CN113971421A - Track slab deformation identification method based on track side vibration acceleration - Google Patents

Abstract

The invention relates to a track slab deformation identification method based on track side vibration acceleration, which comprises the following steps of: step S1, acquiring original rail side vibration signal data; step S2, processing the vibration signal data of the rail side to convert the vibration signal data into a data subset with the same time length; s3, acquiring a characteristic parameter vector matrix of the original rail side vibration signal data based on a local mean decomposition method; s4, constructing a random forest model, and training based on the characteristic parameter vector matrix to obtain a trained random forest; and step S5, acquiring real-time rail side vibration signal data, classifying according to the trained random forest, and completing the identification of the track slab defect information. The method can effectively extract the key information for detecting the deformation of the track slab, and quickly and accurately identify the deformation of the track slab.

Description

Track slab deformation identification method based on track side vibration acceleration

Technical Field

The invention relates to the field of rail detection, in particular to a rail plate deformation identification method based on rail side vibration acceleration.

Background

The high-speed railway is developed vigorously in various countries due to the characteristics of high efficiency, convenience, comfort and low maintenance workload. By the end of 2020, the mileage of the Chinese high-speed rail reaches 3.8 kilometers and is the first place in the world. The ballastless track slab is an important component of various track structures and plays a role of bearing from top to bottom. However, the long-term operation can cause various defects of the track slab, such as warp deformation of the track slab, damage of wide and narrow joints, interlayer debonding, arch deformation of the track slab and the like. Under temperature loads and high frequency train loads, these damages may cause deformation of the rail plate and abnormal vibration of the wheel rail system. Due to the fact that external factors are complex, the performance of the track slab has the characteristic of space-time change, and therefore the occurrence of track slab damage is sudden and unpredictable. The defects of the track slabs are detected in time, the track slabs can be prevented from being out of work, and the safe operation of the train is guaranteed. Therefore, the performance of the track slab is monitored in real time, and the damage state of the track slab is identified, so that the method is important for long-term safe operation of the high-speed rail.

Disclosure of Invention

In view of this, the present invention provides a track slab 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 slab deformation, quickly and accurately identify track slab deformation, and has great advantages in practical applications.

In order to achieve the purpose, the invention adopts the following technical scheme:

a rail plate deformation identification method based on rail side vibration acceleration comprises the following steps:

step S1, acquiring original rail side vibration signal data;

step S2, processing the vibration signal data of the rail side to convert the vibration signal data into a data subset with the same time length;

s3, acquiring a characteristic parameter vector matrix of the original rail side vibration signal data based on a local mean decomposition method;

s4, constructing a random forest model, and training based on the characteristic parameter vector matrix to obtain a trained random forest;

and step S5, acquiring real-time rail side vibration signal data, classifying according to the trained random forest, and completing the identification of the track slab defect information.

Further, the step S1 is specifically: and an optical fiber vibration acceleration sensor is arranged on the rail side, the optical fiber sensor is connected to a data acquisition unit through an optical cable, the data acquisition unit filters the measured data, and the analog signals are converted into digital signals through an A/D conversion circuit to obtain the original rail side vibration signal data.

Further, the step S2 is specifically: original rail side vibration signal data are converted into data subsets with the same time length through methods of data interception, wavelet threshold denoising and fixed window segmentation.

Further, the step S3 is specifically:

step S31, calculating all local extreme points N of the original signal x (t)_iAnd deducing all adjacent local extreme points N_iAnd N_i+1Average value of (d):

will correspond to t (N)_i) And t (N)_i+1) All mean value points m in between_iConnecting with a straight line to obtain a local mean line, and smoothing the local mean line by using a sliding average method to obtain a local mean function m₁₁(t)；

Step S32: from adjacent local extrema N_iAnd N_i+1Obtain the local amplitude a_i:

In the same way, corresponding t (N)_i) And t (N)_i+1) All local amplitudes a in between_iConnecting with a straight line to obtain a local amplitude line, and smoothing the local amplitude line by using a moving average method to obtain an envelope estimation function a₁₁(t)；

