Track slab deformation identification method based on track side vibration acceleration

Through the identification method based on track side vibration acceleration, combined with LMD and time-frequency domain feature extraction, and using the random forest model, the problem of deformation recognition of high-speed railway track plates is solved, and the rapid and accurate identification of track plate defects is achieved, ensuring train safety run.

CN113971421BActive Publication Date: 2024-11-08FUZHOU UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202111391821.9
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2021-11-19
Publication Date
2024-11-08
Estimated Expiration
2041-11-19

AI Technical Summary

Technical Problem

During long-term operation, high-speed railway track plates are prone to defects such as deformation, warping, and interlayer debonding, which leads to abnormal vibration of the wheel and rail system. Due to complex external factors, the damage is sudden and unpredictable, making it difficult to detect and Identify.

Method used

Using the identification method based on the orbital side vibration acceleration, combined with local mean decomposition (LMD) and time-frequency domain feature extraction, classification and identification are performed through a random forest model, and data is collected using optical fiber vibration acceleration sensors to perform data preprocessing and feature parameter extraction to achieve Quick and accurate identification of track plate deformation.

Benefits of technology

Effectively capture key information about track plate deformation, eliminate the influence of environmental noise, realize accurate identification of track plate defects, and ensure the safe operation of trains.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN113971421B_ABST
    Figure CN113971421B_ABST
Patent Text Reader

Abstract

The present invention relates to a track plate deformation identification method based on track side vibration acceleration, comprising the following steps: step S1: obtaining original track side vibration signal data; step S2: processing the track side vibration signal data and converting it into a data subset with the same time length; step S3: obtaining the characteristic parameter vector matrix of the original track side vibration signal data based on the local mean decomposition method; step S4: constructing a random forest model, and training based on the characteristic parameter vector matrix to obtain a trained random forest; step S5: obtaining real-time track side vibration signal data, and classifying according to the trained random forest to complete the identification of track plate defect information. The present invention can effectively extract key information for detecting track plate deformation, and quickly and accurately identify track plate deformation.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention relates to the field of track detection, and in particular to a track plate deformation identification method based on track side vibration acceleration. Background Art

[0002] High-speed railways are booming in various countries with their high efficiency, convenience, comfort and low maintenance workload. By the end of 2020, China's high-speed rail mileage will reach 38,000 kilometers, ranking first in the world. Ballastless track slabs are an important part of various track structures, and they play a role of "bridging the past and the future". However, long-term operation will cause various defects in the track slabs, such as warping deformation of the track slabs, damage to wide and narrow joints, interlayer debonding, and arch deformation of the track slabs. Under the action of temperature loads and high-frequency loads of trains, these damages may cause deformation of the track slabs and abnormal vibration of the wheel-rail system. Due to the complexity of external factors, the performance of the track slabs has the characteristics of temporal and spatial changes, so the occurrence of track slab damage is sudden and unpredictable. Timely detection of track slab defects can prevent the failure of the track slabs and ensure the safe operation of trains. Therefore, real-time monitoring of the performance of the track slabs and identification of the damage status of the track slabs are crucial for the long-term safe operation of high-speed railways. Summary of the invention

[0003] In view of this, the purpose of the present invention is to provide a track plate deformation identification method based on track side vibration acceleration, which combines LMD, time domain feature extraction and time-frequency domain feature extraction, can effectively extract key information for detecting track plate deformation, and quickly and accurately identify the deformation of the track plate, which has great advantages in practical applications.

[0004] To achieve the above object, the present invention adopts the following technical solution:

[0005] A method for identifying track plate deformation based on track side vibration acceleration comprises the following steps:

[0006] Step S1: obtaining original rail side vibration signal data;

[0007] Step S2: Processing the rail side vibration signal data and converting them into data subsets with the same time length;

[0008] Step S3: Based on the local mean decomposition method, the characteristic parameter vector matrix of the original rail side vibration signal data is obtained;

[0009] Step S4: construct a random forest model and train it based on the feature parameter vector matrix to obtain a trained random forest;

[0010] Step S5: Acquire real-time rail side vibration signal data, and classify it according to the trained random forest to complete the identification of track plate defect information.

