Rail transit structure health monitoring method

Abstract

The invention discloses a rail transit structure health monitoring method, and belongs to the technical field of rail transit structure health monitoring. According to the method, vibration sensors are symmetrically arranged on tracks on the two sides to obtain a pair of vibration signals, cepstrum is extracted from each section of vibration signals, real part and imaginary part cepstrum data are obtained, cepstrum features of the real part and imaginary part are extracted respectively, and corresponding feature vectors are constructed. And the mismatching degree of the pair of vibration signals on the feature vector is obtained, and the mismatching degree of the real part and the imaginary part is obtained. Dividing each mismatching degree into three parts according to the time sequence, extracting a mean value, and calculating a change coefficient; and extracting real part and imaginary part feature values of the mismatching degree of each part, performing feature enhancement by using a change coefficient, and finally obtaining a health value of the rail transit structure. According to the method, the vibration signals are processed and analyzed through multiple steps, and the accuracy of rail transit structure health monitoring is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of rail transit structure health monitoring, and in particular to a rail transit structure health monitoring method. Background Art

[0002] In the modern transportation system, rail transit has become an important mode of transportation between cities and regions due to its advantages such as large capacity and high efficiency. However, long-term high-load operation, complex geological conditions and harsh environmental factors all pose challenges to the safety and reliability of rail transit structures. Once the rail transit structure has health problems, it will not only affect the smooth operation of the train, but may also cause serious safety accidents, resulting in huge economic losses and social impacts.

[0003] Traditional rail transit vibration monitoring uses a single-point sensor arrangement. The health of the rail transit structure is determined through simple threshold judgment and spectrum analysis. Specifically, a fixed vibration amplitude threshold is set in advance. Once the vibration amplitude detected by the sensor exceeds the threshold, it is determined that there may be an abnormality in the track structure and an early warning is issued. At the same time, by performing spectrum analysis on the collected vibration signal, an attempt is made to find the abnormal frequency components in the signal to determine whether there are health problems with the track structure.

[0004] However, the operating conditions of rail transit are complex and changeable. Different train types, operating speeds, load conditions, etc. will have different degrees of impact on track vibration. These impacts will cause fluctuations in vibration amplitude and changes in abnormal frequency components. This makes it difficult for traditional monitoring methods based on fixed thresholds and simple spectrum analysis to accurately judge the health of track structures, resulting in low accuracy in rail transit structure health monitoring. Summary of the invention

[0005] In view of the above-mentioned deficiencies in the prior art, a rail transit structure health monitoring method provided by the present invention solves the problem of low accuracy of rail transit structure health monitoring in the prior art.

[0006] In order to achieve the above-mentioned invention object, the technical solution adopted by the present invention is: a rail transit structure health monitoring method, comprising the following steps: Vibration sensors are symmetrically arranged on the tracks on both sides to obtain a pair of vibration signals, wherein the signals collected by the sensors with symmetrical positions are a pair; Extract the inverse spectrum of each signal in the vibration signal to obtain the real inverse spectrum data and the imaginary inverse spectrum data; Extracting cepstrum features from real cepstrum data and imaginary cepstrum data respectively, and constructing real cepstrum feature vectors and imaginary cepstrum feature vectors; Obtaining the mismatch degree of a pair of vibration signals on the real part inverse spectrum eigenvector and the imaginary part inverse spectrum eigenvector, and obtaining the real part mismatch degree and the imaginary part mismatch degree; According to the time sequence, the mismatch degrees of the real and imaginary parts are divided into three parts, the mean value is extracted for each part, and the coefficient of variation of the mismatch degrees of the real and imaginary parts is calculated; The real and imaginary eigenvalues ​​of the real and imaginary mismatch of each part are extracted, and the variation coefficients of the real and imaginary mismatch are used to enhance the characteristics to obtain the health value of the rail transit structure.

[0007] Further, the cepstrum features of the real part cepstrum data include: the mean of the real part cepstrum, the variance of the real part cepstrum and the mean of the key real part cepstrum, wherein the key real part is a real part greater than the mean of the real part cepstrum; The cepstrum features of the imaginary cepstrum data include: a mean of the imaginary cepstrum, a variance of the imaginary cepstrum and a mean of a key imaginary cepstrum, wherein the key imaginary part is an imaginary part greater than the mean of the imaginary cepstrum.

