Physical-data dual-drive track steel spring damage detection and identification method and system

Abstract

The invention discloses a physical-data dual-drive track steel spring damage detection and identification method and system, belongs to the field of vibration and noise reduction and safe operation and maintenance of track traffic, and is characterized in that simulation data are acquired through a vehicle-track coupling dynamic model, and actually measured data are collected by using a fiber grating acceleration sensor array system. The data are preprocessed, and a data set with labels is established; in the aspect of feature extraction, data driving and physical driving methods are combined, and implicit and display features of data are extracted from the angles of time domain, frequency domain, vibration transfer coefficient, wavelet packet energy spectrum and the like. After being fused, the features are used for training a steel spring damage identification model based on an attention mechanism. According to the method and the system, physical and data driving methods are effectively integrated, the accuracy and the efficiency of steel spring damage detection are improved, and the method and the system are suitable for complex engineering environments.

Description

Technical Field

[0001] The present invention relates to the field of vibration and noise reduction and safe operation and maintenance of rail transit, and in particular to a physical-data dual-drive rail spring damage detection and identification method and system. Background Art

[0002] As a vibration-reducing track structure with excellent performance, the steel spring floating plate track supports the floating plate through steel spring isolators to isolate the vibration of the upper part of the track from being transmitted to the lower part. It is mainly used in areas with high vibration reduction requirements. However, long-term operation has found that the steel springs suffer fatigue damage or even break due to combined factors such as train loads and environmental corrosion. Damage and breakage of the steel springs will change the force transmission structure system of the floating plate track, affect the service performance of the floating plate track, and may even threaten driving safety.

[0003] At present, the inspection system for steel spring floating plate structures is divided into daily inspection, regular inspection and special inspection according to the inspection cycle, content and level. The inspection cycle of daily inspection is every day, and the inspection method is visual inspection. The inspection content includes observing the wear of rails and the looseness of fasteners, the appearance of the roadbed structure and the water accumulation in the roadbed area; the regular inspection is to regularly check the steel spring isolator by random inspection. The steps include opening the upper cover of the isolator, releasing the locking system, using special tools to remove the inner sleeve of the isolator, and visually inspecting the corrosion of metal parts and the state of damping materials and steel springs; the special inspection is to take out all steel springs and visually inspect the abnormal sections during the regular inspection. Due to the long mileage of urban rail transit lines, the large number of steel springs, and the steel springs installed between the floating plate and the base, which is a hidden project, it is difficult to quickly and accurately locate the damaged steel springs, the detection and identification are difficult, the work intensity is high, and the time for repairing the skylight is tight. In addition, the traditional manual visual inspection method is extremely dependent on manual experience and has a high rate of missed detection.

[0004] Therefore, the present invention proposes an efficient and intelligent method for intelligent detection and identification of steel spring damage, establishes a set of fiber grating acceleration sensor array system, extracts damage information based on physical and data levels for vibration signals, and assists patrol personnel to quickly locate the floating plate where the damaged steel spring is located through the "mileage pile number" label, and assists in quickly locating the position of the damaged steel spring on the floating plate through the "damage position" label, and gives the damage status identification result through the "damage status" label. Then, the target steel spring is manually taken out for visual inspection and verification. This method can assist in quickly locating the damaged steel spring and judging and identifying the damage status. Summary of the invention

[0005] In view of the shortcomings of the prior art, the purpose of the present invention is to provide a physical-data dual-driven rail steel spring damage detection and identification method, which adopts a data-driven and physically-driven fusion approach and makes full use of the self-attention mechanism in deep learning to mine damage information from the vibration acceleration response data of the floating plate track to achieve damage status and location identification.

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

[0007] A physical-data dual-driven rail spring damage detection and identification method, which obtains simulation data through a vehicle-rail coupling dynamics model and obtains measured data through a fiber Bragg grating acceleration sensor array system;

[0008] Process the data and build a data set;

[0009] Based on physical drive, the corresponding relationship between damage signal and physical characteristics of damage is established according to theoretical model, and the display characteristics of vibration signal are extracted from waveform, spectrum, transfer coefficient, wavelet packet energy spectrum and other aspects;

[0010] Based on data-driven, Conv-1d convolutional neural network is used to mine information from the data level to extract implicit feature matrix;

[0011] The displayed feature vector and the implicit feature matrix are integrated, and the self-attention mechanism is used to build a steel spring damage recognition model. The simulation data is used for pre-training and the measured data is used for fine-tuning. The information such as the steel spring damage state and damage location in the vibration signal is mined from both the displayed features and the implicit features.

[0012] It can quickly locate and identify steel spring damage and guide maintenance, providing a new method, new idea, and new perspective, which will help improve the efficiency and accuracy of detection and identification.