[0011] Furthermore, the step S1 is specifically as follows: installing a fiber optic vibration acceleration sensor on the track side, connecting the fiber optic sensor to a data collector via an optical cable, filtering the measured data, and then converting the analog signal into a digital signal through an A / D conversion circuit to obtain the original track side vibration signal data.

[0012] Furthermore, the step S2 is specifically as follows: by means of data interception, wavelet threshold denoising and fixed window segmentation, the original rail side vibration signal data is converted into data subsets with the same time length.

[0013] Furthermore, the step S3 is specifically as follows:

[0014] Step S31: Calculate all local extreme points N of the original signal x(t) i , and derive all adjacent local extreme points N i and N i+1 The mean of :

[0015]

[0016] The corresponding t(N i ) and t(N i+1 ) between all mean points m i Connect them with a straight line to get the local mean line, and then use the sliding average method to smooth the local mean line to get the local mean function m 11 (t);

[0017] Step S32: From the adjacent local extreme values ​​N i and N i+1 Get the local amplitude a i :

[0018]

[0019] Similarly, the corresponding t(N i ) and t(N i+1 ) between all local amplitudes a i Connect with a straight line to get the local amplitude line, and then use the sliding average method to smooth the local amplitude line to get the envelope estimation function a 11 (t);

[0020] Step S33: The local mean function m 11 (t) is separated from the original signal x(t), and we get:

[0021] h 11 (t) = x(t) - m 11 (t)

[0022] Step S34: h 11(t) and the envelope estimation function a 11 (t) Phase division demodulation h 11 (t), we get:

[0023]

[0024] For 11 (t) Repeat steps S31 and S32 to obtain s 11 (t) envelope estimation function a 12 (t);

[0025] Repeat the above iterative process until a 12 (t) = 1, which means s 11 (t) is a frequency modulated signal;

[0026] Assume 1n (t) is obtained after n iterations, then s 1n (t) envelope estimation function a 1(n+1) (t) satisfies a 1(n+1) (t) = 1, the iterative process can be described as:

[0027] h 11 (t) = x(t) - m 11 (t)

[0028] h 12 (t) = s 11 (t)-m 12 (t)

[0029] h 1n (t) = s 1(n-1) (t)-m 1n (t)

[0030] Where:

[0031] s 11 (t) = h 12 (t) / a 11 (t)

[0032] s 12 (t) = h 12 (t) / a 12 (t)

[0033] s 1n (t) = h 1n (t) / a 1n (t)

[0034] The iteration termination condition is:

[0035]

[0036] Step S35: Multiply the envelope estimation functions obtained in the iterative process to obtain the envelope signal a1(t):

[0037]

[0038] Step S36: The first PF of the original signal is composed of the envelope signal a1(t) and the pure FM signal s 1n (t) and multiply to get:

[0039] PF1(t)=a1(t)s 1n (t)

[0040] Step S37: Separate PF1(t) from the original signal x(t) to obtain a residual signal u(t). The residual signal u(t) is repeated k times as the original signal until u k (t) is a monotonic function:

[0041] u1(t)=x(t)-PF1(t)

[0042] u2(t)=u1(t)-PF2(t) ...

[0044] u k (t) = u k-1 (t)-PF k (t)

[0045] At this point, the original signal x(t) is decomposed into k component signals and 1 residual function u k (t):

[0046]

[0047] Step S37: To quantify PF s The correlation with the original signal introduces the Pearson correlation coefficient:

[0048]

[0049] Where: X represents the original data, Y represents PF s ;

[0050] Step S38: Calculate the time domain characteristic parameters of the vibration signal;

[0051] Step S39: Calculate a time-frequency domain feature parameter of the vibration signal and select feature PF s The energy entropy is used as one of the damage identification features;

[0052] First, calculate the energy E of the qth feature PF q :

[0053]

[0054] Where: l is the length of the characteristic PF;

[0055] Calculate the total energy of these effective characteristic PFs:

[0056]

[0057] The energy entropy of the characteristic PF is:

[0058]

[0059] Where: is the percentage of the energy of the qth characteristic PF relative to the total energy entropy;

[0060] The five time domain features and one time-frequency domain feature extracted from the feature PF finally form a feature parameter vector matrix.