Step S33: the local mean function m₁₁(t) separating from the original signal x (t) to obtain:

h₁₁(t)＝x(t)-m₁₁(t)

step S34: h is to be₁₁(t) and an envelope estimation function a₁₁(t) phase division demodulation h₁₁(t), obtaining:

to s₁₁(t) repeating the steps of S31 and S32 to obtain S₁₁Envelope estimation function a of (t)₁₂(t)；

Repeating the above iterative process until a₁₂(t) 1, in which case s is specified₁₁(t) is a frequency modulated signal;

let s_1n(t) is obtained by n iterations, then s_1nEnvelope estimation function a of (t)_1(n+1)(t) satisfies a_1(n+1)Where (t) is 1, the iterative process can be described as:

h₁₁(t)＝x(t)-m₁₁(t)

h₁₂(t)＝s₁₁(t)-m₁₂(t)

h_1n(t)＝s_1(n-1)(t)-m_1n(t)

in the formula:

s₁₁(t)＝h₁₂(t)/a₁₁(t)

s₁₂(t)＝h₁₂(t)/a₁₂(t)

s_1n(t)＝h_1n(t)/a_1n(t)

the iteration termination condition is as follows:

step S35: multiplying the envelope estimation function obtained in the iteration process to obtain an envelope signal a₁(t)：

Step S36: the first PF of the original signal is derived from the envelope signal a₁(t) with a pure frequency-modulated signal s_1n(t) the multiplication yields:

PF₁(t)＝a₁(t)s_1n(t)

step S37: mixing PF₁(t) separating from the original signal x (t) to obtain a residual signal u (t), and repeating the residual signal u (t) as the original signal k times until u_k(t) is a monotonic function:

u₁(t)＝x(t)-PF₁(t)

u₂(t)＝u₁(t)-PF₂(t)

...

u_k(t)＝u_k-1(t)-PF_k(t)

at this time, the original signal x (t) is decomposed into k component signals and 1 residual function u_k(t)：

Step S37: to quantify PF_sCorrelation degree with original signals, introducing Pearson correlation coefficient:

in the formula: x represents raw data and Y represents PF_s；

Step S38: calculating time domain characteristic parameters of the vibration signals;

step S39: calculating 1 time-frequency domain characteristic parameter of vibration signal, selecting characteristic PF_sAs one of the impairment recognition features;

first, the energy E of the qth characteristic PF is calculated_q：

In the formula: l is the length of the characteristic PF;

the total energy of these valid signatures PF is calculated:

the energy entropy of the feature PF is:

in the formula:

is the percentage of the energy of the qth characteristic PF with respect to the total energy entropy;

and finally forming a characteristic parameter vector matrix through 5 time domain characteristics and 1 time-frequency domain characteristic extracted from the characteristic PF.

Further, the time-domain feature parameters include:

peak-to-peak: v_p-p＝max(PF_q)-min(PF_q)

Variance:

root mean square:

shape factor:

crest factor: c ═ max (PF)_q)/PF_rms

In the formula: PF (particle Filter)_qRepresents the q-th characteristic PF,

the average value of the characteristic PFs is shown, and r is the number of the characteristic PFs.

Further, the step S4 is specifically:

step S41: randomly extracting m from original rail side vibration signal data in a releasing way by a Bootstrap resampling method_treeThe self-help sample set is used as a training data set D, and data which are not obtained form an OOB data set;

step S42: for the extracted m_treeGenerating corresponding m respectively for each sample set_treeClassifying and returning trees;

step S43: with t features, m is randomly drawn at each node of each tree_tryCalculating and selecting the characteristic with the maximum information gain as a classification characteristic by adopting an information gain principle;

step S44: the classification regression tree is not cut, and each tree is made to grow to a preset limit;

step S45: and forming a random forest by the generated trees.

Further, the step S43 is specifically:

firstly, calculating the information entropy:

wherein p is_iThe proportion of each classification characteristic i in the classification system.

Defining the information gain ratio of the feature a to the training data set D as the ratio of the information gain of the feature a to the entropy of the training data set D:

compared with the prior art, the invention has the following beneficial effects:

the method can effectively capture vibration signals containing train vibration, track slab deformation, noise and environmental vibration information, can effectively eliminate the influence of factors such as environmental vibration, noise, time difference and the like on the identification effect by adopting preprocessing methods such as data interception, wavelet threshold denoising, fixed window segmentation and the like, can effectively extract key information for detecting track slab deformation by using the time domain and time frequency domain feature extraction method based on LMD, and can accurately identify the deformation of the track slab through an intelligent identification algorithm of a random forest model.

Drawings

FIG. 1 is a schematic flow diagram of the present invention;

FIG. 2 is a flow chart of an LMD in an embodiment of the present invention;

FIG. 3 is a diagram of a random forest classification model according to an embodiment of the present invention.

Detailed Description

The invention is further explained below with reference to the drawings and the embodiments.

Referring to fig. 1, the present invention provides a track slab deformation identification method based on track side vibration acceleration, including the following steps:

step S1: and (5) acquiring a rail side vibration signal. An optical fiber vibration acceleration sensor is arranged on the side of the track, the optical fiber sensor is connected to a data acquisition unit through an optical cable, the data acquisition unit filters measured data, analog signals are converted into digital signals through an A/D conversion circuit, and finally an embedded system of the acquisition unit sends the acquired vibration data to an upper computer through a wireless network;

step S2: and preprocessing vibration data. Converting original data into a data subset with the same time length by methods of data interception, wavelet threshold denoising and fixed window segmentation;

step S3: and extracting time domain and time-frequency domain characteristics. Separating the frequency modulation signal and the envelope signal from the original signal by a Local Mean Decomposition (LMD) method, and multiplying the signals to obtain an instantaneous frequency with the physical meaning of a Product Function (PFs); iterate until all Product Functions (PF) are separated. Finally, obtaining the time-frequency distribution of the original signal;

step S4: and establishing a random forest model architecture to identify the track slab defect information.

Preferably, in this embodiment, step S2 specifically includes:

and step S21, when the train passes through the sensor, the information captured by the sensor is the wheel track vibration signal containing the track slab deformation information, so that the original data needs to be intercepted.

Defining a vibration acceleration threshold delta, wherein the vibration acceleration threshold delta is set to be 0.01g by comparing vibration acceleration values acquired by a sensor when a train does not pass through and passes through;

when the vibration acceleration x (t) at the time t is larger than or equal to delta, recording the x (t) at the time. Recording x (t +1), x (t +2) and x (t +3) until x (t + n-1) is more than or equal to delta, x (t + n) is less than or equal to delta, and x (t + n +1) is less than delta;

and finally obtaining a vibration acceleration data set with the environment vibration information eliminated: x ═ X (t), X (t +1), X (t +2),. X (t + n-1) ].

Step S22: and denoising the new vibration acceleration set. And removing high-frequency noise in the vibration acceleration data set by using a wavelet decomposition method. Five-layer wavelet threshold denoising is carried out on the new vibration acceleration set by adopting a 'sym 3' wavelet function.

Step S23: and carrying out fixed window segmentation on the denoised signal. The duration of the vibration signal is different when trains of different speeds pass the sensor. In order to eliminate the influence of the time difference on the identification effect, a fixed window is adopted to carry out data segmentation on the de-noised signal.

In this embodiment, it is preferable to make the value of the fixed window length equal to the value of the sampling frequency: l ═ f, and the duration is set to 1 s. If the duration of a piece of data is 6s and 4s, respectively, it can be divided into 6 and 4 data subsets, respectively.

In this embodiment, step S3 specifically includes the following steps:

step S31, first calculate all local extreme points N of the original signal x (t)_iAnd deducing all adjacent local extreme points N_iAnd N_i+1Average value of (d):

will correspond to t (N)_i) And t (N)_i+1) All mean value points m in between_iConnecting with a straight line to obtain a local mean line, and smoothing the local mean line by using a sliding average method to obtain a local mean function m₁₁(t)。

Step S32: from adjacent local extrema N_iAnd N_i+1Obtain the local amplitude a_i:

In the same way, corresponding t (N)_i) And t (N)_i+1) All local amplitudes a in between_iConnecting with a straight line to obtain a local amplitude line, and smoothing the local amplitude line by using a moving average method to obtain an envelope estimation function a₁₁(t)。

Step S33: the local mean function m₁₁(t) from the original signal x (t), we can obtain:

h₁₁(t)＝x(t)-m₁₁(t)

step S34: h is to be₁₁(t) and an envelope estimation function a₁₁(t) phase division demodulation h₁₁(t), it is possible to obtain:

to s₁₁(t) repeating the steps of S31 and S32 to obtain S₁₁Envelope estimation function a of (t)₁₂(t) of (d). Repeating the above iterative process until a₁₂(t) 1, in which case s is specified₁₁(t) is a frequency modulated signal. Suppose s_1n(t) is obtained by n iterations, then s_1nEnvelope estimation function a of (t)_1(n+1)(t) satisfies a_1(n+1)Where (t) is 1, the iterative process can be described as:

h₁₁(t)＝x(t)-m₁₁(t)

h₁₂(t)＝s₁₁(t)-m₁₂(t)

h_1n(t)＝s_1(n-1)(t)-m_1n(t)

in the formula:

s₁₁(t)＝h₁₂(t)/a₁₁(t)

s₁₂(t)＝h₁₂(t)/a₁₂(t)

s_1n(t)＝h_1n(t)/a_1n(t)

the iteration termination condition is as follows:

step S35: multiplying the envelope estimation function obtained in the iteration process to obtain an envelope signal a₁(t)：

Step S36: the first PF of the original signal is derived from the envelope signal a₁(t) with a pure frequency-modulated signal s_1n(t) the multiplication yields:

PF₁(t)＝a₁(t)s_1n(t)

step S37: mixing PF₁(t) separating from the original signal x (t) to obtain a residual signal u (t), and repeating the residual signal u (t) as the original signal k times until u_k(t) is a monotonic function:

u₁(t)＝x(t)-PF₁(t)

u₂(t)＝u₁(t)-PF₂(t)

...

u_k(t)＝u_k-1(t)-PF_k(t)

at this time, the original signal x (t) is decomposed into k component signals and 1 residual function u_k(t)：

Step S37: to quantify PF_sCorrelation degree with original signals, introducing Pearson correlation coefficient:

in the formula: x represents raw data and Y represents PF_s。

The Pearson correlation coefficient tends to-1 or 1 when the correlation of the two variables increases, and tends to 0 when the correlation of the two variables decreases. If the correlation coefficient of the PF component with the original data is less than or equal to 0.2, the PF component is discarded and the remaining PF is called a feature PF.

Step S38: 5 time-domain characteristic parameters of the vibration signal are calculated.

Peak-to-peak: v_p-p＝max(PF_q)-min(PF_q)

Variance:

root mean square:

shape factor:

crest factor: c ═ max (PF)_q)/PF_rms

In the formula: PF (particle Filter)_qRepresents the q-th characteristic PF,

the average value of the characteristic PFs is shown, and r is the number of the characteristic PFs.

Step S39: calculating 1 time-frequency domain characteristic parameter of vibration signal, selecting characteristic PF_sAs one of the impairment recognition features.

First, the energy E of the qth characteristic PF is calculated_q：

In the formula: l is the length of the characteristic PF.

The total energy of these valid signatures PF is calculated:

the energy entropy of the feature PF is:

in the formula:

is the percentage of the energy of the qth characteristic PF with respect to the total energy entropy, and

and finally forming a characteristic parameter vector matrix through 5 time domain characteristics and 1 time-frequency domain characteristic extracted from the characteristic PF. Thus, 6 features can be extracted from the dataset.

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

step S41: from the original training dataset, m is randomly drawn with a back-put by the Bootstrap resampling method_treeAnd (4) forming an OOB data set by using the self-help sample set as a training data set D and the data which is not obtained.

Step S42: for the extracted m_treeGenerating corresponding m respectively for each sample set_treeA classification regression tree. In the invention, m is taken according to classification precision_treeIs 1000.

Step S43: with t features, m is randomly drawn at each node of each tree_tryA characteristic

In general, the splitting feature

Preferably, in this embodiment, t is 6, so m is set_tryBy calculating the amount of information each feature contains, at m, 3_tryAnd selecting one feature with the most classification capability from the features to perform node splitting.

And calculating and selecting the feature with the maximum information gain as the classification feature by adopting an information gain principle.

Firstly, calculating the information entropy:

wherein p is_iThe proportion of each classification characteristic i in the classification system.

Defining the information gain ratio of the feature a to the training data set D as the ratio of the information gain of the feature a to the entropy of the training data set D:

step S44: and (4) not cutting the classification regression tree, and enabling each tree to grow to the maximum. The maximum limit here means that there is only one sample point in all child nodes.

Step S45: and forming a random forest by the generated trees, classifying the data of the test set by using the random forest, determining the classification result according to the voting amount of the tree classifier, and finally finishing the identification of the track slab defect information.

The above description is only a preferred embodiment of the present invention, and all equivalent changes and modifications made in accordance with the claims of the present invention should be covered by the present invention.

Claims (7)

1. A track slab deformation identification method based on track side vibration acceleration is characterized by comprising the following steps:

step S1, acquiring original rail side vibration signal data;

step S2, processing the vibration signal data of the rail side to convert the vibration signal data into a data subset with the same time length;

s3, acquiring a characteristic parameter vector matrix of the original rail side vibration signal data based on a local mean decomposition method;

s4, constructing a random forest model, and training based on the characteristic parameter vector matrix to obtain a trained random forest;

and step S5, acquiring real-time rail side vibration signal data, classifying according to the trained random forest, and completing the identification of the track slab defect information.

2. The track slab deformation identification method based on the track-side vibration acceleration according to claim 1, wherein the step S1 is specifically: and an optical fiber vibration acceleration sensor is arranged on the rail side, the optical fiber sensor is connected to a data acquisition unit through an optical cable, the data acquisition unit filters the measured data, and the analog signals are converted into digital signals through an A/D conversion circuit to obtain the original rail side vibration signal data.

3. The track slab deformation identification method based on the track-side vibration acceleration according to claim 1, wherein the step S2 is specifically: original rail side vibration signal data are converted into data subsets with the same time length through methods of data interception, wavelet threshold denoising and fixed window segmentation.

4. The track slab deformation identification method based on the track-side vibration acceleration according to claim 1, wherein the step S3 is specifically:

step S31, calculating all local extreme points N of the original signal x (t)_iAnd deducing all adjacent local extreme points N_iAnd N_i+1Average value of (d):

will correspond to t (N)_i) And t (N)_i+1) All mean value points m in between_iConnecting with a straight line to obtain a local mean line, and smoothing the local mean line by using a sliding average method to obtain a local mean function m₁₁(t)；

Step S32: from adjacent local extrema N_iAnd N_i+1Obtain the local amplitude a_i:

In the same way, corresponding t (N)_i) And t (N)_i+1) All local amplitudes a in between_iConnecting with a straight line to obtain a local amplitude line, and smoothing the local amplitude line by using a moving average method to obtain an envelope estimation function a₁₁(t)；

Step S33: the local mean function m₁₁(t) with aSeparation of the start signal x (t) yields:

h₁₁(t)＝x(t)-m₁₁(t)

step S34: h is to be₁₁(t) and an envelope estimation function a₁₁(t) phase division demodulation h₁₁(t), obtaining:

to s₁₁(t) repeating the steps of S31 and S32 to obtain S₁₁Envelope estimation function a of (t)₁₂(t)；

Repeating the above iterative process until a₁₂(t) 1, in which case s is specified₁₁(t) is a frequency modulated signal;

let s_1n(t) is obtained by n iterations, then s_1nEnvelope estimation function a of (t)_1(n+1)(t) satisfies a_1(n+1)Where (t) is 1, the iterative process can be described as:

h₁₁(t)＝x(t)-m₁₁(t)

h₁₂(t)＝s₁₁(t)-m₁₂(t)

h_1n(t)＝s_1(n-1)(t)-m_1n(t)

in the formula:

s₁₁(t)＝h₁₂(t)/a₁₁(t)

s₁₂(t)＝h₁₂(t)/a₁₂(t)

s_1n(t)＝h_1n(t)/a_1n(t)

the iteration termination condition is as follows:

step S35: multiplying the envelope estimation function obtained in the iteration process to obtain an envelope signal a₁(t)：

Step S36: the first PF of the original signal is derived from the envelope signal a₁(t) with a pure frequency-modulated signal s_1n(t) the multiplication yields:

PF₁(t)＝a₁(t)s_1n(t)

step S37: mixing PF₁(t) separating from the original signal x (t) to obtain a residual signal u (t), and repeating the residual signal u (t) as the original signal k times until u_k(t) is a monotonic function:

u₁(t)＝x(t)-PF₁(t)

u₂(t)＝u₁(t)-PF₂(t)

...

u_k(t)＝u_k-1(t)-PF_k(t)

at this time, the original signal x (t) is decomposed into k component signals and 1 residual function u_k(t)：

Step S37: to quantify PF_sCorrelation degree with original signals, introducing Pearson correlation coefficient:

in the formula: x represents raw data and Y represents PF_s；

Step S38: calculating time domain characteristic parameters of the vibration signals;

step S39: calculating 1 time-frequency domain characteristic parameter of vibration signal, selecting characteristic PF_sAs one of the impairment recognition features;

first, the energy E of the qth characteristic PF is calculated_q：

In the formula: l is the length of the characteristic PF;

the total energy of these valid signatures PF is calculated:

the energy entropy of the feature PF is:

in the formula:

is the percentage of the energy of the qth characteristic PF with respect to the total energy entropy;

and finally forming a characteristic parameter vector matrix through 5 time domain characteristics and 1 time-frequency domain characteristic extracted from the characteristic PF.

5. The track slab deformation identification method based on track side vibration acceleration according to claim 5, characterized in that the time domain characteristic parameters comprise:

peak-to-peak: v_p-p＝max(PF_q)-min(PF_q)

Variance:

root mean square:

shape factor:

crest factor: c ═ max (P)F_q)/PF_rms

In the formula: PF (particle Filter)_qRepresents the q-th characteristic PF,

the average value of the characteristic PFs is shown, and r is the number of the characteristic PFs.

6. The track slab deformation identification method based on the track-side vibration acceleration according to claim 1, wherein the step S4 is specifically:

step S41: randomly extracting m from original rail side vibration signal data in a releasing way by a Bootstrap resampling method_treeThe self-help sample set is used as a training data set D, and data which are not obtained form an OOB data set;

step S42: for the extracted m_treeGenerating corresponding m respectively for each sample set_treeClassifying and returning trees;

step S43: with t features, m is randomly drawn at each node of each tree_tryCalculating and selecting the characteristic with the maximum information gain as a classification characteristic by adopting an information gain principle;

step S44: the classification regression tree is not cut, and each tree is made to grow to a preset limit;

step S45: and forming a random forest by the generated trees.

7. The track slab deformation identification method based on the track-side vibration acceleration according to claim 1, wherein the step S43 is specifically:

firstly, calculating the information entropy:

wherein p is_iThe proportion of each classification characteristic i in the classification system.

Defining the information gain ratio of the feature a to the training data set D as the ratio of the information gain of the feature a to the entropy of the training data set D:

Images