[0013] Data acquisition. A three-dimensional vehicle-track coupling dynamic model was compiled based on MATLAB, and the steel spring stiffness reduction was used to simulate the steel spring damage. Three damage levels of "healthy, sub-healthy, and damaged" were set, corresponding to the steel spring support stiffness reduction of 0-70%, 70%-95%, and 95%-100%, respectively. The vibration acceleration response of the floating plate and the vibration acceleration response of the foundation under different train running speeds, different track unevenness excitations, different damage locations, and different damage levels were obtained by simulation calculation; the fiber Bragg grating acceleration sensor array system was used to collect the vibration acceleration response signals of the floating plate track and foundation.

[0014] Data processing;

[0015] 1) Screening: Some simulation data may not be ideal, and the data collected by the sensor may be missing, drifting, incomplete, etc. Invalid data should be screened out.

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

[0017] 3) Establish a data set, splice the basic vibration acceleration response signal under the corresponding floating plate vibration acceleration signal, that is, generate two rows of data for each measurement. Taking the kth acquisition as an example, the floating plate vibration acceleration signal is recorded as ka, and the basic vibration acceleration signal is recorded as kb. Splice kb under ka, and construct a time series data set with fixed time window statistical data, including a simulation time series data set S1 and a measured time series data set S2. The time window width is N, and the length of each row of data in the data set S1 / S2 is equal to the time window width N. The training subset Sr1 / Sr2, the verification subset Sv1 / Sv2 and the test subset St1 / St2 are set in a ratio of 6:2:2.

[0018] 4) Add labels, add labels T, T = [time, mileage pile number, damage status, damage location], mark time, mileage, damage status, damage location and other information, establish label sets T1 and T2, T1 corresponds to S1, T2 corresponds to S2. (Note: Here, "damage location" refers to the relative position of the steel spring with reduced stiffness on the floating plate, and the time and mileage pile number in T1 are initially 0.)

[0019] Feature extraction;

[0020] Based on simulation data: Data-driven implicit feature extraction: Based on the simulation time series data set S1, for the ka and kb data, after data normalization, 512 values ​​are randomly selected in order, and the one-dimensional convolutional neural network Conv-1d is used to mine hidden regular information from massive data and upgrade the dimension from 2 dimensions to 128 dimensions to achieve implicit feature extraction and establish an implicit feature matrix F i with a dimension of 128×512. Physically driven display feature extraction: Also based on the simulation time series data set S1, for the ka and kb data, the vibration acceleration waveform, spectrum, wavelet packet energy spectrum, transfer function and other physical quantities are calculated to achieve display feature extraction, and display feature vectors Fk-a and Fk-b are established respectively. Fk-b is spliced ​​under Fk-a to construct a display feature vector Fk with a dimension of 432×1, which reflects the changes in the dynamic characteristics of the floating plate track from the physical essence to mine damage information.

[0021] Based on measured data: Data-driven implicit feature extraction: Based on the measured time series data set S2, for the kc-th and kd-th data, after data normalization, 512 values ​​are randomly selected in order, and the one-dimensional convolutional neural network conv-1d is used to mine hidden regular information from massive data and upgrade from 2 dimensions to 128 dimensions to achieve implicit feature extraction. The implicit feature matrix F i is also established with a dimension of 128×512. Physically driven display feature extraction: Also based on the measured time series data set S2, for the kc-th and kd-th data, the vibration acceleration waveform, spectrum, wavelet packet energy spectrum, transfer function and other physical quantities are calculated to achieve display feature extraction, and display feature vectors Fk-c and Fk-d are established respectively. Fk-d is spliced ​​under Fk-c to construct a display feature vector Fk with a dimension of 432×1, which reflects the changes in the dynamic characteristics of the floating plate track from the physical essence to mine damage information.

[0022] Feature fusion;

[0023] Based on simulation data: The displayed feature vector Fk is repeatedly concatenated under each column of the implicit feature matrix Fi to form a new feature matrix F with a dimension of 560×512.

[0024] Based on measured data: The displayed feature vector Fk is repeatedly concatenated under each column of the implicit feature matrix Fi to form a new feature matrix F with a dimension of 560×512.

[0025] Steel spring damage recognition model. The fusion and weighting of explicit features and implicit features are realized based on the attention mechanism. Taking the k-th vibration acceleration data in the simulation time series data set S1 (or S2) as an example, the display feature vector is first spliced ​​under the elements of the k-th data at each moment, and the display feature vector is integrated into the time series to form an N×217-dimensional matrix. The attention frame with a width of 512 is used for random sampling by translation and sliding. A cls classification head is added to the side of the frame, and the initial value is set to 0. Based on the self-attention mechanism, each sampling is trained and the results are output to the cls classification head. The fully connected layer output is used and the softmax function is used for calculation. Calculate the weights and output the result with the highest probability, that is, the recognition value. For example, output T = [20220410-19:36:24, DK316+500-35, sub-health, g], indicating that the data detection acquisition time is 19:36:24 on April 10, 2022, located at mileage pile number DK316+500 (the 35th floating plate of the sensor system), the damage state is "sub-healthy", and the damage location is "g class", thereby obtaining the [time, mileage pile number, damage state, damage location] information of the detection data. The model is trained based on the simulation time series data set S1, and the model is fine-tuned based on the measured time series data set S2. Use the trained model to monitor the floating plate track steel spring in real time, identify damage information, and notify the track inspection and maintenance department for maintenance.