[0008] Furthermore, the specific process of obtaining the real part mismatch degree is as follows: calculating the similarity of a pair of vibration signals on the real part inverse spectrum feature vector, subtracting the similarity from 1 to obtain the real part mismatch degree; or calculating the Euclidean distance of a pair of vibration signals on the real part inverse spectrum feature vector, normalizing the Euclidean distance, and subtracting the normalized result from 1 to obtain the real part mismatch degree; The similarity of a pair of vibration signals on the imaginary inverse spectrum feature vector is calculated, and the imaginary part mismatch is obtained by subtracting the similarity from 1; or the Euclidean distance of a pair of vibration signals on the imaginary inverse spectrum feature vector is calculated, and the Euclidean distance is normalized, and the imaginary part mismatch is obtained by subtracting the normalized result from 1.

[0009] Furthermore, the formula for calculating the mismatch degree is: or ; Among them, θ is the mismatch degree, A 1 is the real part cepstrum eigenvector or imaginary part cepstrum eigenvector corresponding to a vibration signal, A 2 is the real part cepstrum eigenvector or imaginary part cepstrum eigenvector corresponding to another vibration signal, A 1,i A 1 The i-th element in A 2,i A 2 The i-th element in A, i is the number of the element, f(A 1 ,A 2 ) is for A 1 and A 2 Find the cosine similarity.

[0010] Furthermore, the specific process of calculating the variation coefficient of the real part and the imaginary part mismatch degree is as follows: In chronological order, the real part mismatch degree data and the imaginary part mismatch degree data of all time points are equally divided into three parts corresponding to the starting time period, the middle time period and the ending time period according to the time length; Calculate the mean of each real part mismatch degree and each imaginary part mismatch degree in the initial, middle and final time periods respectively; Calculate the ratio of the mean value of the real part mismatch between the middle and the beginning, and between the end and the middle time periods in turn, add the two ratios together to get the coefficient of variation of the real part mismatch; The ratios of the mean values ​​of the imaginary part mismatch between the middle and the beginning, and between the end and the middle time periods are calculated in turn, and the two ratios are added together to obtain the coefficient of variation of the imaginary part mismatch.

[0011] Furthermore, the specific process of obtaining the variation coefficient of the real part mismatch degree includes: taking the ratio of the mean value of each real part mismatch degree in the middle time period to the mean value of each real part mismatch degree in the starting time period, taking the ratio of the mean value of each real part mismatch degree in the end time period to the mean value of each real part mismatch degree in the middle time period, adding the two ratios to obtain the variation coefficient of the real part mismatch degree; The specific process of obtaining the coefficient of variation of the imaginary part mismatch degree includes: taking the ratio of the mean value of each imaginary part mismatch degree in the middle time period to the mean value of each imaginary part mismatch degree in the starting time period, taking the ratio of the mean value of each imaginary part mismatch degree in the end time period to the mean value of each imaginary part mismatch degree in the middle time period, adding the two ratios to obtain the coefficient of variation of the imaginary part mismatch degree.

[0012] Furthermore, the specific process of obtaining the rail transit structure health value is as follows: In each part, the real part mismatch degree greater than 0.5 is screened out, and the real part outlier value of the part is calculated; In each part, the imaginary part mismatch greater than 0.5 is screened out, and the imaginary part outlier of the part is calculated; The first fully connected layer is used to process the real part outliers of each part to obtain the real part eigenvalues; The second fully connected layer is used to process the imaginary outliers of each part to obtain the imaginary eigenvalues; The real part eigenvalue is enhanced by using the variation coefficient of the real part mismatch degree to obtain the real part enhanced eigenvalue; The imaginary part eigenvalue is enhanced by using the variation coefficient of the imaginary part mismatch degree to obtain the imaginary part enhanced eigenvalue; The third fully connected layer is used to process the real part enhanced eigenvalue and the imaginary part enhanced eigenvalue to obtain the rail transit structure health value.

[0013] Furthermore, the specific process of calculating the real part outlier of the part includes: screening out the real part mismatching degree greater than 0.5 in each part, and obtaining the number M of the real part mismatching degree greater than 0.5 re,H The number of real part mismatches M that are less than or equal to 0.5 re,L , M re,H With M re,L Subtract the two and normalize the subtraction result. Add the normalized result to 1 to get the real part mismatch factor. Take the average of the real part mismatch degrees greater than 0.5, multiply the average by the real part mismatch factor to get the real part outlier of this part.