[0061] Furthermore, the time domain characteristic parameters include:

[0062] Peak-to-peak value: V p-p =max(PF q )-min(PF q )

[0063] variance:

[0064] RMS:

[0065] Form Factor:

[0066] Crest factor: C = max (PF q ) / PF rms

[0067] Where: PF q represents the qth feature PF, is the mean of feature PF, and r is the number of feature PF.

[0068] Furthermore, the step S4 is specifically as follows:

[0069] Step S41: randomly extract m from the original rail side vibration signal data with replacement using the Bootstrap resampling method tree A self-service sample set is used as the training data set D, and the unobtained data constitutes the OOB data set;

[0070] Step S42: extract the m tree The sample sets generate corresponding m tree A classification and regression tree;

[0071] Step S43: If there are t features, then randomly select m features at each node of each tree. try features, and use the information gain principle to calculate and select the feature with the largest information gain as the classification feature;

[0072] Step S44: Do not perform any pruning on the classification and regression trees, and let each tree grow to a preset limit;

[0073] Step S45: The generated multiple trees are organized into a random forest.

[0074] Furthermore, the step S43 is specifically as follows:

[0075] First calculate the information entropy:

[0076]

[0077] Among them, p i is the proportion of each classification feature i in the classification system.

[0078] The information gain ratio of feature A and training data set D is defined as the ratio of the information gain of feature A to the entropy of training data set D:

[0079]

[0080]

[0081]

[0082] Compared with the prior art, the present invention has the following beneficial effects:

[0083] The present invention can effectively capture vibration signals containing train vibration, track plate deformation, noise and environmental vibration information. Preprocessing methods such as data interception, wavelet threshold denoising, and fixed window segmentation can effectively eliminate the influence of environmental vibration, noise, time difference and other factors on the recognition effect. The time domain and time-frequency domain feature extraction methods based on LMD can effectively extract key information for detecting track plate deformation, and the deformation of the track plate can be accurately identified through the intelligent recognition algorithm of the random forest model. BRIEF DESCRIPTION OF THE DRAWINGS

[0084] Figure 1 It is a schematic diagram of the process of the present invention;

[0085] Figure 2 is a flow chart of LMD in one embodiment of the present invention;

[0086] Figure 3 It is the architecture of the random forest classification model in one embodiment of the present invention. DETAILED DESCRIPTION

[0087] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0088] Please refer to Figure 1 The present invention provides a method for identifying track plate deformation based on track side vibration acceleration, comprising the following steps:

[0089] Step S1: Track side vibration signal collection. Install a fiber optic vibration acceleration sensor on the track side. The fiber optic sensor is connected to a data collector via an optical cable. The data collector filters the measured data and then converts the analog signal into a digital signal through an A / D conversion circuit. Finally, the embedded system of the collector sends the collected vibration data to the host computer via a wireless network.

[0090] Step S2: Vibration data preprocessing: The original data is converted into data subsets with the same time length through data interception, wavelet threshold denoising, and fixed window segmentation;

[0091] Step S3: Extract features in the time domain and time-frequency domain. The frequency modulation signal and the envelope signal are separated from the original signal by the local mean decomposition method (LMD), and these signals are multiplied to obtain an instantaneous frequency with the physical meaning of the product function (PFs); iterate until all product functions (PFs) are separated. Finally, the time-frequency distribution of the original signal is obtained;

[0092] Step S4: Establish a random forest model architecture to identify track slab defect information.

[0093] Preferably, in this embodiment, step S2 is specifically:

[0094] Step S21: When the train passes through the sensor, the information captured by the sensor is the wheel-rail vibration signal containing the track plate deformation information, so the original data needs to be intercepted.

[0095] Define the vibration acceleration threshold δ. Here, the vibration acceleration values ​​collected by the sensor when the train is not passing and when the train is passing are compared to set the vibration acceleration threshold δ to 0.01g.

[0096] When the vibration acceleration x(t) ≥ δ at time t, record x(t) at this moment. Also record x(t+1), x(t+2), and x(t+3) until x(t+n-1) ≥ δ, x(t+n) ≤ δ, and x(t+n+1) < δ;

[0097] Finally, we get a vibration acceleration data set with environmental vibration information eliminated: X = [x(t), x(t+1), x(t+2), ... x(t+n-1)].