[0026] Beneficial effects:

[0027] The present invention adopts 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 position recognition;

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

[0029] The present invention can accurately identify the location of a damaged steel spring, 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. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, for ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

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

[0032] Figure 2 It is a schematic diagram of the overall concept of the embodiment of the present disclosure;

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

[0034] The following will be combined with the drawings in the embodiments of the present disclosure to clearly and completely describe the technical solutions in the embodiments of the present disclosure. Obviously, the described embodiments are only part of the embodiments of the present disclosure, rather than all the embodiments. Based on the embodiments in the present disclosure, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present disclosure.

[0035] like Figure 1-Figure 3 As shown, a physical-data dual-driven rail spring damage detection and identification method includes the following steps:

[0036] S1: obtain simulation data through the vehicle-track coupled dynamics model and obtain measured data through the fiber Bragg grating acceleration sensor array system;

[0037] S2: Process the simulation data and measured data to establish a labeled data set;

[0038] S3: Extract data-driven and physical-driven features of the simulation data and measured data based on the data set;

[0039] The data-driven aspect uses Conv-1d convolutional neural network to mine information from the data level to extract implicit feature matrix;

[0040] The physical drive aspect constructs a display feature vector through the waveform, spectrum, vibration acceleration transfer coefficient and wavelet packet energy spectrum of the vibration signal;

[0041] S4: Fusion of displayed feature vector and implicit feature matrix;

[0042] S5: Build a steel spring damage recognition model based on the attention mechanism and use the fused features for training;

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

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

[0045] S11. Constructing a large simulation dataset M1

[0046] Based on the vehicle-track coupling dynamics theory, a vehicle-track three-dimensional spatial coupling dynamics model is built using MATLAB. Figure 3 As shown in the figure, the steel spring damage is simulated by the steel spring stiffness reduction coefficient. Among them, the vehicle system only considers a single vehicle, and its parameters are taken from the B-type subway train; the track system considers a 4.8m short floating plate track, and each floating plate is equipped with 4 pairs of steel spring isolators and 8 pairs of fasteners. Considering the track unevenness, the American AAR6 track spectrum (wavelength range 1~80m) and short-wave unevenness spectrum (wavelength range 0.05~1m) which are closer to the actual line unevenness conditions are used in the calculation. [1] Specific parameters are shown in Table 1. In view of the lack of statistical data on steel spring damage, the damage conditions of steel springs are ideally set as follows: (1) The damage of steel springs is simulated by stiffness reduction, and the stiffness reduction coefficient is the same; (2) Only one or two steel springs are considered to be damaged on each floating plate, and the damage degree is the same. In addition, considering that the working conditions of track lines in actual projects are constantly changing, resulting in complex and changeable track vibration response characteristics, it is necessary to study the damage identification performance of the self-attention mechanism under various working conditions.

[0047] Set the damage status. Set three damage statuses: "healthy, sub-healthy, damaged", corresponding to the reduction of the spring support stiffness by 0-70%, 70%-95% and 95%-100% respectively. In engineering applications, for the "damage" status, an alarm needs to be issued to notify the relevant departments to handle it immediately, and provide damage status and location information to assist workers in quickly locating, inspecting and repairing. Damage status: [healthy, sub-healthy, damaged], a total of 3 types.

[0048] Set the damage position. The so-called "position" has two meanings. One is the mileage pile number of the floating plate where the damaged steel spring is located, which can be known in advance through the sensor number during the data collection stage and added to the label T; here, the "damage position" specifically refers to the relative position of the damaged steel spring on the floating plate. On each floating plate, only one or two steel springs are considered to be damaged at the same time and the degree of damage is the same. The steel spring damage is simulated by stiffness reduction. There are 10 types from a to j. Type: [a~j], a total of 10 types.

[0049] Set the track line conditions.

[0050] Vehicle speed. There are 3 types of vehicle speeds: 40km / h, 60km / h, and 80km / h. Type: [40km / h, 60km / h, 80km / h,], 3 types in total.

[0051] Vehicle load: By setting the vehicle mass and the vehicle nodding moment of inertia, two different load conditions are simulated: AW0 and AW1 (with separate tables indicating empty load, full load, etc.). Type: [AW1, AW2], 2 types in total.

