an initial dataset from the raw vibration signal acquired in Step 1, and labeling the data samples of normal track signals and three types of defect signals; Step 3, selecting M groups each of normal track signals and three types of defect signals from the initial multidimensional fusion feature vector dataset obtained in Step 2. The data is used to construct a multidimensional fusion feature vector dataset. The average standard deviation (ASD) of the multidimensional fusion feature vectors of normal rail signals and three types of defect signals is used as the weight of each feature value of the multidimensional fusion feature vector. The multidimensional fusion feature vectors are pre-weighted to obtain a pre-weighted multidimensional fusion feature vector sample set. Step 4: The pre-weighted multidimensional fusion feature vector sample set obtained in step 3 is input into the support vector machine (SVM) model for classification training. The trained SVM model is used to classify and identify normal rail signals and three types of defect signals, record the location and type of defects, output them, and issue an alarm to achieve real-time defect diagnosis and location output.

[0005] Furthermore, invention patent application number 2024105324443 also discloses an artificial intelligence-based method for diagnosing equipment vibration faults, which includes the following steps: S1: Real-time collection of equipment vibration signals; S2: Preprocessing of the collected vibration signals; S3: Extracting features from the preprocessed signals, and then merging the extracted features into a comprehensive feature vector; S4: Constructing a deep neural network model; S5: When the model identifies an abnormal vibration pattern, an early warning mechanism is immediately triggered; S6: After determining that the equipment has a fault, a diagnostic report is automatically generated based on the analysis results of the deep learning model. By combining advanced artificial intelligence technologies, especially Convolutional Neural Networks (CNN) and Long Short-Term Memory Networks (LSTM), this method 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 enables precise identification of fault types and their causes. Compared with traditional experience-based fault detection methods, it greatly enhances the reliability and effectiveness of fault diagnosis.

[0006] The aforementioned patent application for a railway track defect identification method discloses the use of a DAS system to acquire raw vibration signals and classify and identify normal track signals and three types of defect signals. The aforementioned patent application for a device vibration fault diagnosis method discloses a composite model combining CNN and LSTM to identify vibration signal patterns. However, since optical fibers are generally laid in cable trenches or soil beside the track and are not in direct contact with the railway track, the vibration signals generated by wheel-rail interaction are indirectly transmitted to the optical fiber through the roadbed and soil, constituting indirect measurement. During underground propagation, the vibration signals are inevitably affected by soil particles and soil structure, causing signal attenuation and filtering, reducing the coupling efficiency between wheel-rail vibration and the optical fiber. The useful component of the wheel-rail vibration signal sensed by the DAS system is further reduced compared to the original wheel-rail vibration signal. Therefore, the directly acquired optical fiber vibration signal contains limited wheel-rail interaction information, making it difficult to extract wheel-rail fault characteristics and accurately identify faults, ultimately reducing the accuracy of track defect detection and identification. Furthermore, when the DAS system acquires the original vibration signal, it mostly acquires parameters such as the number of pulses and vibration displacement. However, the vibration displacement signal contains limited information. When the subsequent model identifies and classifies the vibration displacement time-frequency diagram, the model's identification and classification capabilities are limited, resulting in low accuracy in detecting and identifying track defects. Summary of the Invention

[0007] The purpose of this invention is to address the problem of low accuracy in track defect detection and identification due to limited vibration displacement information in existing technologies. This invention provides a track defect identification method, system, device, and medium based on DAS (Discrete Aspect-Oriented System). The invention further differentiates vibration displacement into vibration velocity and vibration acceleration, forming three modal signals: vibration displacement, vibration velocity, and vibration acceleration. The resulting spectral energy distribution characteristics, processed by Variational Mode Decomposition (VMD) technology, are fused with the temporal structure change characteristics output by the model before being input into the fully connected layer of the model for final identification and classification. This allows the model to more easily learn richer information from the velocity and acceleration time-frequency plots, resulting in stronger identification and classification capabilities, accurate detection of track defects, and precise identification of defect types, leading to high accuracy in defect detection.

[0008] To achieve the above objectives, the present invention specifically adopts the following technical solution:

[0009] A DAS-based method for identifying track defects includes the following steps:

[0010] Step S1: Acquire the three-mode signal;

[0011] Vibration displacement sample data was collected, and after two differential processing steps, vibration velocity sample data and vibration acceleration sample data were obtained. The vibration displacement sample data, vibration velocity sample data, and vibration acceleration sample data were then labeled to obtain label data.

[0012] Step S2: Obtain the spectral and temporal characteristics;

[0013] Signal decomposition was performed on the vibration displacement data, vibration velocity data, and vibration acceleration data to be measured to obtain the energy distribution characteristics of the displacement spectrum, velocity spectrum, and acceleration spectrum.

[0014] The vibration displacement data, vibration velocity data, and vibration acceleration data to be measured are processed to obtain the displacement time structure change characteristics, velocity time structure change characteristics, and acceleration time structure change characteristics.

[0015] Step S3, feature fusion;

[0016] The energy distribution characteristics of displacement spectrum, energy distribution characteristics of velocity spectrum, energy distribution characteristics of acceleration spectrum, and time structure variation characteristics of displacement, time structure variation characteristics of velocity, and time structure variation characteristics of acceleration are fused to obtain the fused characteristics.

[0017] Step S4, Identify and classify;

[0018] The fusion features are detected and identified to obtain the classification results.

[0019] Furthermore, in step S1, the specific method for obtaining vibration velocity sample data and vibration acceleration sample data is as follows:

[0020] Step S1-1: The vehicle vibrates as it passes over the track. The DAS system continuously accumulates the original signal trajectory on the time axis, forming a spatiotemporal matrix. The spatial channel where the track defect is located is extracted from the spatiotemporal matrix, resulting in a one-dimensional time series S with a total length of L sampling points. This S is then slidably divided along the time axis into M vibration displacement sample data points of length N. The range of the i-th sample in the time series S is:

[0021]

[0022] Step S1-2: Perform differential calculation on the segmented vibration displacement sample data to obtain vibration velocity sample data; perform differential calculation on the vibration velocity sample data to obtain vibration acceleration sample data; wherein, the formula for calculating the differential is:

[0023]

[0024] in, Indicates the sampling start point. Indicates the sliding offset. Indicates the sample number. Represents the total number of samples. These represent the velocity and acceleration modes, respectively. Indicates a point in time.

[0025] Further, in step S2, variational mode decomposition (VMD) is used to decompose the vibration displacement sample data, vibration velocity sample data, and vibration acceleration sample data into signals, and to calculate the spectral energy distribution characteristics, specifically:

[0026] Step S2-1, the multimodal signal composed of vibration displacement sample data, vibration velocity sample data, and vibration acceleration sample data is represented as follows:

[0027]

[0028] Variational mode decomposition (VMD) is used to decompose multimode signals into... Each IMF component makes each mode function The sum of the estimated bandwidths is minimized, expressed as:

[0029]

[0030]

[0031] Step S2-2, calculate the energy of each IMF component, using the following formula:

[0032]

[0033] The calculated energy is then normalized using the following formula:

[0034]

[0035] Step S2-3: Construct and obtain the corresponding spectral energy distribution characteristics. The calculation formula is as follows:

[0036]

[0037] in, Indicates a point in time The total number, Represents the total number of samples. Indicates a point in time The signal Indicates the sample number. These represent the three modes: displacement, velocity, and acceleration. Representing modes The signal after decomposition One portion, Indicates the number of decomposition layers. , They represent The amplitude and phase, Representing modes The signal after decomposition Each component is a time signal point.

[0038] Furthermore, in step S2-1, variational mode decomposition (VMD) is used to decompose the signals of different modes separately, specifically including the following steps:

[0039] Step S2-1-1: Set the number of decomposition layers For each The Hilbert transform is used to calculate the corresponding analytic signal, and the one-sided spectrum is obtained. The calculation formula is as follows:

[0040]

[0041] Step S2-1-2, for each By aliasing the exponential terms of their corresponding centers, each The spectrum is modulated to the corresponding "baseband", and the calculation formula is:

[0042]

[0043] Step S2-1-3, the square norm of the signal gradient in step S2-1-2. Solve and estimate Given the bandwidth, a constrained variational problem is constructed:

[0044]

[0045]

[0046]

[0047] Step S2-1-4: Solve the variational problem of step S2-1-3 and construct the Lagrange augmented function:

[0048]

[0049] Step S2-1-5: Solve the unconstrained variational problem from step S2-1-4 using the alternating direction multiplier method. , The optimal solution is obtained through iterative updates, and the iterative formula is:

[0050]

[0051]

[0052] Step S2-1-6, for The update is performed using the following formula:

[0053]

[0054] Step S2-1-7, given the determination precision If the convergence condition is met, the iteration stops; otherwise, return to step S2-1-5 to continue iteration. The convergence condition is:

[0055]

[0056] Step S2-1-8, after the iteration is completed Perform an inverse Fourier transform to obtain An IMF component with a certain bandwidth that fluctuates around the center frequency.

[0057] Furthermore, in steps S2, S3, and S4, the track defect detection and identification model is used to obtain time structure change features, perform feature fusion, and identify and classify them.

[0058] The constructed track defect detection and identification model includes three CNN-LSTM networks, one AFF module, and two fully connected layers. The specific training steps for training this model are as follows:

[0059] First, the vibration displacement sample data collected by the DAS system is processed twice to obtain vibration velocity sample data and vibration acceleration sample data;

[0060] The vibration displacement sample data, vibration velocity sample data, and vibration acceleration sample data are then input into the corresponding CNN-LSTM networks. The three CNN-LSTM networks output the temporal structure change features of the vibration displacement sample, the vibration velocity sample, and the vibration acceleration sample, respectively. The vibration displacement sample data, vibration velocity sample data, and vibration acceleration sample data are then decomposed to obtain the spectral energy distribution features of the vibration displacement sample, the vibration velocity sample, and the vibration acceleration sample.

[0061] Then, the spectral energy distribution characteristics of vibration displacement samples, the spectral energy distribution characteristics of vibration velocity samples, the spectral energy distribution characteristics of vibration acceleration samples, the temporal structure change characteristics of vibration displacement samples, the temporal structure change characteristics of vibration velocity samples, and the temporal structure change characteristics of vibration acceleration samples are input into the AFF module for fusion to obtain the fused characteristics.

[0062] Finally, the features are fused into the first fully connected layer, and 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, calculates the loss based on the classification result and the corresponding label data, and updates the parameters of the track defect detection and identification model.

[0063] Further, in step S2, a CNN-LSTM network is used to extract the temporal structure change features of the sample data, including a CNN sub-network and an LSTM layer. The CNN sub-network includes a convolutional layer Conv1, a pooling layer MaxPool1, a ReLU1, a convolutional layer Conv2, a pooling layer MaxPool2, a ReLU2, a convolutional layer Conv3, a pooling layer MaxPool3, and a ReLU3.

[0064] Furthermore, the AFF module includes two convolutional layers, and the specific method for feature fusion in the AFF module is as follows:

[0065] The Xavier initialization method is used to initialize the structural parameters of the convolutional layers in the AFF module;

[0066] The spectral energy distribution characteristics of the three-mode signals are obtained by concatenating them in parallel. The temporal structure features are obtained by concatenating the temporal structure variation characteristics of the three-mode signals in parallel. ;

[0067] Spectral energy characteristics respectively Time structure characteristics Perform convolution operations to obtain spectral energy features. Time structure characteristics ;

[0068] Spectral energy characteristics and time structure features After summing, the local attention features and global attention features are calculated separately.

[0069] The result is obtained by adding the local attention features and the global attention features and then applying the Sigmoid function. And based on the results Obtain fusion features .

[0070] Furthermore, in step S4, after the fused features pass through two fully connected layers, the SoftMax method is used to obtain the final classification result. The time 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 obtain the feature fusion recognition model.

[0071] The design process for a track defect detection and identification model includes three stages: network structure design and parameter setting, network initialization and training, and parameter updating and optimization. The specific methods are as follows:

[0072] Step S4-1, Network structure design and parameter settings;

[0073] After extracting the temporal structure variation features and spectral energy distribution features of the three modal signals, they are simultaneously input into the attention module for fusion. The output of the AFF module is then passed through two fully connected layers and the SoftMax method is used to obtain the final classification result.

[0074] Step S4-2; Network initialization and network training;

[0075] After parameter initialization, the track defect detection and identification model obtains the predicted class probability distribution through forward propagation. The cross-entropy loss function is then used to calculate the loss between the class probabilities output by the fully connected layer and the true probabilities. This loss value is then used for backpropagation to calculate the gradient of each learned parameter. Finally, the model parameters are updated according to the specified learning rate and the principle of gradient descent. The model parameters Including matrix weights and bias Let's take the first iteration of the model's learning process as an example:

[0076] Initialize the parameters of the attention-based fusion network. Good initialization parameters make the model easier to learn and converge faster. This invention uses the Xavier initialization method to initialize the network structure parameters. To ensure that the variance of each layer is consistent during forward and backward propagation, the distribution range of the randomly initialized parameters is the number of input parameters passing through that layer. Number of output parameters The uniform distribution within the obtained distribution range is given by the formula for the distribution range:

[0077]

[0078] Step S4-3; Parameter update and optimization;

[0079] The backpropagation error is calculated based on the set objective function and the obtained predicted labels. This error is then used to update and optimize the parameters of the constructed attention-based fusion network.

[0080] Furthermore, in step S4-3, the specific steps are as follows:

[0081] Step S4-3-1 Fully Connected Layer FC Receiver Fusion Features Output predicted labels The calculation formula is:

[0082]

[0083] in, , , These represent the weight matrix, bias vector, and activation function, respectively. This indicates the predicted label.

[0084] Step S4-3-2: Calculate the loss value between the predicted label and the true label using the cross-information entropy loss function: Calculate the distance between the predicted event label and the true label using the cross-information entropy loss function, and the loss value... The calculation formula is as follows:

[0085]

[0086] in: M Represents the total number of samples. This indicates the actual label.

[0087] Step S4-3-3 uses the loss value to back-calculate the parameter gradient of the attention-based fusion network model, and uses the parameter gradient to update the attention-based fusion network deep learning model; this invention uses the Adam algorithm for update optimization, and the calculation steps are as follows:

[0088]

[0089]

[0090]

[0091] in: To calculate the target gradient, , These are the first and second moment estimates of the gradient, respectively. , which are the exponential decay rates of the first and second moments of the gradient, respectively.

[0092] Let the t-th iteration be... , , The network parameter update formula is:

[0093]

[0094] in: For learning rate, For very small numbers, prevent division by zero.

[0095] A DAS-based track defect identification system includes:

[0096] The three-mode signal acquisition module is used to perform two differential processing on the collected vibration displacement data to obtain the vibration velocity data and vibration acceleration data.

[0097] The spectrum and time feature acquisition module is used to perform signal decomposition on the vibration displacement data, vibration velocity data, and vibration acceleration data to be measured, respectively, to obtain the displacement spectrum energy distribution characteristics, velocity spectrum energy distribution characteristics, and acceleration spectrum energy distribution characteristics.

[0098] The vibration displacement data, vibration velocity data, and vibration acceleration data to be measured are processed to obtain the displacement time structure change characteristics, velocity time structure change characteristics, and acceleration time structure change characteristics.

[0099] The feature fusion module is used to fuse the energy distribution features of displacement spectrum, energy distribution features of velocity spectrum, energy distribution features of acceleration spectrum, time structure variation features of displacement, time structure variation features of velocity, and time structure variation features of acceleration to obtain fused features.

[0100] The identification and classification module is used to detect and identify the fusion features of the signal under test and obtain the classification result.

[0101] A computer device includes a memory and a processor, the memory storing a computer program that, when executed by the processor, causes the processor to perform the steps of the method described above.

[0102] A computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the steps of the above-described method.

[0103] The beneficial effects of this invention are as follows:

[0104] 1. This invention proposes a multimodal vibration information extraction method for the first time, specifically for the application of DAS in railway track defect detection, and applies it to the early detection of track defects and the high-precision identification of defect types. Compared with existing track defect detection technologies, DAS has advantages such as low cost, long distance, high precision, and all-weather capability. Furthermore, research has shown that vibration velocity and vibration acceleration possess richer frequency domain information than vibration displacement. Therefore, this application transforms the use of a single vibration displacement signal into the simultaneous use of displacement, velocity, and acceleration modal signals, effectively enriching the dimensions of the perceived information in the identification model.

[0105] 2. In this invention, VMD is used to replace traditional signal decomposition methods such as spectrum equal division and empirical mode decomposition (EMD), avoiding problems such as mode aliasing and endpoint effects. This provides richer and more accurate frequency domain information for the analysis of vibration signals, effectively improving the frequency domain information representation capability of vibration modal signals in the low-frequency band. A CNN-LSTM combined network is used to extract the depth features of different modal vibration signals, effectively representing the multi-scale temporal information of different disease signals. This network extracts the local structural features of the signal through CNN, while using LSTM to capture the long-term interdependencies of the signal and mine the temporal relationship between short-term signals, significantly enhancing the feature representation capability of disease signals on the time scale.

[0106] 3. In this invention, the attention module is used to fuse the energy distribution characteristics of displacement spectrum, velocity spectrum, acceleration spectrum, 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 features, and achieving accurate detection of disease signals and fine identification of disease types. Attached Figure Description

[0107] Figure 1 A flowchart of the track defect detection and identification method based on DAS multimodal vibration information extraction provided by the present invention;

[0108] Figure 2 This invention describes the structure and working principle of the DAS system.

[0109] Figure 3 This is a schematic diagram of the three modal signals in this invention;

[0110] Figure 4 This is a schematic diagram of the spectrum of the signal VMD decomposition in this invention;

[0111] Figure 5 This is a schematic diagram illustrating the spectral energy distribution characteristics of different disease types in this invention;

[0112] Figure 6 This is the CNN-LSTM network structure in this invention;

[0113] Figure 7 This is the structure of the AFF module in this invention;

[0114] Figure 8 This is the confusion matrix of the identification model of the present invention on the test set of the track defect signal dataset;

[0115] Figure 9 This is a structural diagram of the single-modal input signal recognition model in this invention;

[0116] Figure 10This is a comparison of the accuracy of different models of the present invention on the test set of the track defect signal dataset. Detailed Implementation

[0117] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments.

[0118] Therefore, all other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0119] Example 1

[0120] This embodiment provides a track defect identification method based on DAS (Digital Optical Array). It employs DAS multimodal vibration information extraction. The DAS system utilizes existing communication optical fibers near the train track as the sensing medium, eliminating the need for additional fiber optic cables. The entire signal processing flow is as follows: Figure 1 As shown, the specific steps include:

[0121] Step S1: Acquire the three-mode signal;

[0122] Collect the vibration displacement data to be measured, and perform two differential processing on the collected vibration displacement data to obtain the vibration velocity data and vibration acceleration data to be measured.

[0123] The effective signals of wheel-rail interaction when a train passes are collected using the DAS system and communication optical fibers laid along the track. Three types of defects (including corrugation, empty hanging whitening, and fish scale damage) and one normal type are selected as event labels. The one-dimensional time-series signals at typical defect spatial points are segmented by sliding segmentation, and after two differential operations, three modal signal datasets of vibration displacement, vibration velocity, and vibration acceleration, as well as corresponding label data, are constructed.

[0124] Step S2: Obtain the spectral and temporal characteristics;

[0125] The vibration displacement data, vibration velocity data, and vibration acceleration data to be measured are decomposed into signals to obtain stationary IMF components at different scales, and the energy distribution characteristics of the displacement spectrum, velocity spectrum, and acceleration spectrum are calculated.

[0126] The vibration displacement data, vibration velocity data, and vibration acceleration data to be measured are processed and passed through three parallel CNN-LSTM networks to obtain the displacement time structure change characteristics, velocity time structure change characteristics, and acceleration time structure change characteristics.

[0127] Step S3, feature fusion;

[0128] The energy distribution characteristics of displacement spectrum, energy distribution characteristics of velocity spectrum, energy distribution characteristics of acceleration spectrum, and time structure variation characteristics of displacement, time structure variation characteristics of velocity, and time structure variation characteristics of acceleration are fused to obtain the fused characteristics.

[0129] Step S4, Identify and classify;

[0130] The fusion features are detected and identified to obtain the classification results.

[0131] Example 2

[0132] This embodiment is based on Embodiment 1, and further optimizes the present invention.

[0133] The structure of the DAS system used for signal acquisition is as follows: Figure 2 As shown, the system 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 acquisition unit, and a signal processing host. The narrow-bandwidth laser generates continuous light, which is modulated into pulsed light with a specific period and pulse width by the acousto-optic coupler. The erbium-doped fiber amplifier amplifies the pulsed light, and the amplified pulsed light enters the sensing fiber through the circulator. The pulsed light propagates in the sensing fiber and generates backscattered Rayleigh light. The backscattered Rayleigh light is linearly modulated by the coupler in the phase demodulation device and then converted into the original signal trajectory distributed along the space by the data acquisition unit.

[0134] The specific method for acquiring signals is as follows:

[0135] Step S1-1: The vehicle vibrates as it passes over the track. The DAS system continuously accumulates the original signal trajectory on the time axis, forming a spatiotemporal matrix. The spatial channel where the track defect is located is extracted from the spatiotemporal matrix, resulting in a one-dimensional time series S with a total length of L sampling points. This S is then slidably divided along the time axis into M vibration displacement sample data points of length N. The range of the i-th sample in the time series S is:

[0136]

[0137] Step S1-2: Perform differential calculation on the segmented vibration displacement sample data to obtain vibration velocity sample data; perform differential calculation on the vibration velocity sample data to obtain vibration acceleration sample data; wherein, the formula for calculating the differential is:

[0138]

[0139] in, Indicates the sampling start point. Indicates the sliding offset. Indicates the sample number. Represents the total number of samples. These represent the velocity and acceleration modes, respectively. Indicates a point in time. In this embodiment, the sampling frequency... Set to 1kHz, sample length N is set to 5000, and sliding offset is set to... Set it to 1000.

[0140] Typical signal sample diagrams for the three modes are shown below. Figure 3 As shown, the four columns correspond to four types of track defects (normal, corrugated, bleached, and fish-scale damage), and the four rows correspond to the original signal time-domain waveform, displacement mode time-frequency diagram, velocity mode time-frequency diagram, and acceleration mode time-frequency diagram, respectively. As shown, the main frequency components of the displacement mode signal are concentrated below 100Hz, and it is difficult to distinguish between the four types of samples. The velocity and acceleration modes of different signal types show significant differences in spectral distribution and energy intensity. The low-frequency (below 200Hz) frequency of the normal type velocity mode signal is stronger than the high-frequency (300-500Hz), while the acceleration mode signal has a uniform spectral distribution concentrated between 1-3s. The velocity mode signals of corrugated and fish-scale damage types have a spectral distribution concentrated below 200Hz, while the acceleration mode signals are distributed between 0-200Hz and 400-500Hz, with the fish-scale damage type signal having significantly stronger energy than the corrugated type. The frequency components of the velocity and acceleration modes of the bleached type are evenly distributed across the entire frequency range (0-500Hz). Applying signals of multiple modalities simultaneously to the identification of track defect types effectively enriches the perception capabilities of the identification model and improves its discriminative power.

[0141] Typical track defect labels include three defect types (wave-like wear, empty hanging whitening, 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 7:3 ratio.

[0142] Example 3

[0143] This embodiment is based on Embodiment 2, and further optimizes the present invention.

[0144] For the aforementioned multimodal dataset, step S2 employs Variational Mode Decomposition (VMD) to decompose the vibration displacement sample data, vibration velocity sample data, and vibration acceleration sample data into K stationary IMF components at different scales, and calculates the high-resolution spectral energy distribution characteristics of the signal. The specific steps are as follows:

[0145] Step S2-1, the multimodal signal composed of the measured vibration displacement data, measured vibration velocity data, and measured vibration acceleration data is represented as follows:

[0146]

[0147] Variational mode decomposition (VMD) is used to decompose a multimode signal into K IMF components, such that each mode function... The sum of the estimated bandwidths is minimized, expressed as:

[0148]

[0149]

[0150] Step S2-2, calculate the energy of each IMF component, using the following formula:

[0151]

[0152] The calculated energy is then normalized using the following formula:

[0153]

[0154] Step S2-3: Construct and obtain the corresponding spectral energy distribution characteristics. The calculation formula is as follows:

[0155]

[0156] in, Indicates a point in time The total number, This indicates the total number of vibration displacement data to be measured. Indicates a point in time The signal Indicates the data sequence number. These represent the three modes: displacement, velocity, and acceleration. Representing modes The signal after decomposition One portion, Indicates the number of decomposition layers. , They represent The amplitude and phase, Representing modes The signal after decomposition Each component is a time signal point.

[0157] The core of this embodiment lies in using VMD to decompose the signal into... The basic idea behind the IMF component is to construct and solve variational problems, which effectively addresses the issues of mode overlap and endpoint effects in signal decomposition. Variational mode decomposition (VMD) is used to decompose different modal signals separately, specifically including the following steps:

[0158] Step S2-1-1: Set the number of decomposition layers For each The Hilbert transform is used to calculate the corresponding analytic signal, and the one-sided spectrum is obtained. The calculation formula is as follows:

[0159]

[0160] Step S2-1-2, for each By aliasing the exponential terms of their corresponding centers, each The spectrum is modulated to the corresponding "baseband", and the calculation formula is:

[0161]

[0162] Step S2-1-3, the square norm of the signal gradient in step S2-1-2. Solve and estimate Given the bandwidth, a constrained variational problem is constructed:

[0163]

[0164]

[0165]

[0166] Step S2-1-4: Solve the variational problem of step S2-1-3 and construct the Lagrange augmented function:

[0167]

[0168] Step S2-1-5: Solve the unconstrained variational problem from step S2-1-4 using the alternating direction multiplier method. , The optimal solution is obtained through iterative updates, and the iterative formula is:

[0169]

[0170]

[0171] Step S2-1-6, for The update is performed using the following formula:

[0172]

[0173] Step S2-1-7, given the determination precision If the convergence condition is met, the iteration stops; otherwise, return to step S2-1-5 to continue iteration. The convergence condition is:

[0174]

[0175] Step S2-1-8, after the iteration is completed Perform an inverse Fourier transform to obtain An IMF component with a certain bandwidth that fluctuates around the center frequency;

[0176] in, Represents the impulse function. represents an imaginary number, Indicates a point in time. This indicates the center frequency corresponding to the IMF component. This represents the K IMF components obtained from the decomposition. Indicates the center frequency of the IMF component. This indicates a secondary penalty item. , Both represent Lagrange multipliers. This represents taking the derivative of the function with respect to time t. Represents multimodal signals. Indicates frequency, , , Representing components respectively ,Signal Lagrange multipliers Fourier transform, Indicates the number of iterations. This represents the noise tolerance parameter. Indicates the precision of the judgment.

[0177] The number of VMD decompositions is set to The spectrum diagram of each IMF component after decomposition is shown below. Figure 4 As shown, 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, with signal energy concentrated in components IMF6-IMF12. The energy characteristics of signals from different damage types are as follows: Figure 5 As shown, the signal energy of the normal type and the empty-hanging white type is mainly distributed in IMF9-IMF12. The energy characteristic values ​​of each IMF are similar, but the normal type is stronger than the empty-hanging white type. The signal energy of the wave-patterned 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 effects, and effectively uncovers weak defect information hidden in the complex structure of the signal. Using spectral energy distribution characteristics for characterization and quantification provides richer and more accurate frequency domain information for vibration signal analysis, effectively improving the frequency domain information representation capability of vibration modal signals in the low-frequency band.

[0178] Example 4

[0179] This embodiment is based on Embodiment 2, and further optimizes the present invention.

[0180] In this invention, steps S2, S3, and S4 all use a track defect detection and identification model to obtain time structure change features, perform feature fusion, and identify and classify them.

[0181] The constructed track defect detection and identification model includes three CNN-LSTM networks, one AFF module, and two fully connected layers. The specific training steps for training this model are as follows:

[0182] First, the vibration displacement sample data collected by the DAS system is processed twice to obtain vibration velocity sample data and vibration acceleration sample data;

[0183] The vibration displacement sample data, vibration velocity sample data, and vibration acceleration sample data are then input into the corresponding CNN-LSTM networks. The three CNN-LSTM networks output the temporal structure change features of the vibration displacement sample, the vibration velocity sample, and the vibration acceleration sample, respectively. The vibration displacement sample data, vibration velocity sample data, and vibration acceleration sample data are then decomposed to obtain the spectral energy distribution features of the vibration displacement sample, the vibration velocity sample, and the vibration acceleration sample.

[0184] Then, the spectral energy distribution characteristics of vibration displacement samples, the spectral energy distribution characteristics of vibration velocity samples, the spectral energy distribution characteristics of vibration acceleration samples, the temporal structure change characteristics of vibration displacement samples, the temporal structure change characteristics of vibration velocity samples, and the temporal structure change characteristics of vibration acceleration samples are input into the AFF module for fusion to obtain the fused characteristics.

[0185] Finally, the features are fused into the first fully connected layer, and 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, calculates the loss based on the classification result and the corresponding label data, and updates the parameters of the track defect detection and identification model.

[0186] Example 5

[0187] This embodiment is based on Embodiment 4, and further optimizes the present invention.

[0188] In this invention, three parallel CNN-LSTM networks are constructed, and signals of vibration displacement, vibration velocity, and vibration acceleration are used as inputs to the three CNN-LSTM networks respectively. That is, a signal of one mode corresponds to one CNN-LSTM network. The signals are sequentially passed through three CNN-LSTM networks with identical structures, sharing the network loss. The CNN-LSTM networks output the temporal structure variation characteristics of the signal of that mode. .

[0189] The specific structure of this CNN-LSTM network is as follows: Figure 6 As shown, the network includes a CNN subnetwork and an LSTM layer. The CNN subnetwork includes a convolutional layer Conv1, a pooling layer MaxPool1, a ReLU1, a convolutional layer Conv2, a pooling layer MaxPool2, a ReLU2, a convolutional layer Conv3, a pooling layer MaxPool3, and a ReLU3. The specific network structure parameters of the CNN-LSTM network are as follows: 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.

[0190] Example 6

[0191] This embodiment is based on Embodiments 3, 4 and 5, and further optimizes the present invention.

[0192] Based on the spectral energy distribution features extracted in Example 3 and the temporal structure change features extracted in Example 5, and by fusing the temporal structure change features and spectral energy distribution features of the displacement, velocity, and acceleration modal signals using the Attention Filter (AFF) module, the specific steps are as follows:

[0193] The spectral energy distribution characteristics of three modal signals spliced ​​in parallel are denoted as...

[0194]

[0195] The temporal structure variation characteristics of three modal signals spliced ​​in parallel are denoted as...

[0196]

[0197] in These correspond to displacement, velocity, and acceleration modes, respectively. and Perform a corresponding convolution operation on each, making them have the same dimension. Let the spectral energy distribution characteristics after the convolution operation be... The characteristics of time structure change are Attention module, AFF module, etc. Figure 7 As shown, and After summing, calculate the local attention features and the global attention features separately; let the input features be... Then the corresponding formula for calculating the local attention features is:

[0198]

[0199] The formula for calculating the features of global attention is:

[0200]

[0201] in, This is a feature of local attention. For global attention features, This is a convolution operation, with a convolution size of 1×1. This indicates the BatchNorm layer. Represents the ReLU activation function. This represents the average pooling operation. Therefore, based on... and Calculate the local attention features and global attention features, add them together, and then apply the result to the Sigmoid function. for:

[0202]

[0203] The fused features after the attention fusion module are denoted as follows: The results are as follows:

[0204]

[0205] The network model parameters of the AFF module are set to: 1*5 / 1*1 / 0, 1*5 / 1*1 / 0 (kernel / stride / padding).

[0206] Example 7

[0207] This embodiment is based on Embodiment 6, and further optimizes the present invention.

[0208] In step S4, the fused features are processed through two fully connected layers and the SoftMax method is used to obtain the final classification result. The time structure change feature extraction module, feature fusion module and recognition module are trained offline based on the same objective function to obtain the optimal model and obtain the feature fusion recognition model.

[0209] The design process for a track defect detection and identification model includes three stages: network structure design and parameter setting, network initialization and training, and parameter updating and optimization. The specific methods are as follows:

[0210] Step S4-1, Network structure design and parameter settings;

[0211] After extracting the temporal structure variation features and spectral energy distribution features of the three modal signals, they are simultaneously input into the attention module for fusion. The output of the AFF module is then passed through two fully connected layers and the SoftMax method is used to obtain the final classification result. The parameters of the fully connected layers are set as follows: Linear / 32*64 / 1*128, Linear / 1*128 / 1*4 (activation function / input dimension / output dimension).

[0212] Step S4-2, network initialization and network training;

[0213] After parameter initialization, the track defect detection and identification model obtains the predicted class probability distribution through forward propagation. The cross-entropy loss function is then used to calculate the loss between the class probabilities output by the fully connected layer and the true probabilities. This loss value is then used for backpropagation to calculate the gradient of each learned parameter. Finally, the model parameters are updated according to the specified learning rate and the principle of gradient descent. The model parameters Including matrix weights and bias Let's take the first iteration of the model's learning process as an example:

[0214] Initialize the parameters of the attention-based fusion network. Good initialization parameters make the model easier to learn and converge faster. This invention uses the Xavier initialization method to initialize the network structure parameters. To ensure that the variance of each layer is consistent during forward and backward propagation, the distribution range of the randomly initialized parameters is the number of input parameters passing through that layer. Number of output parameters The uniform distribution within the obtained distribution range is given by the formula for the distribution range:

[0215]

[0216] Step S4-3: Parameter update and optimization;

[0217] The backpropagation error is calculated based on the set objective function and the obtained predicted labels. This error is then used to update and optimize the parameters of the constructed attention-based fusion network.

[0218] Furthermore, in step S4-3, the specific steps are as follows:

[0219] Step S4-3-1 Fully Connected Layer FC Receiver Fusion Features Output predicted labels The calculation formula is:

[0220]

[0221] in, , , These represent the weight matrix, bias vector, and activation function, respectively. This indicates the predicted label.

[0222] Step S4-3-2: Calculate the loss value between the predicted label and the true label using the cross-information entropy loss function: Calculate the distance between the predicted event label and the true label using the cross-information entropy loss function, and the loss value... The calculation formula is as follows:

[0223]

[0224] in: M Represents the total number of samples. This indicates the actual label.

[0225] Step S4-3-3 uses the loss value to back-calculate the parameter gradient of the attention-based fusion network model, and uses the parameter gradient to update the attention-based fusion network deep learning model; this invention uses the Adam algorithm for update optimization, and the calculation steps are as follows:

[0226]

[0227]

[0228]

[0229] in: To calculate the target gradient, , These are the first and second moment estimates of the gradient, respectively. , which are the exponential decay rates of the first and second moments of the gradient, respectively.

[0230] Let the t-th iteration be... , , The network parameter update formula is:

[0231]

[0232] in: For learning rate, For very small numbers, prevent division by zero.

[0233] In the feature fusion network model based on the AFF module, the model parameters are first used. The network model is updated. Then, based on the change in the training loss value, it is determined whether the network model has converged. Once convergence is achieved, the training process stops; otherwise, it jumps to step S4-3-2 and continues iterative updates until the set maximum number of iterations is reached. When the loss function value is less than a certain threshold or the number of iterations exceeds a preset threshold, the model is considered to have converged, and the above iterative process stops. Finally, the model with the best result is selected as the final event recognition model.

[0234] Example 8

[0235] This embodiment is based on Embodiment 7, and performs classification and recognition tests on the trained model.

[0236] Based on the preceding dataset, the model was trained and tested. The confusion matrix for the test set in the track defect dataset is as follows: Figure 8 As shown. To evaluate the performance of this model, a comparative analysis of its recognition effects with several other models based on different modal signals and feature extraction methods is conducted:

[0237] 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: Figure 9 As shown, the methods for extracting spectral energy distribution characteristics and temporal structure change characteristics are the same as those in Examples 3 and 5. Figure 10 (a) is a histogram showing the distribution of recognition accuracy for different input signal modes. The results show that, compared with the single-mode signal model, the multi-mode signal of model A improves the recognition rate by 0.57%-3.27%, indicating that the introduction of multi-mode signals in DAS enriches the dimensions of network perception information and effectively improves the model's recognition ability.

[0238] 2) Remove the LSTM layer from the deep network in Example 3 and use only the CNN layer (e.g., Figure 6 (As shown in the dashed box) to directly extract the time structure change features, called Model C. The method for extracting the spectral energy distribution features is the same as in Example 3. Figure 10 (b) is a histogram showing the recognition accuracy distribution of different temporal structure change feature extraction methods. The results show that the CNN-LSTM network can uncover the global temporal relationship of the signal that the CNN network may ignore, thereby improving the model's discriminability.

[0239] 3) Compare Model A with different spectral energy distribution feature extraction models (Model D-1: no signal multi-scale decomposition and spectral energy distribution feature extraction, only using the temporal structure change features of three modal signals for identification and classification; Model D-2: no signal multi-scale decomposition, uniformly dividing the signal spectrum into K frequency bands and calculating the energy of each frequency band to constitute spectral energy distribution features; Model D-3: replacing the signal decomposition method from VMD to EMD, using the first 10 IMFs to calculate spectral energy distribution features, the calculation steps are the same as step S3-2). The temporal structure change feature extraction method and CNN-LSTM network parameters are the same as in Examples 4 and 5. Figure 10 (c) is a histogram of the recognition accuracy distribution of different spectral energy distribution feature extraction models. VMD decomposes complex signals into stationary components of different scales and refines the decomposition of the lower frequency band containing the main frequency components of wheel-rail interaction, thereby revealing the weak disease information hidden between different components. At the same time, it avoids the potential problems of mode mixing and endpoint effects of traditional decomposition methods such as spectrum equal division and EMD.

[0240] In summary, the track defect identification method proposed in this embodiment has advantages in enriching the dimensions of network sensing information and improving the feature representation ability of signals in the time scale and frequency domain. The identification accuracy reaches 98.17%, which proves the efficiency and reliability of the method in the field of track defect detection.

[0241] Example 9

[0242] This embodiment provides a DAS-based track defect identification system, including:

[0243] The three-mode signal acquisition module is used to perform two differential processing on the collected vibration displacement data to obtain the vibration velocity data and vibration acceleration data.

[0244] The spectrum and time feature acquisition module is used to perform signal decomposition on the vibration displacement data, vibration velocity data, and vibration acceleration data to be measured, respectively, to obtain the displacement spectrum energy distribution characteristics, velocity spectrum energy distribution characteristics, and acceleration spectrum energy distribution characteristics.

[0245] The vibration displacement data, vibration velocity data, and vibration acceleration data to be measured are processed to obtain the displacement time structure change characteristics, velocity time structure change characteristics, and acceleration time structure change characteristics.

[0246] The feature fusion module is used to fuse the energy distribution features of displacement spectrum, energy distribution features of velocity spectrum, energy distribution features of acceleration spectrum, time structure variation features of displacement, time structure variation features of velocity, and time structure variation features of acceleration to obtain fused features.

[0247] The identification and classification module is used to detect and identify the fused features to obtain the classification results.

[0248] Example 10

[0249] A computer device includes a memory and a processor, the memory storing a computer program that, when executed by the processor, causes the processor to perform steps of a DAS-based track defect identification method.

[0250] The computer device can be a desktop computer, laptop, handheld computer, or cloud server, etc. The computer device can interact with the user via a keyboard, mouse, remote control, touchpad, or voice control.

[0251] The memory includes at least one type of readable storage medium, including flash memory, hard disk, multimedia card, card-type memory (e.g., SD or D-interface display memory), 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 the 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, smart media card (SMC), secure digital (SD) card, flash card, etc., equipped on the computer device. Of course, the memory may include both the internal storage unit and the external storage device of the computer 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 defect identification method. In addition, the memory can also be used to temporarily store various types of data that have been output or will be output.

[0252] 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 typically used to control the overall operation of the computer device. In this embodiment, the processor is used to run program code stored in the memory or process data, for example, to run the program code of the DAS-based track defect identification method.

[0253] Example 11

[0254] A computer-readable storage medium storing a computer program that, when executed by a processor, causes the processor to perform the steps of a DAS-based track defect identification method.

[0255] The computer-readable storage medium stores an interface display program that can be executed by at least one processor to cause the at least one processor to perform the steps of the DAS-based track defect identification method described above.

[0256] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they 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 this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk), and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, or network device, etc.) to execute the DAS-based track defect identification method described in the embodiments of this application.

Claims

1. A DAS-based track defect detection and identification method, characterized by, Includes the following steps: Step S1: Acquire the three-mode signal; Collect the vibration displacement data to be measured, and perform two differential processing on the collected vibration displacement data to obtain the vibration velocity data and vibration acceleration data to be measured. Step S2: Obtain the spectral and temporal characteristics; Variational mode decomposition technology is used to decompose the measured vibration displacement data, measured vibration velocity data, and measured vibration acceleration data into signals, 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. The vibration displacement data, vibration velocity data, and vibration acceleration data to be measured are processed to obtain the displacement time structure change characteristics, velocity time structure change characteristics, and acceleration time structure change characteristics. Step S3, feature fusion; The energy distribution characteristics of displacement spectrum, energy distribution characteristics of velocity spectrum, energy distribution characteristics of acceleration spectrum, and time structure variation characteristics of displacement, time structure variation characteristics of velocity, and time structure variation characteristics of acceleration are fused to obtain the fused characteristics. Step S4, Identify and classify; The fused features are detected and identified to obtain classification results; In steps S2, S3, and S4, the track defect detection and identification model is used to obtain time structure change features, perform feature fusion, and identify and classify them. The constructed track defect detection and identification model includes three CNN-LSTM networks, one AFF module, and two fully connected layers. The specific training steps for training this model are as follows: First, the vibration displacement sample data collected by the DAS system is processed twice to obtain vibration velocity sample data and vibration acceleration sample data; The vibration displacement sample data, vibration velocity sample data, and vibration acceleration sample data are then input into the corresponding CNN-LSTM networks. The three CNN-LSTM networks output the temporal structure change features of the vibration displacement sample, the vibration velocity sample, and the vibration acceleration sample, respectively. The vibration displacement sample data, vibration velocity sample data, and vibration acceleration sample data are then decomposed to obtain the spectral energy distribution features of the vibration displacement sample, the vibration velocity sample, and the vibration acceleration sample. Then, the spectral energy distribution characteristics of vibration displacement samples, the spectral energy distribution characteristics of vibration velocity samples, the spectral energy distribution characteristics of vibration acceleration samples, the temporal structure change characteristics of vibration displacement samples, the temporal structure change characteristics of vibration velocity samples, and the temporal structure change characteristics of vibration acceleration samples are input into the AFF module for fusion to obtain the fused characteristics. Finally, the features are fused into the first fully connected layer, and 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, calculates the loss based on the classification result and the corresponding label data, and updates the parameters of the track defect detection and identification model.

2. The DAS-based track rail defect detection and identification method of claim 1, wherein, In step S1, the specific method for obtaining the three-mode signal is as follows: Step S1-1: The vehicle vibrates as it passes over the track. The DAS system continuously accumulates the original signal trajectory on the time axis, forming a spatiotemporal matrix. The spatial channel where the track defect is located is extracted from the spatiotemporal matrix, resulting in a one-dimensional time series S with a total length of L sampling points. This S is then slidably divided along the time axis into M vibration displacement data points of length N. The range of the i-th vibration displacement data point in the time series S is: Step S1-2: Perform differential calculation on the segmented vibration displacement data to obtain the vibration velocity data; perform differential calculation on the vibration velocity data to obtain the vibration acceleration data; wherein, the formula for calculating the differential is: in, Indicates the sampling start point. Indicates the sliding offset. Indicates the data sequence number. This represents the total number of vibration displacement data to be measured. These represent the velocity and acceleration modes, respectively. Indicates a point in time.

3. The method for detecting and identifying track defects based on DAS as described in claim 1, characterized in that, In step S2, when using variational mode decomposition technology to decompose the measured vibration displacement data, measured vibration velocity data, and measured vibration acceleration data into signals, the specific steps are as follows: Step S2-1, the multimodal signal composed of the measured vibration displacement data, measured vibration velocity data, and measured vibration acceleration data is represented as follows: Variational mode decomposition (VMD) is used to decompose multimode signals into... Each IMF component makes each mode function The sum of the estimated bandwidths is minimized, expressed as: Step S2-2, calculate the energy of each IMF component, using the following formula: The calculated energy is then normalized using the following formula: Step S2-3: Construct and obtain the corresponding spectral energy distribution characteristics. The calculation formula is as follows: in, Indicates a point in time The total number, This represents the total number of vibration displacement data to be measured. Indicates a point in time The signal Indicates the data sequence number. These represent the three modes: displacement, velocity, and acceleration. Representing modes The signal after decomposition One portion, Indicates the number of decomposition layers. , They represent The amplitude and phase, Representing modes The signal after decomposition Each component is a time signal point.

4. The DAS-based method for detecting and identifying track defects as described in claim 3, characterized in that, In step S2-1, variational mode decomposition (VMD) is used to decompose the signals of different modes separately, specifically including the following steps: Step S2-1-1: Set the number of decomposition layers For each The Hilbert transform is used to calculate the corresponding analytic signal, and the one-sided spectrum is obtained. The calculation formula is as follows: Step S2-1-2, for each By aliasing each of its corresponding central exponent terms, The spectrum is modulated to the corresponding "baseband", and the calculation formula is: Step S2-1-3, the square norm of the signal gradient in step S2-1-2. Solve and estimate Given the bandwidth, a constrained variational problem is constructed: Step S2-1-4: Solve the variational problem of step S2-1-3 and construct the Lagrange augmented function: Step S2-1-5: Solve the unconstrained variational problem from step S2-1-4 using the alternating direction multiplier method. , The optimal solution is obtained through iterative updates, and the iterative formula is: Step S2-1-6, for The update is performed using the following formula: Step S2-1-7, given the determination precision If the convergence condition is met, the iteration stops; otherwise, return to step S2-1-5 to continue iteration. The convergence condition is: Step S2-1-8, after the iteration is completed Perform an inverse Fourier transform to obtain An IMF component with a certain bandwidth that fluctuates around the center frequency; in, Represents the impulse function. represents an imaginary number, Indicates a point in time. This indicates the center frequency corresponding to the IMF component. The result of decomposition One IMF component, Indicates the center frequency of the IMF component. This indicates a secondary penalty item. , Both represent Lagrange multipliers. This represents taking the derivative of the function with respect to time t. Represents multimodal signals. Indicates frequency, , , Representing components respectively ,Signal Lagrange multipliers Fourier transform, Indicates the number of iterations. This represents the noise tolerance parameter. Indicates the precision of the judgment.

5. The DAS-based track defect detection and identification method as described in claim 1, characterized in that, The CNN-LSTM network consists of a CNN sub-network and an LSTM layer arranged sequentially. The CNN sub-network includes a convolutional layer Conv1, a pooling layer MaxPool1, a ReLU1, a convolutional layer Conv2, a pooling layer MaxPool2, a ReLU2, a convolutional layer Conv3, a pooling layer MaxPool3, and a ReLU3.

6. The DAS-based track defect detection and identification method as described in claim 1, characterized in that, The AFF module consists of two convolutional layers. The specific method for feature fusion in the AFF module is as follows: The Xavier initialization method is used to initialize the structural parameters of the convolutional layers in the AFF module; The spectral energy distribution characteristics of the three-mode signals are obtained by concatenating them in parallel. ; The temporal structure features are obtained by concatenating the temporal structure variation characteristics of the three-mode signals in parallel. ; Spectral energy characteristics respectively Time structure characteristics Perform convolution operations to obtain spectral energy features. Time structure characteristics ; Spectral energy characteristics and time structure features After summing, the local attention features and global attention features are calculated separately. The result is obtained by adding the local attention features and the global attention features and then applying the Sigmoid function. And based on the results Obtain fusion features .

7. A DAS-based track defect detection and identification system, characterized in that, include: The three-mode signal acquisition module is used to perform two differential processing on the collected vibration displacement data to obtain the vibration velocity data and vibration acceleration data. The spectrum and time feature acquisition module is used to perform signal decomposition on the vibration displacement data, vibration velocity data, and vibration acceleration data under test using variational mode decomposition technology, and obtain the displacement spectrum energy distribution characteristics, velocity spectrum energy distribution characteristics, and acceleration spectrum energy distribution characteristics, respectively. The vibration displacement data, vibration velocity data, and vibration acceleration data to be measured are processed to obtain the displacement time structure change characteristics, velocity time structure change characteristics, and acceleration time structure change characteristics. The feature fusion module is used to fuse the energy distribution features of displacement spectrum, energy distribution features of velocity spectrum, energy distribution features of acceleration spectrum, time structure variation features of displacement, time structure variation features of velocity, and time structure variation features of acceleration to obtain fused features. The identification and classification module is used to detect and identify the fused features to obtain the classification results; The spectrum and time feature acquisition module, feature fusion module, and identification and classification module all use the track defect detection and identification model to acquire time structure change features, perform feature fusion, and identify and classify them. The constructed track defect detection and identification model includes three CNN-LSTM networks, one AFF module, and two fully connected layers. The specific training steps for training this model are as follows: First, the vibration displacement sample data collected by the DAS system is processed twice to obtain vibration velocity sample data and vibration acceleration sample data; The vibration displacement sample data, vibration velocity sample data, and vibration acceleration sample data are then input into the corresponding CNN-LSTM networks. The three CNN-LSTM networks output the temporal structure change features of the vibration displacement sample, the vibration velocity sample, and the vibration acceleration sample, respectively. The vibration displacement sample data, vibration velocity sample data, and vibration acceleration sample data are then decomposed to obtain the spectral energy distribution features of the vibration displacement sample, the vibration velocity sample, and the vibration acceleration sample. Then, the spectral energy distribution characteristics of vibration displacement samples, the spectral energy distribution characteristics of vibration velocity samples, the spectral energy distribution characteristics of vibration acceleration samples, the temporal structure change characteristics of vibration displacement samples, the temporal structure change characteristics of vibration velocity samples, and the temporal structure change characteristics of vibration acceleration samples are input into the AFF module for fusion to obtain the fused characteristics. Finally, the features are fused into the first fully connected layer, and 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, calculates the loss based on the classification result and the corresponding label data, and updates the parameters of the track defect detection and identification model.

8. A computer device, characterized in that: It includes a memory and a processor, the memory storing a computer program that, when executed by the processor, causes the processor to perform the steps of the method as described in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that: The system stores a computer program that, when executed by a processor, causes the processor to perform the steps of the method as described in any one of claims 1 to 6.

Images