[0014] Furthermore, the specific process of calculating the imaginary part outlier of the part includes: screening out the imaginary part mismatching degree greater than 0.5 in each part, and obtaining the number M of imaginary part mismatching degrees greater than 0.5 im,H The number of imaginary mismatches M that is less than or equal to 0.5 im,L , M im,H With M im,L Subtract the two and normalize the subtraction result. Add the normalized result to 1 to get the imaginary part mismatch factor. Take the average of the imaginary part mismatch degrees greater than 0.5, multiply the average by the imaginary part mismatch factor to get the imaginary part outlier of this part.

[0015] Furthermore, the variation coefficient of the real part mismatch degree is multiplied by the real part eigenvalue to obtain the real part enhanced eigenvalue; The imaginary part enhanced eigenvalue is obtained by multiplying the variation coefficient of the imaginary part mismatch degree by the imaginary part eigenvalue.

[0016] The beneficial effects of the present invention are: 1. The present invention arranges vibration sensors symmetrically on both sides of the track to obtain a pair of vibration signals. Compared with the traditional single-point arrangement, this symmetrical sensor arrangement can simultaneously obtain vibration information on both sides of the track. The vibration conditions on both sides of the track can verify and complement each other. When there is local damage or abnormality on one side of the track, the signal difference at the symmetrical position can be more clearly reflected, thereby capturing the vibration characteristics of the track more comprehensively and accurately, avoiding the omission of key information due to the limitations of single-point monitoring, and helping to more accurately judge the health status of the track structure.

[0017] 2. The present invention extracts the cepstrum of the vibration signal and further obtains the real and imaginary cepstrum data, and then constructs the real and imaginary cepstrum feature vectors respectively to reflect the distribution of the real and imaginary cepstrum.

[0018] 3. The present invention obtains the real part mismatch and the imaginary part mismatch by obtaining the mismatch of a pair of vibration signals on the real and imaginary inverse spectrum feature vectors, and quantifies the difference in inverse spectrum characteristics of vibration signals at symmetrical positions on both sides of the track. Under normal circumstances, the vibration signals at symmetrical positions on both sides of the track have similarities in inverse spectrum characteristics. When an abnormality occurs in the track structure on one side, the vibration response will change significantly, resulting in an increase in mismatch. Using mismatch as a judgment indicator can more sensitively identify abnormal conditions of the track structure. Compared with the traditional threshold judgment based only on a single signal, it can more accurately distinguish normal signal characteristics from fault signal characteristics, reducing the possibility of misjudgment and missed judgment.

[0019] 4. The present invention divides the mismatch of each real part and imaginary part into three parts according to the time sequence, extracts the mean value of each part and calculates the coefficient of variation. It can perform trend analysis on the health status of the track structure. When the coefficient of variation gradually increases, it means that the mismatch of the track structure is increasing.

[0020] 5. The present invention extracts the real eigenvalue and imaginary eigenvalue of each part of the real and imaginary mismatch, and uses the variation coefficient of the real and imaginary mismatch to strengthen the features, and finally obtains the rail transit structure health value. The variation coefficient reflects the degree of change of the mismatch. Using it for feature strengthening can make the features related to the health status of the rail structure more prominent and obvious, and improve the monitoring accuracy of the rail transit structure health value. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 The figure is a flow chart of a rail transit structure health monitoring method.

[0022] Figure 2 A schematic diagram of the layout. DETAILED DESCRIPTION

[0023] The specific implementation modes of the present invention are described below so that those skilled in the art can understand the present invention. However, it should be clear that the present invention is not limited to the scope of the specific implementation modes. For those of ordinary skill in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the attached claims, these changes are obvious, and all inventions and creations utilizing the concept of the present invention are protected.

[0024] like Figure 1 As shown, a rail transit structure health monitoring method includes the following steps: Vibration sensors are symmetrically arranged on the tracks on both sides to obtain a pair of vibration signals, wherein the signals collected by the sensors with symmetrical positions are a pair; Extract the inverse spectrum of each signal in the vibration signal to obtain the real inverse spectrum data and the imaginary inverse spectrum data; Extracting cepstrum features from real cepstrum data and imaginary cepstrum data respectively, and constructing real cepstrum feature vectors and imaginary cepstrum feature vectors; Obtaining the mismatch degree of a pair of vibration signals on the real part inverse spectrum eigenvector and the imaginary part inverse spectrum eigenvector, and obtaining the real part mismatch degree and the imaginary part mismatch degree; According to the time sequence, the mismatch degrees of the real and imaginary parts are divided into three parts, the mean value is extracted for each part, and the coefficient of variation of the mismatch degrees of the real and imaginary parts is calculated; The real and imaginary eigenvalues ​​of the real and imaginary mismatch of each part are extracted, and the variation coefficients of the real and imaginary mismatch are used to enhance the characteristics to obtain the health value of the rail transit structure.

[0025] like Figure 2 As shown, a vibration sensor is arranged on each side track, and another vibration sensor is arranged on the other side track. Figure 2 This is only a schematic diagram. In actual layout, the vibration sensors are all set on the side of the track and in the non-contact position between the track and the train to avoid being affected by the train operation.

[0026] When a train is running, the track will transmit mechanical vibration. Under ideal working conditions and with little external interference, the distance at which the vibration sensor can significantly detect the vibration signal is usually tens of meters. However, this detection distance is affected by many factors such as track material and train type.

[0027] The present invention collects vibration signals generated during the process of the train running towards the vibration sensor.

[0028] In this embodiment, the collected vibration signal is divided into multiple segments, and the inverse spectrum is extracted for each segment. For example, a vibration signal of 10 minutes is divided into 10 or 20 segments, and a sampling frequency of 2000 Hz is set for each segment.

[0029] The cepstrum features of the real part cepstrum data include: the mean of the real part cepstrum, the variance of the real part cepstrum and the mean of the key real part cepstrum, wherein the key real part is a real part greater than the mean of the real part cepstrum; The cepstrum features of the imaginary cepstrum data include: a mean of the imaginary cepstrum, a variance of the imaginary cepstrum and a mean of a key imaginary cepstrum, wherein the key imaginary part is an imaginary part greater than the mean of the imaginary cepstrum.

[0030] When calculating the cepstrum features, the present invention considers the real part and the imaginary part respectively, and extracts the mean, variance, and mean of the key parts. The mean reflects the overall level of the cepstrum, and the variance measures the dispersion of the cepstrum data. After the cepstrum transformation, the larger cepstrum components correspond to the important periodicity or harmonic information of the signal. By taking the part above the mean, noise interference can be reduced, and more attention can be paid to the main structural characteristics of the signal.

[0031] The cepstrum includes: real part and imaginary part. The present invention processes the real part and imaginary part of the cepstrum separately. The statistical characteristics of the real cepstrum can reveal the power spectrum information of the signal, such as the energy distribution of the fundamental wave and harmonics. The statistical characteristics of the imaginary cepstrum can be used for phase structure analysis. For example, in vibration signal analysis, the phase information can provide important information about the signal source.

[0032] In this embodiment, the specific process of obtaining the real part mismatch degree is: calculating the similarity of a pair of vibration signals on the real part inverse spectrum feature vector, and subtracting the similarity from 1 to obtain the real part mismatch degree; or calculating the Euclidean distance of a pair of vibration signals on the real part inverse spectrum feature vector, and normalizing the Euclidean distance, and subtracting the normalized result from 1 to obtain the real part mismatch degree;

[0033] The similarity of a pair of vibration signals on the imaginary inverse spectrum feature vector is calculated, and the imaginary part mismatch is obtained by subtracting the similarity from 1; or the Euclidean distance of a pair of vibration signals on the imaginary inverse spectrum feature vector is calculated, and the Euclidean distance is normalized, and the imaginary part mismatch is obtained by subtracting the normalized result from 1.

[0034] In this embodiment, the formula for calculating the mismatch degree is: or ; Among them, θ is the mismatch degree, A 1 is the real part cepstrum eigenvector or imaginary part cepstrum eigenvector corresponding to a vibration signal, A 2 is the real part cepstrum eigenvector or imaginary part cepstrum eigenvector corresponding to another vibration signal, A 1,i A 1 The i-th element in A 2,i A 2 The i-th element in A, i is the number of the element, f(A 1 ,A 2 ) is for A 1 and A 2 Calculate the cosine similarity. When calculating the real part mismatch, A 1 is the real part inverse spectrum eigenvector corresponding to a vibration signal, A 2 is the real part inverse spectrum eigenvector corresponding to another vibration signal; when calculating the imaginary part mismatch, A1 is the imaginary inverse spectrum eigenvector corresponding to a vibration signal, A 2 is the imaginary inverse spectrum eigenvector corresponding to another vibration signal.

[0035] The present invention calculates the similarity of a pair of vibration signals on the real and imaginary inverse spectrum feature vectors, which can directly reflect the similarity of the two signal features. Under normal operating conditions, the vibration signals at symmetrical positions on both sides of the track have a high similarity. When the track structure is abnormal, such as track wear, roadbed disease, etc., the characteristics of the vibration signals on both sides will change and the similarity will decrease. By subtracting the similarity from 1 to obtain the mismatch, the mismatch will increase accordingly, which can clearly reflect the abnormal changes in the track structure, thereby more accurately detecting the occurrence of faults.

[0036] The Euclidean distance measures the distance between two eigenvectors in space and can intuitively reflect the degree of difference in signal characteristics. Normalizing the Euclidean distance eliminates the influence of different data scales on the results, making the calculation of mismatch more objective and accurate. The mismatch obtained by subtracting the normalized result from 1 can clearly reflect the difference in vibration signal characteristics.

[0037] In this embodiment, the specific process of calculating the variation coefficient of the real part and the imaginary part mismatch degree is: In chronological order, the real part mismatch degree data and the imaginary part mismatch degree data of all time points are equally divided into three parts corresponding to the starting time period, the middle time period and the ending time period according to the time length; Calculate the mean of each real part mismatch degree and each imaginary part mismatch degree in the initial, middle and final time periods respectively; Calculate the ratio of the mean value of the real part mismatch between the middle and the beginning, and between the end and the middle time periods in turn, add the two ratios together to get the coefficient of variation of the real part mismatch; The ratios of the mean values ​​of the imaginary part mismatch between the middle and the beginning, and between the end and the middle time periods are calculated in turn, and the two ratios are added together to obtain the coefficient of variation of the imaginary part mismatch.

[0038] In this embodiment, the specific process of obtaining the variation coefficient of the real part mismatch degree includes: taking the ratio of the mean value of each real part mismatch degree in the middle time period to the mean value of each real part mismatch degree in the starting time period, taking the ratio of the mean value of each real part mismatch degree in the end time period to the mean value of each real part mismatch degree in the middle time period, adding the two ratios to obtain the variation coefficient of the real part mismatch degree; The specific process of obtaining the coefficient of variation of the imaginary part mismatch degree includes: taking the ratio of the mean value of each imaginary part mismatch degree in the middle time period to the mean value of each imaginary part mismatch degree in the starting time period, taking the ratio of the mean value of each imaginary part mismatch degree in the end time period to the mean value of each imaginary part mismatch degree in the middle time period, adding the two ratios to obtain the coefficient of variation of the imaginary part mismatch degree.

[0039] The present invention divides the mismatch data into three time periods and calculates the coefficient of variation, which can clearly capture the dynamic changes of the real and imaginary mismatches over time. For example, in a track monitoring scenario, if the coefficient of variation of the real mismatch gradually increases, it means that the difference in the real part characteristics of the vibration signals on both sides of the track is expanding, indicating that new problems have occurred in the track structure or the original problems are worsening.

[0040] In this embodiment, another implementation method for calculating the variation coefficient of the real part and imaginary part mismatch degree is given. The real part or imaginary part mismatch degree data of all time points are arranged in chronological order and recorded as y k , the corresponding time point is recorded as t k , k=1, 2, ..., K. Use the least squares method to perform linear fitting on the data and obtain the fitted line equation y=at+b. Where a is the slope of the fitted line and b is the intercept. In order to obtain the coefficient of variation, three specific time points t can be selected. 1 ,t 2 ,t 3 (For example, select at equal intervals) and calculate the fitting values ​​y corresponding to these three time points respectively 1 =at 1 +b,y 2 =at 2 +b,y 3 =at 3 +b. Then the coefficient of variation C=(y 2 -y 1 ) / y 1 +( y 3 -y 2 ) / y 2 , k is a positive integer, and K is the amount of data.

[0041] In this embodiment, the specific process of obtaining the rail transit structure health value is as follows: In each part, the real part mismatch degree greater than 0.5 is screened out, and the real part outlier value of the part is calculated; In each part, the imaginary part mismatch greater than 0.5 is screened out, and the imaginary part outlier of the part is calculated; The first fully connected layer is used to process the real part outliers of each part to obtain the real part eigenvalues; The second fully connected layer is used to process the imaginary outliers of each part to obtain the imaginary eigenvalues; The real part eigenvalue is enhanced by using the variation coefficient of the real part mismatch degree to obtain the real part enhanced eigenvalue; The imaginary part eigenvalue is enhanced by using the variation coefficient of the imaginary part mismatch degree to obtain the imaginary part enhanced eigenvalue; The third fully connected layer is used to process the real part enhanced eigenvalue and the imaginary part enhanced eigenvalue to obtain the rail transit structure health value.

[0042] The present invention divides all real part mismatches and imaginary part mismatches into three parts, namely, the starting time period, the middle time period and the ending time period. Therefore, the first fully connected layer processes the real part outliers of the starting time period, the middle time period and the ending time period, and the second fully connected layer processes the imaginary part outliers of the starting time period, the middle time period and the ending time period.

[0043] The present invention calculates outliers by screening out the real and imaginary mismatches greater than 0.5, and can focus on data that deviates greatly from the normal range, highlighting possible structural problems. The present invention calculates the outliers of the real and imaginary parts separately, and can capture abnormal information in vibration signals from two different dimensions. The present invention uses a fully connected layer to process the real and imaginary outliers to obtain eigenvalues, and uses the variation coefficient of the real and imaginary mismatches to strengthen the corresponding eigenvalues, which can further highlight the correlation between the eigenvalues ​​and the structural health status. The variation coefficient reflects the changing trend of the mismatch in different time periods. Combining it with the eigenvalue can make the eigenvalue more comprehensively reflect the dynamic changes of the structural health status, enhance the expressive power of the features, and improve the accuracy of the assessment of the structural health status. The present invention obtains the rail transit structural health value by processing the real and imaginary enhanced eigenvalues ​​and the imaginary enhanced eigenvalues ​​through the third fully connected layer, realizing the fusion of real and imaginary information. The present invention extracts characteristic information from different angles and comprehensively considers the structural health status reflected by the real and imaginary parts, which can more comprehensively and accurately evaluate the overall health status of the rail transit structure, avoiding the one-sidedness of evaluation based on single-dimensional information.

[0044] In this embodiment, another implementation scheme for obtaining the health value of the rail transit structure is given: the real mismatch degree is arranged into a matrix in each part to obtain the real mismatch degree matrix, the imaginary mismatch degree is arranged into a matrix in each part to obtain the imaginary mismatch degree matrix, the real mismatch degree matrix and the imaginary mismatch degree matrix are processed respectively by principal component analysis (PCA) to obtain the real main component and the imaginary main component, the variation coefficient of the real mismatch degree is multiplied with the real main component, the variation coefficient of the imaginary mismatch degree is multiplied with the imaginary main component, and the multiplied features are input into the support vector regression (SVR) model. SVR is a machine learning algorithm for regression analysis, which predicts the health value of the rail transit structure by finding the optimal hyperplane to fit the data. When training the SVR model, historical real outliers, imaginary outliers, variation coefficients and corresponding known health value data can be used for training to determine the parameters of the model, and then the health value of the new data is predicted.

[0045] In this embodiment, the real part outlier is equal to the mean of the real part mismatching degrees greater than 0.5. In a more preferred solution, the specific process of calculating the real part outlier of the part includes: screening out the real part mismatching degrees greater than 0.5 in each part, and obtaining the number M of real part mismatching degrees greater than 0.5. re,H The number of real part mismatches M that are less than or equal to 0.5 re,L , M re,H With M re,L Subtract the two and normalize the subtraction result. Add the normalized result to 1 to get the real part mismatch factor. Take the average of the real part mismatch degrees greater than 0.5, multiply the average by the real part mismatch factor to get the real part outlier of this part.

[0046] In this embodiment, the formula for calculating the real part outlier value of this part is: ; in, is the real part outlier, ε re is the mean of the real part mismatches greater than 0.5, M re,H is the number of real part mismatches greater than 0.5, M re,L is the number of real part mismatches less than or equal to 0.5; In this embodiment, the imaginary part outlier is equal to the average of the imaginary part mismatches greater than 0.5. In a more preferred solution, the specific process of calculating the imaginary part outlier of the part includes: screening out the imaginary part mismatches greater than 0.5 in each part, and obtaining the number M of imaginary part mismatches greater than 0.5. im,H The number of imaginary mismatches M that is less than or equal to 0.5 im,L , M im,H With M im,LSubtract the two and normalize the subtraction result. Add the normalized result to 1 to get the imaginary part mismatch factor. Take the average of the imaginary part mismatch degrees greater than 0.5, multiply the average by the imaginary part mismatch factor to get the imaginary part outlier of this part.

[0047] In this embodiment, the formula for calculating the imaginary part outlier value of this part is: ; in, is the imaginary part outlier, ε im is the mean value of the imaginary mismatch greater than 0.5, M im,H is the number of imaginary mismatches greater than 0.5, M im,L is the number of imaginary mismatches less than or equal to 0.5.

[0048] In the present invention, since the mismatch degree ranges from 0 to 1, the similarity of the inverse spectra of the two vibration sensors is high when it is close to 0, and the similarity of the inverse spectra of the two vibration sensors is low when it is close to 1. Therefore, the middle value is taken as the threshold, which is not limited to the value selected by the present invention and can be adjusted as needed.

[0049] The present invention calculates the difference between the number of real part (or imaginary part) mismatches greater than 0.5 and the number of mismatches less than or equal to 0.5, and performs normalization processing to obtain the mismatch factor. This method can effectively reflect the proportion of abnormal data in the overall data. When the number of mismatches greater than 0.5 is relatively large, the mismatch factor will be larger, so that the final calculated abnormal value will also be larger, highlighting the impact of the abnormal data. For example, in a certain part of the data, if most of the real part mismatches are greater than 0.5, then the real part mismatch factor will be significantly greater than 1, and the real part abnormal value obtained after multiplying with the mean will also increase significantly, intuitively reflecting that the degree of abnormality of this part of the data is high.

[0050] In this embodiment, the coefficient of variation of the real part mismatch degree is multiplied by the real part eigenvalue to obtain the real part enhanced eigenvalue; The imaginary part enhanced eigenvalue is obtained by multiplying the variation coefficient of the imaginary part mismatch degree by the imaginary part eigenvalue.

[0051] The first fully connected layer, the second fully connected layer, and the third fully connected layer can be trained using historical real outliers, imaginary outliers, coefficients of variation, and corresponding known health value data (the training process is implemented using the existing gradient descent method) to determine the parameters of the fully connected layers, and then predict the health values ​​of new data.

[0052] The above are only preferred embodiments of the present invention and are not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A rail transit structure health monitoring method, characterized in that: The following steps are involved: Vibration sensors are symmetrically arranged on the tracks on both sides to obtain a pair of vibration signals, wherein the signals collected by the sensors with symmetrical positions are a pair; Extract the inverse spectrum of each signal in the vibration signal to obtain the real inverse spectrum data and the imaginary inverse spectrum data; Extracting cepstrum features from real cepstrum data and imaginary cepstrum data respectively, and constructing real cepstrum feature vectors and imaginary cepstrum feature vectors; Obtaining the mismatch degree of a pair of vibration signals on the real part inverse spectrum eigenvector and the imaginary part inverse spectrum eigenvector, and obtaining the real part mismatch degree and the imaginary part mismatch degree; According to the time sequence, the mismatch degrees of the real and imaginary parts are divided into three parts, the mean value is extracted for each part, and the coefficient of variation of the mismatch degrees of the real and imaginary parts is calculated; The real and imaginary eigenvalues ​​of the real and imaginary mismatch of each part are extracted, and the variation coefficients of the real and imaginary mismatch are used to enhance the characteristics to obtain the health value of the rail transit structure.

2. The rail transit structure health monitoring method according to claim 1, characterized in that: The cepstrum features of the real part cepstrum data include: the mean of the real part cepstrum, the variance of the real part cepstrum and the mean of the key real part cepstrum, wherein the key real part is a real part greater than the mean of the real part cepstrum; The cepstrum features of the imaginary cepstrum data include: a mean of the imaginary cepstrum, a variance of the imaginary cepstrum and a mean of a key imaginary cepstrum, wherein the key imaginary part is an imaginary part greater than the mean of the imaginary cepstrum.

3. The rail transit structure health monitoring method according to claim 1, characterized in that: The specific process of obtaining the real part mismatch degree is as follows: calculating the similarity of a pair of vibration signals on the real part inverse spectrum feature vector, subtracting the similarity from 1 to obtain the real part mismatch degree; or calculating the Euclidean distance of a pair of vibration signals on the real part inverse spectrum feature vector, normalizing the Euclidean distance, and subtracting the normalized result from 1 to obtain the real part mismatch degree; The similarity of a pair of vibration signals on the imaginary inverse spectrum feature vector is calculated, and the imaginary part mismatch is obtained by subtracting the similarity from 1; or the Euclidean distance of a pair of vibration signals on the imaginary inverse spectrum feature vector is calculated, and the Euclidean distance is normalized, and the imaginary part mismatch is obtained by subtracting the normalized result from 1.

4. The rail transit structure health monitoring method according to claim 1, characterized in that: The formula for calculating the mismatch is: or ; Where θ is the mismatch degree, A1 is the real part cepstrum eigenvector or imaginary part cepstrum eigenvector corresponding to one vibration signal, A2 is the real part cepstrum eigenvector or imaginary part cepstrum eigenvector corresponding to another vibration signal, and A 1,i is the i-th element in A1, A 2,i is the i-th element in A2, i is the element number, and f(A1,A2) is the cosine similarity between A1 and A2.