[0098] Step S22: De-noise the new vibration acceleration data set. Use the wavelet decomposition method to remove high-frequency noise in the vibration acceleration data set. Use the "sym3" wavelet function to perform five-layer wavelet threshold denoising on the new vibration acceleration data set.

[0099] Step S23: Perform fixed window segmentation on the denoised signal. When trains of different speeds pass through the sensor, the duration of the vibration signal is different. In order to eliminate the influence of time difference on the recognition effect, a fixed window is used to perform data segmentation on the denoised signal.

[0100] In this embodiment, preferably, the fixed window length is equal to the sampling frequency: L=f, and the duration is set to 1s. If the duration of a piece of data is 6s and 4s respectively, it can be divided into 6 and 4 data subsets respectively.

[0101] In this implementation, step S3 specifically includes the following steps:

[0102] Step S31: First calculate all local extreme value points N of the original signal x(t) i , and derive all adjacent local extreme points N i and N i+1 The mean of :

[0103]

[0104] The corresponding t(N i ) and t(N i+1 ) between all mean points m i Connect them with a straight line to get the local mean line, and then use the sliding average method to smooth the local mean line to get the local mean function m 11 (t).

[0105] Step S32: From the adjacent local extreme values ​​N i and N i+1 Get the local amplitude a i :

[0106]

[0107] Similarly, the corresponding t(N i ) and t(N i+1 ) between all local amplitudes a i Connect with a straight line to get the local amplitude line, and then use the sliding average method to smooth the local amplitude line to get the envelope estimation function a 11 (t).

[0108] Step S33: The local mean function m 11 (t) is separated from the original signal x(t), we can get:

[0109] h 11 (t) = x(t) - m 11 (t)

[0110] Step S34: h 11 (t) and the envelope estimation function a 11 (t) Phase division demodulation h 11 (t), we can get:

[0111]

[0112] For 11 (t) Repeat steps S31 and S32 to obtain s 11 (t) envelope estimation function a 12 (t). Repeat the above iterative process until a 12 (t) = 1, which means s 11 (t) is a frequency modulated signal. Assume that s 1n (t) is obtained after n iterations, then s 1n (t) envelope estimation function a 1(n+1) (t) satisfies a 1(n+1) (t) = 1, the iterative process can be described as:

[0113] h 11 (t) = x(t) - m 11 (t)

[0114] h 12 (t) = s 11 (t)-m 12 (t)

[0115] h 1n (t) = s 1(n-1) (t)-m 1n (t)

[0116] Where:

[0117] s 11 (t) = h 12 (t) / a 11 (t)

[0118] s 12 (t) = h 12 (t) / a 12 (t)

[0119] s 1n (t) = h 1n (t) / a 1n (t)

[0120] The iteration termination condition is:

[0121]

[0122] Step S35: Multiply the envelope estimation functions obtained in the iterative process to obtain the envelope signal a1(t):

[0123]

[0124] Step S36: The first PF of the original signal is composed of the envelope signal a1(t) and the pure FM signal s 1n (t) and multiply to get:

[0125] PF1(t)=a1(t)s 1n (t)

[0126] Step S37: Separate PF1(t) from the original signal x(t) to obtain a residual signal u(t). The residual signal u(t) is repeated k times as the original signal until u k (t) is a monotonic function:

[0127] u1(t)=x(t)-PF1(t)

[0128] u2(t)=u1(t)-PF2(t) ...

[0130] u k (t) = u k-1 (t)-PF k (t)

[0131] At this point, the original signal x(t) is decomposed into k component signals and 1 residual function u k (t):

[0132]

[0133] Step S37: To quantify PF s The correlation with the original signal introduces the Pearson correlation coefficient:

[0134]

[0135] Where: X represents the original data, Y represents PF s .

[0136] When the correlation between the two variables increases, the Pearson correlation coefficient tends to -1 or 1, and when the correlation between the two variables decreases, the Pearson correlation coefficient tends to 0. If the correlation coefficient between the PF component and the original data is less than or equal to 0.2, the PF component is discarded, and the retained PF is called the characteristic PF.

