A physical-data dual driving track steel spring damage detection and identification method and system

Abstract

A physical-data dual driving rail steel spring damage detection and identification method and system belong to the field of rail transit vibration and noise reduction and safety operation and maintenance. Simulation data is obtained through a vehicle-track coupling dynamics model, and real measurement data is collected by using a fiber grating acceleration sensor array system. The data is preprocessed to establish a labeled data set. In the feature extraction aspect, the implicit and explicit features of the data are extracted from the time domain, frequency domain, vibration transmission coefficient and wavelet packet energy spectrum by combining the data-driven and physical-driven methods. After the fusion of these features, a steel spring damage identification model based on the attention mechanism is trained. The method and system effectively integrate the physical and data-driven methods, improve the accuracy and efficiency of steel spring damage detection, and are suitable for complex engineering environments.

Description

Technical Field

[0001] This invention relates to the field of vibration reduction, noise reduction, and safe operation and maintenance in rail transit, specifically to a physical-data dual-drive method and system for detecting and identifying damage to rail springs. Background Technology

[0002] Floating slab track, a high-performance vibration-damping track structure, uses steel spring isolators to support the floating slab, isolating vibrations from the upper track and transmitting them downwards. It is primarily used in sections with high vibration reduction requirements. However, long-term operation has revealed that the steel springs suffer fatigue damage and even fracture under the combined effects of train loads and environmental corrosion. Damage and fracture of the steel springs alter the force transmission structure of the floating slab track, affecting its service performance and potentially threatening train safety.

[0003] Currently, the maintenance system for steel spring floating slab structures is divided into daily inspections, periodic inspections, and special inspections based on inspection cycle, content, and level. Daily inspections are conducted daily using visual inspection methods, including observing rail wear, fastener loosening, the appearance of the track bed structure, and water accumulation in the track bed section. Periodic inspections involve randomly selecting steel spring isolators during regular inspections. The procedures include opening the isolator's top cover, releasing the locking system, using specialized tools to remove the inner sleeve of the isolator, and visually inspecting the corrosion of metal components and the condition of damping materials and steel springs. Special inspections target abnormal sections identified during periodic inspections, involving the removal of all steel springs and visual inspection. Due to the long mileage and large number of steel springs in urban rail transit lines, and the fact that the steel springs are installed between the floating slab and the base and are considered concealed works, damaged steel springs are difficult to locate quickly and accurately, making detection and identification challenging, labor-intensive, and time-sensitive during maintenance windows. Furthermore, traditional manual visual inspection methods heavily rely on human experience, resulting in a high rate of missed inspections.

[0004] Therefore, this invention proposes an efficient and intelligent method for detecting and identifying damage to steel springs. A fiber optic accelerometer array system is established to extract damage information from vibration signals based on physical and data-level data extraction. "Mileage marker" tags assist inspectors in quickly locating the floating plate where the damaged steel spring is located, "Damage Location" tags assist in quickly locating the damaged steel spring's position on the floating plate, and "Damage Status" tags provide the damage status identification result. Then, the target steel spring is manually removed for visual inspection and verification. This method can assist in quickly locating damaged steel springs and identifying their damage status. Summary of the Invention

[0005] To address the shortcomings of existing technologies, the purpose of this disclosure is to provide a physical-data dual-drive method for detecting and identifying damage to track steel springs. This method combines data-driven and physical-driven approaches and fully utilizes the self-attention mechanism in deep learning to extract damage information from the vibration acceleration response data of floating slab tracks, thereby achieving damage status and location identification.

[0006] The objective of this disclosure can be achieved through the following technical solutions:

[0007] A physical-data dual-drive method for detecting and identifying damage to track steel springs is proposed, which obtains simulation data through a vehicle-track coupled dynamics model and acquires measured data through a fiber optic grating accelerometer sensor array system.

[0008] For data processing, create a dataset;

[0009] Based on physics-driven theoretical models, the correspondence between damage signals and physical characteristics of damage is established, and the display features of vibration signals are extracted from multiple aspects such as waveform, spectrum, transmission coefficient, and wavelet packet energy spectrum.

[0010] Based on data-driven approaches, Conv-1d convolutional neural networks are used to mine information from the data level and extract implicit feature matrices.

[0011] The model integrates the explicit feature vector with the implicit feature matrix and uses a self-attention mechanism to build a steel spring damage identification model. It is then pre-trained using simulation data and fine-tuned using measured data to extract information such as the steel spring damage state and damage location from both explicit and implicit features.

[0012] This method provides a new approach, a new idea, and a new perspective for quickly locating and identifying damage to steel springs and guiding repairs, which helps improve the efficiency and accuracy of detection and identification.

[0013] Data Acquisition. A three-dimensional spatially coupled dynamic model of the vehicle-track system was developed using MATLAB. Spring stiffness reduction was used to simulate spring damage, with three damage levels: "healthy," "sub-healthy," and "damaged," corresponding to spring support stiffness reductions of 0–70%, 70%–95%, and 95%–100%, respectively. Simulations were used to calculate the vibration acceleration response of the floating slab and the foundation under different train speeds, track irregularities, damage locations, and damage levels. A fiber optic accelerometer array system was used to collect the vibration acceleration response signals of the floating slab, track, and foundation.

[0014] Data processing;

[0015] 1) Screening: Simulation data may be partially unsatisfactory, and data collected by sensors may have a large number of missing, drifting, or incomplete data. Invalid data will be filtered out.

[0016] 2) Normalization: Normalize the data amplitude to improve the convergence ability of the network model.