5. The rail transit structure health monitoring method according to claim 1, characterized in that: The specific process of calculating the coefficient of variation of the real and imaginary mismatch is: In chronological order, the real part mismatch degree data and the imaginary part mismatch degree data of all time points are equally divided into three parts corresponding to the starting time period, the middle time period and the ending time period according to the time length; Calculate the mean of each real part mismatch degree and each imaginary part mismatch degree in the initial, middle and final time periods respectively; Calculate the ratio of the mean value of the real part mismatch between the middle and the beginning, and between the end and the middle time periods in turn, add the two ratios together to get the coefficient of variation of the real part mismatch; The ratios of the mean values ​​of the imaginary part mismatch between the middle and the beginning, and between the end and the middle time periods are calculated in turn, and the two ratios are added together to obtain the coefficient of variation of the imaginary part mismatch.

6. The rail transit structure health monitoring method according to claim 1, characterized in that: The specific process of obtaining the variation coefficient of the real part mismatch degree includes: taking the ratio of the mean value of each real part mismatch degree in the middle time period to the mean value of each real part mismatch degree in the starting time period, taking the ratio of the mean value of each real part mismatch degree in the end time period to the mean value of each real part mismatch degree in the middle time period, adding the two ratios to obtain the variation coefficient of the real part mismatch degree; The specific process of obtaining the coefficient of variation of the imaginary part mismatch degree includes: taking the ratio of the mean value of each imaginary part mismatch degree in the middle time period to the mean value of each imaginary part mismatch degree in the starting time period, taking the ratio of the mean value of each imaginary part mismatch degree in the end time period to the mean value of each imaginary part mismatch degree in the middle time period, adding the two ratios to obtain the coefficient of variation of the imaginary part mismatch degree.

7. The rail transit structure health monitoring method according to claim 1, characterized in that: The specific process of obtaining the health value of rail transit structure is as follows: In each part, the real part mismatch degree greater than 0.5 is screened out, and the real part outlier value of the part is calculated; In each part, the imaginary part mismatch greater than 0.5 is screened out, and the imaginary part outlier of the part is calculated; The first fully connected layer is used to process the real part outliers of each part to obtain the real part eigenvalues; The second fully connected layer is used to process the imaginary outliers of each part to obtain the imaginary eigenvalues; The real part eigenvalue is enhanced by using the variation coefficient of the real part mismatch degree to obtain the real part enhanced eigenvalue; The imaginary part eigenvalue is enhanced by using the variation coefficient of the imaginary part mismatch degree to obtain the imaginary part enhanced eigenvalue; The third fully connected layer is used to process the real part enhanced eigenvalue and the imaginary part enhanced eigenvalue to obtain the rail transit structure health value.

8. The rail transit structure health monitoring method according to claim 7, characterized in that: The specific process of calculating the real part outlier of the part includes: screening out the real part mismatch greater than 0.5 in each part, and obtaining the number M of real part mismatch greater than 0.5 re,H The number of real part mismatches M that are less than or equal to 0.5 re,L , M re,H With M re,L Subtract the two and normalize the subtraction result. Add the normalized result to 1 to get the real part mismatch factor. Take the average of the real part mismatch degrees greater than 0.5, multiply the average by the real part mismatch factor to get the real part outlier of this part.

9. The rail transit structure health monitoring method according to claim 7, characterized in that: The specific process of calculating the imaginary part outlier of the part includes: screening out the imaginary part mismatch greater than 0.5 in each part, and obtaining the number M of imaginary part mismatch greater than 0.5 im,H The number of imaginary mismatches M that is less than or equal to 0.5 im,L , M im,H With M im,L Subtract the two and normalize the subtraction result. Add the normalized result to 1 to get the imaginary part mismatch factor. Take the average of the imaginary part mismatch degrees greater than 0.5, multiply the average by the imaginary part mismatch factor to get the imaginary part outlier of this part.

10. The rail transit structure health monitoring method according to claim 7, characterized in that: The real part enhanced eigenvalue is obtained by multiplying the variation coefficient of the real part mismatch degree by the real part eigenvalue; The imaginary part enhanced eigenvalue is obtained by multiplying the variation coefficient of the imaginary part mismatch degree by the imaginary part eigenvalue.

Images