[0137] Step S38: Calculate five time domain characteristic parameters of the vibration signal.

[0138] Peak-to-peak value: V p-p =max(PF q )-min(PF q )

[0139] variance:

[0140] RMS:

[0141] Form Factor:

[0142] Crest factor: C = max (PF q ) / PF rms

[0143] Where: PF q represents the qth feature PF, is the mean of feature PF, and r is the number of feature PF.

[0144] Step S39: Calculate a time-frequency domain feature parameter of the vibration signal and select feature PF s The energy entropy is used as one of the damage identification features.

[0145] First, calculate the energy E of the qth feature PF q :

[0146]

[0147] Where: l is the length of the characteristic PF.

[0148] Calculate the total energy of these effective characteristic PFs:

[0149]

[0150] The energy entropy of the characteristic PF is:

[0151]

[0152] Where: is the percentage of the energy of the qth feature PF relative to the total energy entropy, and

[0153] The 5 time domain features and 1 time-frequency domain feature extracted from the feature PF finally form a feature parameter vector matrix. Therefore, 6 features can be extracted from the data set.

[0154] In this embodiment, step S4 specifically includes the following steps:

[0155] Step S41: From the original training data set, randomly select m with replacement using the Bootstrap resampling method tree A self-service sample set is used as the training data set D, and the unobtained data constitute the OOB data set.

[0156] Step S42: extract the m tree The sample sets generate corresponding m tree In the present invention, m is selected according to the classification accuracy. tree is 1000.

[0157] Step S43: If there are t features, then randomly select m features at each node of each tree. try Features

[0158] Generally speaking, split characteristics Preferably, in this embodiment, t=6, so let m try =3, by calculating the amount of information contained in each feature, in m try Select a feature with the best classification ability from the features to split the node.

[0159] The information gain principle is used to calculate and select the feature with the largest information gain as the classification feature.

[0160] First calculate the information entropy:

[0161]

[0162] Among them, p i is the proportion of each classification feature i in the classification system.

[0163] The information gain ratio of feature A and training data set D is defined as the ratio of the information gain of feature A to the entropy of training data set D:

[0164]

[0165]

[0166]

[0167] Step S44: Do not perform any pruning on the classification and regression trees, and let each tree grow to its maximum limit. The maximum limit here means that there is only one sample point in all child nodes.

[0168] Step S45: The generated multiple trees are organized into a random forest, and the data of the test set is classified using the random forest. The classification result is determined by the number of votes of the tree classifier, and finally the identification of the track plate defect information is completed.

[0169] The above description is only a preferred embodiment of the present invention. All equivalent changes and modifications made according to the scope of the patent application of the present invention should fall within the scope of the present invention.