[0017] 3) Establish a dataset by splicing the basic vibration acceleration response signal below the corresponding floating plate vibration acceleration signal. That is, two rows of data are generated for each measurement. Taking the k-th acquisition as an example, the floating plate vibration acceleration signal is denoted as ka, and the basic vibration acceleration signal is denoted as kb. kb is spliced ​​below ka. Construct a time series dataset using statistical data with a fixed time window, including the simulated time series dataset S1 and the measured time series dataset S2. The time window width is N. The length of each row of data in datasets S1 / S2 is equal to the time window width N. Set the training subset Sr1 / Sr2, the validation subset Sv1 / Sv2, and the test subset St1 / St2 in a ratio of 6:2:2.

[0018] 4) Add labels. Add a label T, where T = [time, mileage, damage status, damage location]. Label the time, mileage, damage status, and damage location. Create label sets T1 and T2, where T1 corresponds to S1 and T2 corresponds to S2. (Note: Here, "damage location" refers to the relative position of the stiffness-reduced steel spring on the floating plate. The time and mileage in T1 are initially set to 0.)

[0019] Feature extraction;

[0020] Based on simulation data: For data-driven implicit feature extraction, using the simulation time-series dataset S1, after data normalization, 512 values ​​were randomly selected sequentially for the ka-th and kb-th data points. A one-dimensional convolutional neural network, Conv-1d, was used to mine hidden patterns from the massive data and increase the dimensionality from 2D to 128D, achieving implicit feature extraction and establishing an implicit feature matrix Fi with dimensions 128×512. For physics-driven explicit feature extraction, also using the simulation time-series dataset S1, for the ka-th and kb-th data points, physical quantities such as vibration acceleration waveform, spectrum, wavelet packet energy spectrum, and transfer function were calculated to achieve explicit feature extraction. Explicit feature vectors Fk-a and Fk-b were established, and Fk-b was concatenated under Fk-a to construct an explicit feature vector Fk with dimensions 432×1. This reflects the changes in the dynamic characteristics of the floating slab track from a physical perspective, mining damage information.

[0021] Based on measured data: For data-driven implicit feature extraction, using the measured time-series dataset S2, after data normalization, 512 values ​​were randomly selected sequentially for the kc and kd-th data points. A one-dimensional convolutional neural network (conv-1d) was used to mine hidden patterns from the massive data and increase the dimensionality from 2D to 128D, achieving implicit feature extraction. Similarly, an implicit feature matrix Fi with dimensions 128×512 was established. For physics-driven explicit feature extraction, also based on the measured time-series dataset S2, for the kc and kd-th data points, physical quantities such as vibration acceleration waveform, spectrum, wavelet packet energy spectrum, and transfer function were calculated to achieve explicit feature extraction. Explicit feature vectors Fk-c and Fk-d were established respectively. Fk-d was concatenated under Fk-c to construct an explicit feature vector Fk with dimensions 432×1, reflecting the changes in the dynamic characteristics of the floating slab track from a physical perspective to mine damage information.

[0022] Feature fusion;

[0023] Based on simulation data: The display feature vector Fk is repeatedly concatenated below each column of the implicit feature matrix F i to form a brand new feature matrix F with dimensions 560×512.

[0024] Based on actual test data: The display feature vector Fk is repeatedly concatenated below each column of the implicit feature matrix F i to form a brand new feature matrix F with dimensions 560×512.

[0025] A steel spring damage identification model is proposed. This model integrates and weights explicit and implicit features based on an attention mechanism. Taking the k-th vibration acceleration data point in a simulated time-series dataset S1 (or S2) as an example, the explicit feature vector is first appended to the elements of each time step of the k-th data point, thus incorporating the explicit feature vector into the time series and forming an N×217 dimensional matrix. A 512-dimensional attention box is used for translational sliding random sampling, with a CLS classification head added to the side of the box. The initial value is set to 0. Training is performed on each sampling iteration based on the self-attention mechanism, and the results are output to the CLS classification head. A fully connected layer is used for output, and the softmax function is applied to calculate... The weights are calculated, and the result with the highest probability is output, i.e., the identification value. For example, the output T = [20220410-19:36:24, DK316+500-35, Sub-healthy, g] indicates that the data detection acquisition time was 19:36:24 on April 10, 2022, located at mileage station DK316+500 (the 35th floating plate of the sensing system), the damage state was "sub-healthy", and the damage location was "g". Thus, the [time, mileage station, damage state, damage location] information of the detection data is obtained. The model is trained based on the simulated time series dataset S1 and fine-tuned based on the measured time series dataset S2. The trained model is used to monitor the steel springs of the floating plate track in real time, identify damage information, and notify the track inspection and maintenance department for repair.

[0026] Beneficial effects:

[0027] This invention employs a physical-data dual-drive method, combined with the self-attention mechanism in deep learning, to effectively mine damage information in vibration signals and improve the accuracy of steel spring damage state and location identification.

[0028] This invention obtains simulation data through a three-dimensional spatial coupling dynamic model of vehicle-track, simulates various real driving conditions, overcomes the problem of scarcity of measured data, and provides a wealth of data samples;

[0029] This invention can accurately identify the location of damaged steel springs, which helps to carry out targeted maintenance and replacement, improves inspection efficiency, reduces the time and workload of manual inspection, and achieves high efficiency and accuracy in maintenance work. Attached Figure Description

[0030] To more clearly illustrate the technical solutions in the embodiments of this disclosure or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, those skilled in the art can obtain other drawings based on these drawings without creative effort.

[0031] Figure 1 This is a flowchart of an embodiment of the present disclosure;

[0032] Figure 2 This is a schematic diagram illustrating the overall concept of an embodiment of this disclosure;

[0033] Figure 3 This is a three-dimensional spatial coupling dynamic model of the vehicle-floating slab track according to an embodiment of the present disclosure (side view, end view). Detailed Implementation

[0034] The technical solutions of the embodiments of this disclosure will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this disclosure, and not all embodiments. Based on the embodiments of this disclosure, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this disclosure.

[0035] like Figures 1-3 As shown, a physical-data dual-drive method for detecting and identifying damage to track steel springs includes the following steps:

[0036] S1: Simulation data is obtained through a vehicle-track coupled dynamics model, and measured data is obtained through a fiber optic grating acceleration sensor array system;

[0037] S2: Process the simulation data and measured data to create a labeled dataset;

[0038] S3: Based on the dataset, extract data-driven and physics-driven features from the simulation data and the measured data respectively;

[0039] The data-driven aspect employs a Conv-1d convolutional neural network to mine information from the data level and extract implicit feature matrices.

[0040] The physical driving aspect constructs a display feature vector using the waveform, spectrum, vibration acceleration transmission coefficient, and wavelet packet energy spectrum of the vibration signal.

[0041] S4: Fuse the explicit feature vectors with the implicit feature matrix;

[0042] S5: A steel spring damage recognition model is built based on the attention mechanism and trained using the fused features;

[0043] S6: Use the model to identify the new floating slab track spring data.

[0044] In some embodiments, S1 data acquisition includes the following steps;

[0045] S11. Construct the simulation dataset M1

[0046] Based on the vehicle-track coupled dynamics theory, a three-dimensional spatial coupled dynamics model of the vehicle and track is built using MATLAB, such as... Figure 3 As shown, the damage of steel springs is simulated using the steel spring stiffness reduction factor. The vehicle system considers only a single car, with parameters taken from a Type B subway train. The track system considers a 4.8m short floating slab track, with each slab equipped with 4 pairs of steel spring vibration isolators and 8 pairs of fasteners. Track irregularities are taken into account; the calculations use the US AAR 6 level track spectrum (wavelength range 1–80m) and shortwave irregularity spectrum (wavelength range 0.05–1m), which are closer to the actual track irregularity conditions. [1] Specific parameters are shown in Table 1. Given the scarcity of statistical data on steel spring damage, the damage conditions for steel springs are idealized as follows: (1) Steel spring damage is simulated by stiffness reduction, with the same stiffness reduction coefficient; (2) Only one or two steel springs with the same degree of damage are considered on each floating plate. Furthermore, considering the constantly changing track conditions in actual engineering, leading to complex and variable track vibration response characteristics, it is necessary to study the damage identification performance of the self-attention mechanism under various working conditions.

[0047] Set damage status. Three damage statuses are set: "Healthy," "Sub-healthy," and "Damaged," corresponding to a reduction in steel spring support stiffness of 0-70%, 70%-95%, and 95%-100%, respectively. In engineering applications, an alarm should be issued to notify relevant departments for immediate handling in the "Damaged" status, providing damage status and location information to assist workers in quickly locating, inspecting, and repairing the damage. Damage status: [Healthy, Sub-healthy, Damaged], a total of 3 types.

[0048] Set the damage location. The term "location" has two meanings: firstly, the mileage marker of the floating platform where the damaged spring is located, which can be pre-determined through sensor numbers and added to tag T during the data acquisition phase; secondly, "damage location" specifically refers to the relative position of the damaged spring on the floating platform. Only one or two springs on each floating platform are considered to be damaged simultaneously with the same degree of damage. Spring damage is simulated through stiffness reduction, resulting in 10 types from a to j. Types: [a~j], 10 types in total.

[0049] Set the track line operating conditions.

[0050] Vehicle speed. There are three speed limits: 40km / h, 60km / h, and 80km / h. Type: [40km / h, 60km / h, 80km / h], 3 speed limits in total.

[0051] Regarding vehicle load, two different load conditions are simulated by setting the vehicle body mass and the vehicle body's nose-nodding moment of inertia: AW0 and AW1 (separate tables represent unloaded, fully loaded, etc.). Types: [AW1, AW2], a total of 2 types.

[0052] Vibration response extraction point: Extract the vibration acceleration response at the same location on the floating slab and the foundation.

[0053] Sampling settings: The sampling frequency (i.e., the reciprocal of the integral step size for solving the coupled dynamics model) is set to fs = 4096Hz, and the sampling time (i.e., the solution time of the coupled dynamics model) is set to 4s.

[0054] There are a total of 3 × 10 × 3 × 2 = 180 working conditions, and each working condition is run 10 times. Since track irregularity is a non-stationary random excitation, the track irregularity excitation generated each time is of the same level but random. Therefore, it can be guaranteed that 1800 sets (3600 data points) of differentiated simulation raw data can be obtained. Each set of simulation raw data includes one floating plate vibration acceleration data and one foundation vibration acceleration data.

[0055] S12. Construct the experimental large dataset M2

[0056] A fiber optic accelerometer (FBG) array system was constructed using wavelength division multiplexing (WDM) technology. A main optical fiber is laid along the track line. A coupler is installed at the corresponding location of each floating slab to add a branch optical path. Two FBG accelerometers are connected to this branch optical path to measure the vibration acceleration of the track slab and the base, respectively, constructing a large-scale measured dataset M2. Taking full advantage of the linear distribution of the track line and the ease of WDM multiplexing of FBG sensors, a single light source, wavelength demodulation system, and a single main optical fiber are used to achieve long-distance, long-term monitoring of damage to the steel springs of the floating slab track, avoiding the cable redundancy problems of traditional accelerometers.

[0057] Set track operation conditions;

[0058] A section of track with floating steel spring slabs was selected, and a fiber optic grating accelerometer array system was used to monitor and collect vibration data in real time. Each slab was monitored to ensure that vibration data from damaged steel springs accounted for approximately 5% of the total data sample. That is, if there is k... error =If one plate is damaged, then the monitoring range is k. sum = 20 floating slab track lengths, with the monitored quantity being the vibration acceleration response of the floating slabs and foundation. Vibration acceleration data can also be collected under other conditions, such as when steel spring damage exists.

[0059] Since track maintenance must be scheduled during nighttime maintenance windows, there is a time difference between track inspection and track maintenance. This time difference can be fully utilized to collect vibration data samples containing damaged steel springs. Furthermore, let's assume an ideal scenario: the rails, fasteners, and floating slabs of the monitored section are deemed to be in good working order by routine manual inspection, but a small number of floating slabs have damaged steel springs. Train speeds passing through this section fluctuate, and vehicle loads differ during peak, off-peak, and low-peak traffic periods.

[0060] Sampling settings: Sampling frequency is set to fs = 1024Hz, triggered continuous sampling. Assume that 20 high-quality vibration data points triggered by passing trains are obtained from 100 sampling sessions. Each sampling session collects one vibration acceleration response from both the floating slab and the foundation, meaning that each floating slab in the monitoring system has 20 valid samples, with each sample containing one vibration acceleration response from both the floating slab and the foundation.

[0061] The monitoring system contains 20 floating slabs, and a total of 400 sets (800 data points) of raw vibration data of the floating slabs were obtained. Among them, 5% (20 sets, 200 data points) showed damage to the steel springs.

[0062] In some embodiments, S2 data processing includes the following steps;

[0063] S21: Simulation data preprocessing;

[0064] S211: Data Screening. Simulation results may be unsatisfactory due to factors such as improper integration parameter settings, necessitating preliminary data screening.

[0065] S212: Constructing the Dataset. For the simulation calculations, the basic vibration acceleration response signal is concatenated with the corresponding floating slab vibration acceleration signal to construct a time-series dataset S1, with dimensions 1080×N. Setting up Training, Validation, and Test Subsets. Based on the simulation time-series dataset S1, the training subset Sr1, validation subset Sv1, and test subset St1 are set up in a 6:2:2 ratio, with subset sizes of 1080, 360, and 360, respectively.

[0066] S212: Add labels. Define and label the degree of damage corresponding to each acceleration time series X(t), and establish a label T1 = [text(i)], text(i) = [time, mileage marker, damage state, damage location], where i = 1, 2, ..., 1080. Example: text(3) = [#, #, sub-health, g], where the time information is empty, the mileage marker information is empty, the damage state is sub-health, and the damage location is g. Label T1 is divided into sub-labels Tr1, Tv1, and Tt1.

[0067] S22: Preprocessing of measured data.

[0068] S221: Data Filtering. Actual measurements are highly random, and the data collected by the sensor may contain a large number of missing, drifting, or incomplete data. Invalid data is removed.

[0069] S221: Constructing the dataset. Based on the measured data, the basic vibration acceleration response signal is concatenated with the corresponding floating slab vibration acceleration signal to construct a time-series dataset S2, with a dimension of 400×N. Setting up training, validation, and test subsets. Based on the measured time-series dataset S2, the training subset Sr2, validation subset Sv2, and test subset St2 are set up in a 6:2:2 ratio, with subset sizes of 240, 80, and 80, respectively.

[0070] S222: Add labels. Define and label the damage degree corresponding to each acceleration time series X(t), and establish a label T2 = [text(i)], text(i) = [time, mileage station, damage state, damage location], where i = 1, 2, ..., 400. Example: text(3) = [20220410-19:36:24, DK316+500-15, sub-healthy, g], indicating that the data collection time is 19:36:24 on April 10, 2022, the mileage station is DK316+500 (the 15th floating plate in the sensing system), the damage state is sub-healthy, and the damage location is g. Label T2 is divided into sub-labels Tr2, Tv2, and Tt2.

[0071] In some embodiments, S3 feature extraction includes the following steps;

[0072] S31: Feature extraction based on simulation data.

[0073] S311 is based on data-driven aspects:

[0074] This paper utilizes a one-dimensional convolutional neural network architecture to mine hidden patterns from massive datasets, achieving implicit feature extraction and establishing an implicit feature matrix. The processed data undergoes one-dimensional convolution operations using the Conv-1d function. For the ka and kb data points in the dataset, one-dimensional convolutions are used to project the basic data information onto the high-dimensional feature spaces fa and fb, respectively, and then concatenated to form the corresponding implicit features. Specifically, 512 consecutive data points are randomly selected from the ka and kb datasets. The size of the one-dimensional convolution kernel is (1,), the stride is 1, and the number of kernels is 64. Therefore, the dimensions of fa and fb obtained after one-dimensional convolution are 64×512, and the final concatenated hidden feature has a dimension of 128×512. This achieves the extraction of implicit features.

[0075] Specifically, for every two rows of data, such as ka and kb, 512 consecutive columns of data are randomly selected. The number of convolutional kernels is set to 128, the kernel size is set to 1×2, and the stride is set to 1. The process is as follows: the 128 convolutional kernels form a column that constitutes a 128×2 dimensional convolutional kernel matrix, which is multiplied by a 2×512 dimensional data matrix to obtain a 128×512 dimensional score matrix, thus realizing implicit feature extraction.

[0076] S312 is based on physical drive aspects:

[0077] Based on the ka and kb data in the dataset, global features of the data are obtained from the perspectives of waveform, spectrum, transmission coefficient and wavelet packet energy spectrum, which reflect the physical characteristics of the floating plate. Two 216-dimensional display feature vectors are then obtained and concatenated to form a 432-dimensional display feature vector.

[0078] Specifically, this involves extracting feature values ​​from the waveform or time domain of the signal to uncover damage information in the steel spring, and constructing a 2D feature vector using the average peak value and mean square value of vibration acceleration. Signal feature extraction is performed from the frequency domain perspective, extracting the first three dominant frequencies and amplitudes to construct a 6-dimensional feature vector. Damage information of steel springs is analyzed using vibration transmission coefficient as a feature index. Vibration transmission coefficient in the range of [1, 200 Hz] is extracted, and a 200-dimensional feature vector is constructed. Using wavelet energy spectrum as signal features, non-stationary and abrupt information of vibration signals is extracted. By performing three-level wavelet packet decomposition on the vibration signal, an 8-dimensional feature vector is constructed using the wavelet energy spectrum of the eight frequency bands in the third level as an index. Finally, the two 216-dimensional display feature vectors are concatenated to form the display feature vector.

[0079] S3121: Waveform. The acquired time series datasets S1 and S2 are time series signals of vibration acceleration of the floating plate and foundation. Feature values ​​are extracted from the waveform or time domain of the signals to mine information on steel spring damage.

[0080] Step 1: Average Peak Value of Vibration Acceleration. Calculate the average peak value of vibration acceleration X. max ,

[0081]

[0082] Where, x m，i Let represent the i-th peak value among the discrete points of the signal, and k represent the number of amplitude values.

[0083] Step 2: Calculate the mean square value of vibration acceleration. rms ,

[0084]

[0085] Where, x i This represents the i-th amplitude among the discrete points of the signal, and n represents the number of sampling points.

[0086] Step 3: Construct the feature vector. Feature vector

[0087] S3122: Spectrum. Extracting signal features from the frequency domain by performing Fourier transform on the signal's spectrum is the most common processing method.

[0088] Step 1: Extract the first three dominant frequencies and amplitudes. Calculate the frequencies and amplitudes of the first three dominant frequencies w1 and X(w1), w2 and X(w2), and w3 and X(w3). The spectrum X(w) of signal X(t) is given by...

[0089]

[0090] Where x(w1), x(w2), and x(w3) are the 1st, 2nd, and 3rd peaks of the spectrum results, respectively, and k corresponds to their frequency. Step 2: Construct the feature vector. Feature vector

[0091] S3123: Vibration acceleration transmission coefficient. Steel spring isolators support the track slab, reducing the transmission of vibration from the upper track to the foundation. Damage to the steel springs alters the vibration transmission characteristics; therefore, the vibration transmission coefficient is used as a characteristic index to analyze steel spring damage information. The track structure is a layered structure and can be considered a linear invariant system with uniform and superimposed characteristics. In the zero state, the relationship between its output response and the input excitation does not change with the time starting point of the input excitation acting on the system. X(n) and Y(x) are the system's input and output, respectively, and the frequency response characteristic H(w) is independent of the input and output. The estimated value of the transmission coefficient of the track system. Represented as:

[0092]

[0093] Where X(ω) and Y(ω) are the spectral curves of X(n) and Y(x) after Fourier transform, and n is the number of sampling points.

[0094] Step 1: Extract the vibration transmission coefficient in the range of [1, 200 Hz]. According to the requirements of my country's "Technical Specification for Floating Slab Track" (CJJ / T 191-2012), the frequency range for measuring the vibration reduction effect of floating slabs is 1 Hz to 200 Hz. The vibration acceleration transmission coefficient is extracted at 1 Hz intervals.

[0095] Step 2: Construct the feature vector. Feature vector

[0096] S3124: Wavelet packet energy spectrum;

[0097] When the steel spring is damaged, the energy distribution of the floating slab track vibration changes in the frequency domain, containing rich damage information. Due to track irregularities and other combined influences, the track vibration response can be considered to lie between stationary and non-stationary signals. Furthermore, steel spring damage is a localized injury to the track structure, causing local abrupt changes in the vibration response. Simply using Fourier transform cannot handle non-stationary and abrupt changes; it requires the use of time-frequency domain analysis. Wavelet packet decomposition, considered an evolution of wavelet analysis, is an excellent time-frequency analysis method. It not only possesses the time- and frequency-domain localization characteristics and variable resolution of wavelet analysis but also simultaneously decomposes the low- and high-frequency components of the signal. The energy of each frequency band in the wavelet packet decomposition result is called the wavelet energy spectrum. Using this as a signal feature allows for in-depth mining of the non-stationary and abrupt information of the vibration signal, reflecting the steel spring damage.

[0098] Based on time series datasets S1 and S2, a three-level wavelet packet decomposition was performed using db10 and Shannon entropy to extract wavelet energy spectrum information. This enabled feature extraction of non-stationary and abrupt signals, uncovering the steel spring damage information contained in the energy changes of the vibration signal.

[0099] Step 1: Perform 3-level wavelet packet decomposition on the vibration signal. The frequency range of the 8 frequency bands in the 3rd level is: [(n-1)2 -j- 1 fs,n2 -j-1 fs], n = 1, 2, 3…8, where fs is the sampling frequency and the wavelet packet decomposition coefficients in each frequency band are...

[0100] Step 2: Reconstruct the wavelet packet coefficients and extract the signal within each frequency band. The total signal is then represented as:

[0101] Step 3: Calculation of energy within each frequency band and reconstruction of the signal within each frequency band. Corresponding energy Then we have:

[0102]

[0103] Where, x jk (j = 0, 1, 2, ..., 7; k = 0, 1, 2, ..., m) represents the amplitude of discrete points of the reconstructed signal, and m is the number of sampling points.

[0104] Step 4: Constructing the feature vector. After the steel spring is damaged, the signal energy in each frequency band of the track vibration signal will change significantly. A feature vector is constructed using the wavelet energy spectrum of the 8 frequency bands in the 3rd layer as an indicator.

[0105] S3125: Constructing the display feature vector Fe based on physical driving.

[0106] Where ω1~ω3 are the 3rd order dominant frequencies, i=1,2,...,200Hz, F e There are a total of 216 dimensions.

[0107] S32: Feature extraction based on measured data.

[0108] S321: Data-driven aspect: Same as S311.

[0109] S322: Based on physical drive: Same as S312.

[0110] In some embodiments, the specific steps of the S4 implicit and explicit feature fusion include:

[0111] The explicit feature vectors are concatenated with the implicit feature matrix. Specifically, the explicit feature vectors are directly concatenated below each column of the implicit feature matrix to form a 560×512 dimensional fused feature matrix.

[0112] In some embodiments, the specific steps of the S5 hidden steel spring damage identification model include:

[0113] S51: Model building.

[0114] S511: Normalization of the fused feature matrix;

[0115] The fused feature matrix is ​​standardized using the Z-score method. The original dataset is normalized to a dataset with a mean of 0 and a variance of 1. The normalization formula is as follows:

[0116]

[0117] Where x is the observed value, μ is the mean of all sample data, and σ is the standard deviation of all sample data.

[0118] S512: Concatenate CLS vectors.

[0119] A CLS header, initially a 512-dimensional zero vector, is constructed and concatenated to the first column of the fusion feature matrix to form a 560*513-dimensional input matrix.

[0120] S513: Build a self-attention mechanism processing layer.

[0121] Calculate Self-Attention using the following formula:

[0122]

[0123] Where Q is the Query vector, K is the Key vector, and V is the Value vector; K T This is the transpose of the K vector. The softmax function is used to obtain the weights of the values, acting as a scaling factor.

[0124] Each Self-Attention iteration is recorded as a Block, and a total of 4 Self-Attention iterations are performed, which means there are 4 Blocks.

[0125] Based on the self-attention mechanism, the feature information of each part inside the input matrix is ​​obtained, and the feature information of each part is reflected in the CLS head. Finally, only the CLS head vector is output.

[0126] S514: Build a fully connected layer and a softmax layer.

[0127] The CLS head vector is processed through a fully connected layer to generate a 1×(10+3) vector (corresponding to 10 damage locations + 3 damage states). This vector is then fed into a softmax layer to obtain the weights for each damage location and damage state, resulting in 10 categories + 3 types, i.e., 10 types of steel spring damage locations and 3 types of damage states. The category corresponding to the highest probability is the output category, ultimately outputting the damage state and location of the steel spring. Simultaneously, the steel spring damage location and state in the corresponding labels are updated, and the final complete output label T = [data number, floating plate number, damage state, damage location], such as T = [20220410-19:36:24, DK316+500-15, sub-health, g].

[0128] S515: Model training.

[0129] The process is divided into two phases: model training and model fine-tuning. Model fine-tuning is essential because, compared to the idealized assumptions of the vehicle-track three-dimensional coupled dynamics model, the actual support stiffness of each fastener on a real track may differ, and the support stiffness of the slab may vary, making real track tracks more complex. Therefore, a large measured dataset based on actual tracks is established, and explicit / implicit feature values ​​are extracted again to fine-tune the steel spring damage identification model. This helps improve the model's generalization ability and adaptability to real-world engineering applications.

[0130] S:5151: Model Training (Phase I): Based on the simulated time series dataset, a subset Sr1 is trained to extract both explicit and implicit features. Implicit feature extraction requires preprocessing of the simulated data, while explicit feature extraction does not. This data is then input into the steel spring damage recognition model for pre-training. After pre-training, the model is input into the validation subset Sv1 for cross-validation and parameter tuning, following the same procedure. Finally, the model is input into the test set St1 for performance testing. Through continuous optimization, the steel spring damage recognition model achieves high detection accuracy, efficiency, and robustness.

[0131] S5152: Model Fine-tuning (Phase II): After preprocessing the measured data, based on the measured time series dataset S2, explicit and implicit features are extracted respectively. Implicit feature extraction requires data normalization preprocessing. The data is then input into the steel spring damage identification model for retraining, enabling the model to acquire measured information and perform transfer learning. The validation subset Sv2 and test subset St2 are then input into the steel spring damage identification model sequentially, following the same steps as in S321. By inputting measured data, the model is fine-tuned to account for potential differences in reality, thus improving its practical application performance.

[0132] In some embodiments, the specific steps of S6 damage identification and guided maintenance include:

[0133] S61: Damage Identification. The trained model is applied to the actual floating slab track steel spring damage detection project. A fiber optic grating accelerometer array system is used to detect the vibration acceleration signal of the floating slab track. The collected vibration acceleration data of the floating slab and base are input into the model to determine the damage state and location. The damage state includes three states: "healthy", "sub-healthy", and "damaged". The location information includes the mileage information of the floating slab where the damage is located and the location information of the damage on the floating slab.

[0134] S62: Guidance for inspection and maintenance. Damage identification information is sent to the track maintenance department to quickly locate the damaged position and condition of the steel spring, enabling further maintenance work to be arranged.

[0135] Working Principle: Damage to the steel springs in floating slab tracks is a type of hidden track defect, making detection and identification extremely difficult. Conventional manual inspection methods face significant challenges because the steel spring damage is located beneath the floating slab and is difficult to observe. Furthermore, the stiffness reduction caused by fatigue damage cannot be visually identified. Additionally, the track lines are long, and the inspection window is short, further hindering the effectiveness of this method. Current deep learning-based methods also have limitations. First, the data is exceptionally complex: Track structures are multi-layered, heterogeneous systems. Inhomogeneous materials within components such as concrete slabs, voids or cracks, fluctuating structural strength, loose fasteners, and random excitation from track irregularities all contribute to vibration signals containing a large amount of interference. Second, there is a lack of physical characteristics: Feature values ​​extracted solely from the data through deep learning lack physical meaning and interpretability. Abnormal data is difficult to match with actual physical structural anomalies (such as rail corrugation, fastener damage or loosening, slab cracking, and steel spring damage). These factors all contribute to the limitations of purely data-driven identification methods. Numerous studies in the field of structural health monitoring have shown that mining structural damage information solely from the data level is not feasible. It is necessary to comprehensively mine information from multiple aspects, including data and physical factors, to improve the accuracy of identification.

[0136] This solution provides a physical-data dual-driven identification method. First, simulation data is obtained through a vehicle-track coupled dynamics model, and measured data is obtained through a fiber optic accelerometer sensor array system. The data is processed to establish a dataset. Then, based on the theoretical model, the correspondence between damage signals and physical damage characteristics is established. Based on physical driving, the visible features of the vibration signal are extracted from multiple aspects such as waveform, spectrum, transmission coefficient, and wavelet packet energy spectrum. Next, based on data driving, a Conv-1d convolutional neural network is used to mine information from the data level to extract the implicit feature matrix. Then, the visible feature vector and the implicit feature matrix are fused, and a steel spring damage identification model is built using a self-attention mechanism (this model is pre-trained using simulation data and fine-tuned using measured data). The model mines information such as the steel spring damage state and damage location within the vibration signal from both visible and implicit features, providing a new method, new ideas, and new perspective, which helps to improve the detection and identification efficiency and accuracy.

[0137] (2) Obtaining simulation data through a three-dimensional coupled dynamic model of vehicle-track has the following advantages: Although damage or fracture of steel springs in urban rail transit can seriously endanger driving safety, such defects are not common, and detection and identification are difficult. Therefore, obtaining a large number of high-quality samples of track vibration acceleration data for steel spring damage is extremely difficult. In addition, data collected only on a single line has relatively simple characteristics such as plate length and steel spring damage mode, and the obtained data is not representative. This scheme, based on a three-dimensional coupled dynamic model of vehicle-track, sets different working conditions such as "damage state, damage location, vehicle speed, and vehicle load" to comprehensively simulate various real driving conditions, obtain rich simulation data samples, and solve the problem of the scarcity of measured data on steel spring damage.

[0138] (3) The beneficial effects of this method are that, through model building and model training, based on big data operation analysis, the steel spring damage identification results are output, the mileage station number of the floating plate where the damaged steel spring is located and the damaged position of the steel spring on the floating plate are clearly identified, which can greatly improve the inspection efficiency, save manual inspection time, effectively reduce the workload of daily inspection and regular inspection by inspection and maintenance personnel, help inspection and maintenance personnel to carry out special inspections in a targeted manner, and repair and replace the steel spring, so as to achieve targeted repair of steel spring.

[0139] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0140] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention.

Claims

1. A physical-data dual-driven rail steel spring damage detection and identification method, comprising the following steps: S1: obtaining simulation data through a vehicle-rail coupling dynamics model, and obtaining measured data through a fiber Bragg grating acceleration sensor array system; S2: processing the simulation data and the measured data to establish a labeled data set; S3: extracting features in terms of data driving and physical driving respectively from the simulation data and the measured data based on the data set; the data driving aspect uses a Conv-1d convolutional neural network to extract an implicit feature matrix from the data level; the physical driving aspect constructs a display feature vector through the waveform, frequency spectrum, vibration acceleration transfer coefficient, and wavelet packet energy spectrum of the vibration signal; S4: fusing the display feature vector and the implicit feature matrix; S5: building a steel spring damage identification model based on an attention mechanism and training the model using the fused features; S6: identifying new floating slab rail steel spring data using the model; the simulation data obtaining step comprises: building a simulation big data set, building a vehicle-rail three-dimensional space coupling dynamics model using MATLAB based on vehicle-rail coupling dynamics theory, and simulating steel spring damage by using a steel spring stiffness reduction coefficient; setting damage state, damage location, track line working condition, vibration response extraction point, and sampling setting, which respectively correspond to the percentage of steel spring support stiffness reduction, the relative position of damaged steel spring on the floating slab, vehicle speed, vehicle load, vibration acceleration response of the floating slab and the foundation at the same position, sampling frequency, and sampling time; running the vehicle-rail coupling dynamics model under various working conditions to obtain multiple sets of differentiated simulation raw data, each set of simulation raw data containing one piece of floating slab vibration acceleration data and one piece of foundation vibration acceleration data; the specific steps of building the steel spring damage identification model comprise: using the Z-score method to standardize the fused feature matrix; constructing and splicing a CLS vector to the fused feature matrix to form an input matrix; applying a self-attention mechanism processing layer to extract features through multiple Self-Attention calculations; processing the CLS head vector using a fully connected layer and a softmax layer to output the damage state and position of the steel spring. 2.The physical-data dual-driven rail steel spring damage detection and identification method according to claim 1, wherein the measured data is obtained by building a fiber Bragg grating acceleration sensor array system. 3.The physical-data dual-driven rail steel spring damage detection and identification method according to claim 1, wherein in step S2, the data processing comprises: simulation data preprocessing, including data screening, constructing a time series data set and dividing it into training, validation, and test subsets, labeling and dividing into sub-labels; measured data preprocessing, including data screening, constructing a time series data set and dividing it into training, validation, and test subsets, labeling and dividing into sub-labels. 4.The physical-data dual-driven rail steel spring damage detection and identification method according to claim 1, wherein the feature extraction step comprises: ​ ​ ​ The implicit feature matrix is constructed by one-dimensional convolutional neural network for implicit feature extraction of the simulation data set; The explicit features are extracted from the time domain, frequency domain, vibration transmission coefficient, and wavelet packet energy spectrum, etc. The extracted implicit and explicit features are integrated to form a feature vector. The same method is applied to the measured data set for feature extraction.

5. The physical-data dual driving rail steel spring damage detection and identification method according to claim 4, wherein the implicit feature extraction step comprises: Using one-dimensional convolutional neural network architecture to mine hidden regular information from massive data to realize implicit feature extraction and establish an implicit feature matrix; Using Conv-1d function to process the processed data through one-dimensional convolution operation; For k-a and k-b data of the data set, one-dimensional convolution is used to project the basic data information to high-dimensional feature space f-a and f-b, respectively, and they are spliced to form corresponding implicit features.

6. The physical-data dual driving rail steel spring damage detection and identification method according to claim 4, wherein the implicit feature extraction step comprises: Waveform: extracting feature values from the waveform or time domain of the signal, and constructing a 2-dimensional feature vector through the average peak value of vibration acceleration and the mean square value of vibration acceleration; Frequency spectrum: extracting the first three order frequencies and amplitudes from the frequency domain to construct a 6-dimensional feature vector; Vibration acceleration transmission coefficient: analyzing the steel spring damage information by taking the vibration transmission coefficient as the characteristic index, extracting the vibration transmission coefficient in the range of [1, 200Hz], and constructing a 200-dimensional feature vector; Wavelet packet energy spectrum: taking the wavelet energy spectrum as the signal feature to extract the non-stationary and mutation information of the vibration signal, and constructing an 8-dimensional feature vector by taking the 3rd layer 8-band wavelet energy spectrum as the index; Based on the physical driving, the two 216-dimensional explicit feature vectors are spliced to form a 432-dimensional explicit feature vector.

7. The physical-data dual driving rail steel spring damage detection and identification method according to claim 1, wherein the steel spring damage identification model training comprises two stages: The first stage is model training, including extracting explicit and implicit features using simulation time series data set, pre-training, and using validation subset for cross-validation and parameter adjustment; The second stage is model fine-tuning, including feature extraction and data normalization preprocessing using measured time series data set, and pre-training the model again for transfer learning; The model is fine-tuned using the validation and test subsets. The system comprises the following components: Data acquisition unit: configured with a vehicle-track coupling dynamics model to obtain simulation data, and equipped with a fiber Bragg grating acceleration sensor array system to collect measured data; Data processing unit: for processing simulation data and measured data to establish a labeled data set; 8. A physical-data dual driven rail steel spring damage detection and identification system, characterized in that, Feature extraction unit: for feature extraction of simulation data and measured data in terms of data-driven and physical-driven aspects according to the data set; ​ ​ ​ The data-driven aspect adopts a Conv-1d convolutional neural network to mine information from the data layer and extract an implicit feature matrix; The physically-driven aspect constructs a display feature vector through the waveform, frequency spectrum, vibration acceleration transmission coefficient and wavelet packet energy spectrum of the vibration signal; The feature fusion unit is used for fusing the display feature vector and the implicit feature matrix; The steel spring damage identification model unit is constructed based on an attention mechanism, trained using the fused features, and used for identifying new floating slab track steel spring data; The simulation data acquisition steps include: A simulation big data set is constructed, a vehicle-track coupling dynamics model is built using MATLAB based on vehicle-track coupling dynamics theory, and the steel spring damage condition is simulated using a steel spring stiffness reduction coefficient; The damage state, damage position, track line condition, vibration response extraction point and sampling settings are set, corresponding to the steel spring support stiffness reduction percentage, the relative position of the damaged steel spring on the floating slab, the vehicle speed, the vehicle load, the vibration acceleration response of the floating slab and the foundation at the same position, the sampling frequency and the sampling time; The vehicle-track coupling dynamics model is run under various conditions to obtain multiple sets of differentiated simulation raw data, each set of simulation raw data containing a floating slab vibration acceleration data and a foundation vibration acceleration data; The specific steps for building the steel spring damage identification model include: The fused feature matrix is standardized using the Z-score method; The CLS vector is constructed and spliced into the fused feature matrix to form an input matrix; The self-attention mechanism processing layer is applied to extract features through multiple Self-Attention calculations; The CLS head vector is processed using a fully connected layer and a softmax layer to output the damage state and position of the steel spring.

Images