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

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

[0054] There are 3×10×3×2=180 working conditions in total, and each group of working conditions is run 10 times. Since track irregularity is a non-stationary random excitation, the track irregularity excitation generated each time by the same level of track irregularity spectrum is of the same level but random, so it is guaranteed that 1800 groups (3600) of differentiated simulation raw data can be obtained, and each group of simulation raw data contains a floating plate vibration acceleration data and a basic vibration acceleration data.

[0055] S12. Constructing the measured large data set M2

[0056] Fiber Bragg grating (FBG) accelerometers are used to build a fiber Bragg grating accelerometer array system based on wavelength division multiplexing (WDM) technology. In the system, a main optical fiber is laid along the direction of the track line, and a coupler is set at the corresponding position of each floating plate to add a branch optical path. Two fiber Bragg grating accelerometers are connected to the branch optical path to measure the vibration acceleration of the track plate and the base respectively, and a large measured data set M2 is constructed. By fully combining the linear distribution characteristics of the track line and the easy wavelength division multiplexing characteristics of the fiber Bragg grating sensor, a single set of light source, wavelength demodulation system and a main optical fiber are used to realize long-distance and long-term monitoring of floating plate track steel spring damage, avoiding the problem of cable redundancy of traditional accelerometers.

[0057] Set track line conditions;

[0058] A section of track with steel spring floating slabs was selected, and a fiber Bragg grating acceleration sensor array system was used to monitor and collect vibration data in real time. Each slab was monitored to ensure that the vibration data of damaged steel springs accounted for about 5% of the total data samples. That is, if there are k error = 1 board is damaged, then the monitoring range is k sum = 20 floating slab track lengths, the monitored quantity is the vibration acceleration response of the floating slab and foundation. The line vibration acceleration data can also be collected based on other situations when there is steel spring damage.

[0059] Since track maintenance must be arranged during the nighttime window, there is a time difference between track inspection and track maintenance, which can be fully utilized to collect vibration data samples containing damaged steel springs. And assuming a relatively ideal situation, the rails, fasteners, floating plates and other components of the monitored section are considered to be in good working order through routine manual inspections, but a small number of floating plates have damaged steel springs, the speed of trains passing through the section fluctuates, and the vehicle loads are different during peak, off-peak and off-peak periods.

[0060] Sampling settings: The sampling frequency is set to fs = 1024Hz, and the sampling is triggered continuously. Assume that 20 high-quality vibration data excited by train passing are obtained in 100 acquisitions. Each time, one vibration acceleration response of the floating plate and the foundation is collected, that is, 20 groups of valid samples of each floating plate in the monitoring system, each group contains one vibration acceleration of the floating plate and the foundation.

[0061] There are 20 floating plates in the monitoring system, and a total of 400 groups (800 records) of floating plate vibration raw data are obtained, of which 5% (20 groups, 200 records) have damaged steel springs.

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

[0063] S21: simulation data preprocessing;

[0064] S211: Data screening. The simulation calculation may lead to unsatisfactory solution results due to improper setting of integral parameters and other factors, so preliminary data screening is required.

[0065] S212: Construct a data set. For simulation calculations, the basic vibration acceleration response signal is spliced ​​after the corresponding floating plate vibration acceleration signal to construct a time series data set S1, and the dimension of S1 is 1080×N. Set training, validation, and test subsets. Based on the simulation time series data set S1, set the training subset Sr1, validation subset Sv1, and test subset St1 in a ratio of 6:2:2, with subset capacities of 1080, 360, and 360, respectively.

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

[0067] S22: Preprocessing of measured data.

[0068] S221: Data screening. Actual measurements are highly accidental, and the data collected by the sensor may be missing, drifting, incomplete, etc. Invalid data should be removed.

[0069] S221: Construct a data set. Based on the measured data, the basic vibration acceleration response signal is spliced ​​after the corresponding floating plate vibration acceleration signal to construct a time series data set S2, and the dimension of S2 is 400×N. Set training, validation, and test subsets. Based on the measured time series data set S2, set the training subset Sr2, validation subset Sv2, and test subset St2 in a ratio of 6:2:2, with subset capacities of 240, 80, and 80, respectively.

[0070] S222: Add labels. Define and mark the damage degree corresponding to each acceleration time series X(t), and establish the label T2 = [text(i)], text(i) = [time, mileage pile number, damage state, damage location], where i = 1, 2, ..., 400. Example: text(3) = [20220410-19:36:24, DK316+500-15, sub-health, g], indicating that the data collection time is 19:36:24 on April 10, 2022, the pile number mileage is DK316+500 (the 15th floating plate in the sensor system), the damage state is sub-health, 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] Use the one-dimensional convolutional neural network architecture to mine hidden regular information from massive data, realize implicit feature extraction, and establish an implicit feature matrix. Perform a one-dimensional convolution operation on the processed data and use the Conv-1d function to process it. For the ka and kb data of the data set, use one-dimensional convolution to project the basic data information into the high-dimensional feature space fa and fb, respectively, and splice them to form the corresponding implicit features. The specific process is to randomly cut 512 consecutive data from the data ka and kb, the size of the one-dimensional convolution kernel is (1,), the step size is 1, and the number is 64, that is, the dimensions of fa and fb obtained after one-dimensional convolution are 64×512, and the dimension of the hidden features finally spliced ​​together is 128×512. Realize 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 cut, the number of convolution kernels is set to 128, the convolution kernel size is set to 1×2, and the step size is set to 1. The specific process is that 128 convolution kernels form a column to form a 128×2-dimensional convolution kernel matrix, which is multiplied by the 2×512-dimensional data matrix to obtain a 128×512-dimensional score matrix to achieve implicit feature extraction.

[0076] S312 is based on physical drive aspects:

[0077] According to the ka and kb data of the data set, the global characteristics of the data are obtained from the perspectives of waveform, spectrum, transfer coefficient and wavelet packet energy spectrum that reflect the physical characteristics of the floating plate, and then two 216-dimensional display feature vectors are obtained. The two are spliced ​​as a whole to form a 432-dimensional display feature vector.

[0078] Specifically, it extracts characteristic values ​​from the waveform angle or time domain angle of the signal, mines the steel spring damage information, and constructs a 2D characteristic vector through the average peak value of vibration acceleration and the mean square value of vibration acceleration. Extract signal features from the frequency domain, extract the first three main frequencies and amplitudes, and construct a 6-dimensional feature vector The vibration transfer coefficient is used as a characteristic index to analyze the steel spring damage information, extract the vibration transfer coefficient in the range of [1,200Hz], and construct a 200-dimensional feature vector The wavelet energy spectrum is used as the signal feature to extract the non-stationary and mutation information of the vibration signal. The vibration signal is decomposed into three layers of wavelet packets and an 8-dimensional feature vector is constructed using the third layer of 8-band wavelet energy spectrum as an indicator. Finally, the two 216-dimensional display feature vectors are concatenated as a whole to form a display feature vector.

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

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

[0081]

[0082] Among them, x m，i represents the i-th peak value in the discrete points of the signal, and k represents the number of amplitudes.

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

[0084]

[0085] Among them, x i Represents the i-th amplitude in the signal discrete point, and n represents the number of sampling points.

[0086] Step 3: Construct the eigenvector.

[0087] S3122: Spectrum. The most common processing method is to perform Fourier transform on the signal to extract signal features from the frequency domain.

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

[0089]

[0090] Among them, x(w1), x(w2), and x(w3) are the 1st, 2nd, and 3rd peaks of the spectrum results, respectively, and k corresponds to their frequencies. Step 2: Construct the eigenvector. Eigenvector

[0091] S3123: Vibration acceleration transfer coefficient. The steel spring isolators support the track plate and reduce the vibration of the upper part of the track to the lower foundation. Damage to the steel spring will change the vibration transfer characteristics, so the vibration transfer coefficient is used as a characteristic indicator to analyze the steel spring damage information. The track structure is a layered structure and can be regarded as a linear invariant system with uniform and superposition characteristics. In the zero state, the relationship between its output response and input excitation does not change with the starting time of the input excitation acting on the system. X(n) and Y(x) are the input and output of the system, respectively, and the frequency response characteristic H(w) is independent of the input and output. Estimated value of the transfer coefficient of the track system It is expressed as:

[0092]

[0093] Among them, X(ω) and Y(ω) are the frequency spectrum curves of X(n) and Y(x) after Fourier transformation, and n is the number of sampling points.

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

[0095] Step 2: Construct the eigenvector.

[0096] S3124: Wavelet packet energy spectrum;

[0097] When the steel spring is damaged, the energy distribution of the floating plate track vibration in the frequency domain will change, which contains rich damage information. Due to the combined influence of track unevenness and other factors, the vibration response of the track can be considered to be between a stationary signal and a non-stationary signal. The steel spring damage is a local damage to the track structure, which will cause local mutations in the vibration response. The simple use of Fourier transform cannot handle non-stationary information and local mutation information. It needs to be processed in combination with time-frequency domain analysis. As an excellent time-frequency analysis method, wavelet packet decomposition is considered to be the evolution of wavelet analysis. It not only has the localization characteristics of wavelet analysis in the time and frequency domains and the variable resolution characteristics, but also can simultaneously decompose the low-frequency and high-frequency parts of the signal. The energy of each frequency band in the wavelet packet decomposition result is called the wavelet energy spectrum. Using it as a signal feature can deeply mine the non-stationary and mutation information of the vibration signal and reflect the damage of the steel spring.

[0098] Based on the time series data sets S1 and S2, a three-layer wavelet packet decomposition is performed based on db10 and Shannon entropy to extract the wavelet energy spectrum information of the data, realize the feature extraction of non-stationary signals and mutation signals, and mine the steel spring damage information contained in the energy change of the vibration signal

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

[0100] Step 2: Reconstruct the wavelet packet decomposition coefficients and extract the signals in each frequency band The total signal is expressed as:

[0101] Step 3: Calculate the energy in each frequency band and reconstruct the signal in each frequency band The corresponding energy Then we have:

[0102]

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

[0104] Step 4: Construct a eigenvector. After the steel spring is damaged, the signal energy in each frequency band of the rail vibration signal will change significantly. A eigenvector is constructed using the 8-band wavelet energy spectrum of the third layer as an indicator.

[0105] S3125: Construct a display feature vector Fe based on physical drive.

[0106] , where ω1~ω3 are the 3rd order main frequency, i=1, 2, ..., 200Hz, F e A total of 216 dimensions.

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

[0108] S321: Based on data-driven aspects: same as S311.

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

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

[0111] The displayed feature vector is concatenated with the implicit feature matrix. Specifically, the displayed feature vector is directly concatenated below each column of the implicit feature matrix to form a 560×512-dimensional fusion feature matrix.

[0112] In some embodiments, the S5 hidden steel spring damage identification model specifically includes the following steps:

[0113] S51: Model building.

[0114] S511: Normalization of fusion feature matrix;

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

[0116]

[0117] Among them, 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] Construct a CLS header with an initial 512-dimensional zero vector and concatenate it to the first column of the fused feature matrix to form a 560*513-dimensional input matrix.

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

[0121] Self-Attention is calculated using the following formula:

[0122]

[0123] Among them, Q is the Query vector, K is the Key vector, and V is the Value vector; K T is the transpose of the K vector, is the scaling factor and the softmax function is used to obtain the weights of the values.

[0124] Each Self-Attention is recorded as a Block, and a total of 4 Self-Attentions are calculated, which includes 4 Blocks.

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

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

[0127] After the CLS head vector is processed by a fully connected layer, a vector of 1×(10+3) (corresponding to 10 damage locations + 3 damage states) is generated, which is then input into the softmax layer to obtain the weights of each damage location and each damage state, which is divided into 10 categories + 3 types, namely 10 steel spring damage location categories and 3 damage states. The category corresponding to the maximum probability is the output category, and the damage state and position of the steel spring are finally output. At the same time, the steel spring damage location and state in the corresponding label 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] It is divided into two stages: model training and model fine-tuning. Model fine-tuning is very necessary. Compared with the idealized assumptions of the vehicle-track three-dimensional spatial coupling dynamics model, the actual support stiffness of each fastener of the actual track line may be different, the support stiffness of the plate may be different, and the actual track line is more complex. Therefore, a large measured data set is established based on the actual line, and the explicit / implicit eigenvalues ​​are extracted again to achieve fine-tuning of the steel spring damage identification model, which helps to improve the generalization ability of the model and its ability to adapt to actual engineering.

[0130] S:5151: Model training (Phase I): Based on the simulation time series data set training subset Sr1, explicit features and implicit features are extracted respectively. The implicit feature extraction part needs to be preprocessed by simulation data first, while explicit feature extraction does not need preprocessing. Then input it into the steel spring damage recognition model for pre-training; after pre-training, input the validation subset Sv1, cross-validate and adjust the parameters, and the operation is the same as above; finally, input the test set St1 for performance testing. Through continuous optimization, the steel spring damage recognition model has high detection accuracy, efficiency and robustness.

[0131] S5152: Model fine-tuning (Phase II): After pre-processing the measured data, based on the measured time series data set S2, the explicit features and implicit features are extracted, where the implicit feature extraction needs to be pre-processed with data normalization, and then input into the steel spring damage recognition model, and pre-train again, so that the model can obtain the measured information and perform transfer learning; then the verification subset Sv2 and the test subset St2 are respectively input into the steel spring damage recognition model in sequence, and the steps are the same as S321. By inputting the measured data, the model is fine-tuned for the actual differences that may occur, and fits the actual situation, thereby improving the actual application effect of the model.

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

[0133] S61: Damage identification. The trained model is applied to the actual floating plate track steel spring damage detection project. The fiber grating accelerometer array system is used to detect the floating plate track vibration acceleration signal. The collected floating plate and base vibration acceleration data 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 plate where the damage is located and the location information of the damage on the floating plate.

[0134] S62: Guide inspection and maintenance. Send the damage identification information to the track maintenance department to quickly locate the steel spring damage position and quickly identify the damage status so as to arrange further maintenance work.

[0135] Working principle: The steel spring damage of floating plate track is a hidden track disease, which is extremely difficult to detect and identify. For the conventional manual inspection method, the steel spring damage is located below the floating plate and is not easy to observe, and the stiffness reduction caused by fatigue damage cannot be identified by the naked eye. In addition, the track line is long and the track inspection operation window time is short, so this method faces great resistance. The various new methods based on deep learning proposed at present also have limitations. First, the data is abnormally complex: Since the track structure is a multi-layer heterogeneous complex structural system, the internal materials of components such as concrete slabs are not uniform, there are voids or cracks, the structural strength fluctuates, the fasteners and other connections may be loose, and the track is uneven. Random excitation, etc., all of these will cause the vibration signal to contain a lot of interference information. Secondly, lack of physical features: The feature values ​​extracted from the data level only through deep learning have no physical meaning and are not interpretable. The abnormal data is difficult to match with the actual physical structure abnormalities (such as rail corrugation, fastener damage or looseness, plate cracking, steel spring damage, etc.). All these lead to limitations in identification methods based solely on data-driven methods. Many studies in the field of structural health monitoring have shown that it is not feasible to mine structural damage information simply from the data level. It is necessary to comprehensively mine from multiple aspects such as data and physics to improve identification accuracy.

[0136] This solution provides a physical-data dual-driven identification method. First, simulation data is obtained through a vehicle-track coupling dynamics model and measured data is obtained through a fiber Bragg grating accelerometer array system. The data is processed to establish a data set. Then, the correspondence between the damage signal and the physical characteristics of the damage is established based on the theoretical model. Based on physical drive, the display characteristics of the vibration signal are extracted from multiple aspects such as waveform, spectrum, transfer coefficient, wavelet packet energy spectrum, etc.; then, based on data drive, a Conv-1d convolutional neural network is used to mine information from the data level to extract an implicit feature matrix; then, the display feature vector is fused with the implicit feature matrix, and a steel spring damage identification model is built using a self-attention mechanism (the model is pre-trained with simulation data and fine-tuned with measured data in turn), and information such as the steel spring damage state and damage location in the vibration signal is mined from both display features and implicit features, providing a new method, new idea, and new perspective that helps to improve detection and identification efficiency and accuracy.

[0137] (2) The beneficial effect of obtaining simulation data through the vehicle-track three-dimensional spatial coupling dynamics model is that although the damage or breakage of steel springs in urban rail transit will endanger driving safety and have extremely serious consequences, this kind of disease is not common, and it is not easy to detect and identify, so it is extremely difficult to obtain a large number of high-quality steel spring damage track vibration acceleration data samples. In addition, data is only collected on a certain line, and the length of the plate, steel spring damage mode, etc. are relatively simple, and the data obtained are not representative. This solution, based on the vehicle-track three-dimensional spatial coupling dynamics model, sets different working conditions such as "damage state, damage location, vehicle speed, vehicle load", comprehensively simulates various real driving conditions, obtains rich simulation data samples, and solves the problem of scarcity of actual measured data on steel spring damage.

[0138] (3) The beneficial effect of this method is that, through model building and model training, based on big data operation analysis, the steel spring damage identification result is output, the mileage pile 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, and effectively reduce the workload of daily inspections and regular inspections by inspection and maintenance personnel, and help inspection and maintenance personnel to carry out special inspections in a targeted manner, repair and replace steel springs, and achieve targeted repair of steel springs.

[0139] In the description of this specification, the description with reference to the terms "one embodiment", "example", "specific example", etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representation of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner.

[0140] The above shows and describes 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, and the above embodiments and descriptions are only for explaining the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, and these changes and improvements all fall within the scope of the present invention to be protected.

Claims

1. A physical-data dual-drive rail spring damage detection and identification method, comprising the following steps: S1: obtain simulation data through the vehicle-track coupled dynamics model and obtain measured data through the fiber Bragg grating acceleration sensor array system; S2: Process the simulation data and measured data to establish a labeled data set; S3: Extract data-driven and physical-driven features of the simulation data and measured data based on the data set; The data-driven aspect uses Conv-1d convolutional neural network to mine information from the data level to extract implicit feature matrix; The physical drive aspect constructs a display feature vector through the waveform, spectrum, vibration acceleration transfer coefficient and wavelet packet energy spectrum of the vibration signal; S4: Fusion of displayed feature vector and implicit feature matrix; S5: Build a steel spring damage recognition model based on the attention mechanism and use the fused features for training; S6: Use the model to identify new floating slab track spring data.

2. A physical-data dual-drive rail spring damage detection and identification method according to claim 1, characterized in that: The step of acquiring the simulation data comprises: Construct a large simulation data set, use MATLAB to build a three-dimensional vehicle-track coupling dynamics model based on the vehicle-track coupling dynamics theory, and simulate the steel spring damage through the steel spring stiffness reduction coefficient; Set the damage state, damage location, track line condition, vibration response extraction point and sampling settings, which correspond to the percentage of spring support stiffness reduction, the relative position of the damaged spring on the floating slab, vehicle speed, vehicle load, vibration acceleration response at the same position of the floating slab and foundation, sampling frequency and sampling time respectively; The vehicle-track coupling dynamics model is run under various working conditions to obtain multiple sets of differentiated simulation raw data. Each set of simulation raw data includes a floating plate vibration acceleration data and a basic vibration acceleration data.

3. The physical-data dual-drive rail spring damage detection and identification method according to claim 1 is characterized in that: The measured data is obtained by building a fiber grating acceleration sensor array system.

4. The physical-data dual-drive rail spring damage detection and identification method according to claim 1 is characterized in that: In step S2, the data processing includes: Simulation data preprocessing, including data screening, building a time series data set and dividing it into training, validation and testing subsets, labeling and dividing it into sub-labels; Measured data preprocessing includes data screening, building a time series dataset and dividing it into training, validation, and testing subsets, labeling, and dividing it into sub-labels.

5. The physical-data dual-drive rail spring damage detection and identification method according to claim 1 is characterized in that: The feature extraction step comprises: The implicit features of the simulation data set are extracted through a one-dimensional convolutional neural network to construct an implicit feature matrix; Combining data-driven and physics-driven methods, the display characteristics of data are extracted from the perspectives of time domain, frequency domain, vibration transfer coefficient and wavelet packet energy spectrum; Integrate the extracted implicit and explicit features to form a feature vector; The same method is applied to extract features from the measured data set.

6. A physical-data dual-drive rail spring damage detection and identification method according to claim 5, characterized in that: The implicit feature extraction step comprises: Use the one-dimensional convolutional neural network architecture to mine hidden regular information from massive data, extract implicit features, and establish an implicit feature matrix; Use the processed data to perform a one-dimensional convolution operation using the Conv-1d function; For the ka and kb data of the dataset, one-dimensional convolution is used to project the basic data information into the high-dimensional feature space fa and fb respectively, and they are concatenated to form the corresponding implicit features.

7. The physical-data dual-drive rail spring damage detection and identification method according to claim 5 is characterized in that: The implicit feature extraction steps include: Waveform: extract characteristic values ​​from the waveform angle or time domain angle of the signal, and construct a 2D characteristic vector through the average peak value of vibration acceleration and the mean square value of vibration acceleration; Spectrum, extract signal features from the frequency domain, extract the first three main frequencies and amplitudes, and construct a 6-dimensional feature vector; Vibration acceleration transfer coefficient: Use vibration transfer coefficient as a characteristic index to analyze the steel spring damage information, extract the vibration transfer coefficient in the range of [1,200Hz], and construct a 200-dimensional feature vector; Wavelet packet energy spectrum, using wavelet energy spectrum as signal feature, extracts non-stationary and mutation information of vibration signal, and constructs 8-dimensional feature vector by decomposing vibration signal into 3 layers of wavelet packets and using 8-band wavelet energy spectrum of the 3rd layer as index; The display feature vector is constructed based on the physical drive, and the two obtained 216-dimensional display feature vectors are integrally spliced ​​to form a 432-dimensional display feature vector.

8. The physical-data dual-drive rail spring damage detection and identification method according to claim 1 is characterized in that: The specific steps of building the steel spring damage identification model include: Use the Z-score method to standardize the fusion feature matrix; Construct and concatenate the CLS vector to the fusion feature matrix to form the input matrix; Apply the self-attention mechanism processing layer to extract features through multiple Self-Attention calculations; The CLS head vector is processed using the fully connected layer and the softmax layer to output the damage state and location of the steel spring.

9. The physical-data dual-drive rail spring damage detection and identification method according to claim 1 is characterized in that: The training of the steel spring damage identification model consists of two stages: The first stage is model training, which includes using the simulated time series dataset to extract explicit features and implicit features, pre-training, and using the validation subset for cross-validation and parameter adjustment; The second stage is model fine-tuning, which includes feature extraction and data normalization preprocessing using the measured time series dataset, and pre-training the model again for transfer learning; Fine-tune the model using the validation and test subsets.

10. A physical-data dual-drive rail spring damage detection and identification system, characterized in that Includes the following components: Data acquisition unit: equipped with a vehicle-track coupled dynamics model to obtain simulation data, and equipped with a fiber Bragg grating acceleration sensor array system to collect measured data; Data processing unit: used to process simulation data and measured data and establish labeled data sets; Feature extraction unit: used to extract data-driven and physical-driven features of simulation data and measured data according to the data set; The data-driven aspect uses Conv-1d convolutional neural network to mine information from the data level to extract implicit feature matrix; The physical drive aspect constructs a display feature vector through the waveform, spectrum, vibration acceleration transfer coefficient and wavelet packet energy spectrum of the vibration signal; Feature fusion unit: used to fuse the displayed feature vector with the implicit feature matrix; Steel spring damage identification model unit: built based on the attention mechanism and trained using the fused features to identify new floating slab track steel spring data.

Images