Claims

1. A track plate deformation identification method based on track side vibration acceleration, It is characterized in that The following steps are involved: Step S1: obtaining original rail side vibration signal data; Step S2: Processing the rail side vibration signal data and converting them into data subsets with the same time length; Step S3: Based on the local mean decomposition method, the characteristic parameter vector matrix of the original rail side vibration signal data is obtained; Step S4: construct a random forest model and train it based on the feature parameter vector matrix to obtain a trained random forest; Step S5: acquiring real-time rail side vibration signal data, and classifying it according to the trained random forest to complete the identification of track plate defect information; The step S2 specifically includes: converting the original rail side vibration signal data into data subsets with the same time length by means of data interception, wavelet threshold denoising, and fixed window segmentation; The step S3 is specifically as follows: Step S31: Calculate all local extreme points N of the original signal x(t) i , and derive all adjacent local extreme points N i and N i+1 The mean of : Will correspond to N i With N i+1 All mean points between i Connect them with a straight line to get the local mean line, and then use the sliding average method to smooth the local mean line to get the local mean function m 11 (t); Step S32: From the adjacent local extreme values ​​N i and N i+1 Get the local amplitude a i : Similarly, the corresponding N i With N i+1 All local amplitudes a between i Connect with a straight line to get the local amplitude line, and then use the sliding average method to smooth the local amplitude line to get the envelope estimation function a 11 (t); Step S33: The local mean function m 11 (t) is separated from the original signal x(t), and we get: h 11 (t)=x(t)-m 11 (t) Step S34: h 11 (t) and the envelope estimation function a 11 (t) Phase division demodulation h 11 (t), we get: For 11 (t) Repeat steps S31 and S32 to obtain s 11 (t) envelope estimation function a 12 (t); until a 12 (t) = 1, which means s 11 (t) is a frequency modulated signal; Assume 1n (t) is obtained after n iterations, then s 1n (t) envelope estimation function a 1(n+1) (t) satisfies a 1(n+1) (t) = 1, the iterative process can be described as: h 11 (t)=x(t)-m 11 (t) h 12 (t)=s 11 (t)-m 12 (t) h 1n (t)=s 1(n-1) (t)-m 1n (t) Where: s 11 (t)=h 12 (t) / a 11 (t) s 12 (t)=h 12 (t) / a 12 (t) s 1n (t)=h 1n (t) / a 1n (t) The iteration termination condition is: Step S35: Multiply the envelope estimation functions obtained in the iterative process to obtain the envelope signal a 1 (t): Step S36: The first PF of the original signal is composed of the envelope signal a 1 (t) and pure FM signal s 1n (t) and multiply to get: PF 1 (t)=a 1 (t)s 1n (t) Step S37: PF 1 (t) is separated from the original signal x(t) to obtain the residual signal u 1 (t), residual signal u 1 (t) is repeated k times as the original signal until u k (t) is a monotonic function: you 1 (t)=x(t)-PF 1 (t) you 2 (t)=u 1 (t)-PF 2 (t) ... you k (t)=u k-1 (t)-PF k (t) At this point, the original signal x(t) is decomposed into k component signals and 1 residual function u k (t): Step S37: In order to quantify the correlation between PF and the original signal, the Pearson correlation coefficient is introduced: Where: X represents the original data, Y represents PF; Step S38: Calculate the time domain characteristic parameters of the vibration signal; Step S39: Calculate a time-frequency domain feature parameter of the vibration signal and select feature PF s The energy entropy is used as one of the damage identification features; First, calculate the energy E of the qth feature PF q : Where: l is the length of the characteristic PF; Calculate the total energy of these effective characteristic PFs: The energy entropy of the characteristic PF is: Where: is the percentage of the energy of the qth characteristic PF relative to the total energy entropy; The 5 time domain features and 1 time-frequency domain feature extracted from the feature PF finally form a feature parameter vector matrix; The time domain characteristic parameters include: Peak-to-peak value: V p-p =max(PF q )-min(PF q ) variance: RMS: Form Factor: Crest factor: C = max(PF q ) / PF rms Where: PF q represents the qth feature PF, is the mean of feature PF, and r is the number of feature PF.

2. The track plate deformation identification method based on track side vibration acceleration according to claim 1, It is characterized in that The step S1 specifically includes: installing a fiber optic vibration acceleration sensor on the track side, connecting the fiber optic sensor to a data collector via an optical cable, filtering the measured data, and then converting the analog signal into a digital signal through an A / D conversion circuit to obtain the original track side vibration signal data.

3. The track plate deformation identification method based on track side vibration acceleration according to claim 1, It is characterized in that The step S4 is specifically as follows: Step S41: randomly extract m from the original rail side vibration signal data with replacement using the Bootstrap resampling method tree A self-service sample set is used as the training data set D, and the unobtained data constitutes the OOB data set; Step S42: extract the m tree The sample sets generate corresponding m tree A classification and regression tree; Step S43: If there are t features, then randomly select m features at each node of each tree. try features, and use the information gain principle to calculate and select the feature with the largest information gain as the classification feature; Step S44: Do not perform any pruning on the classification and regression trees, and let each tree grow to a preset limit; Step S45: The generated multiple trees are organized into a random forest.

4. The track plate deformation identification method based on track side vibration acceleration according to claim 3, It is characterized in that The step S43 is specifically as follows: First calculate the information entropy: Among them, p i is the proportion of each classification feature i in the classification system; The information gain ratio of feature A and training data set D is defined as the ratio of the information gain of feature A to the entropy of training data set D: