Method for Monitoring Service State of Switch Rail Based on Feature Fusion

Abstract

A method for monitoring a service state of a switch rail based on feature fusion includes: mounting an ultrasonic guided wave monitoring apparatus on the switch rail to perform online defect identification and simulation of the switch rail, and establishing a baseline signal library, where the ultrasonic guided wave monitoring apparatus generates a guided wave signal propagating along the switch rail and receives an echo signal of the guided wave signal; performing feature extraction; extracting a healthy feature vector, and nondimensionalizing; selecting a defect-sensitive feature, and acquiring, by a BPSO algorithm, a best feature subset; training an LS-SVM through the best feature subset by a cross-validation method to acquire an automatic online defect identification model of the switch rail based on the LS-SVM; and monitoring the switch rail by the automatic online defect identification model of the switch rail based on the LS-SVM.

Claims

1. A method for monitoring a service state of a switch rail based on feature fusion, comprising: S1: mounting an ultrasonic guided wave monitoring apparatus on the switch rail to perform online defect identification and simulation of the switch rail and establishing a baseline signal library, wherein the ultrasonic guided wave monitoring apparatus generates a guided wave signal propagating along the switch rail and receives an echo signal of the guided wave signal: S1.0: acquiring, by the ultrasonic guided wave monitoring apparatus, an echo signal x′(n) in real time, wherein the echo signal is acquired by sampling multiple sample points continuously in a time series; n denotes a serial number of the sample point in the echo signal; the echo signal comprises a healthy echo signal x′.sub.1(n) of the switch rail in a healthy state and a defective echo signal x′.sub.2(n) of the switch rail in a simulated defective state; and the healthy echo signal x′.sub.1(n) and the defective echo signal x′.sub.2(n) have an identical count; S1.1: preprocessing the acquired echo signal x′(n) by normalization: $x\left(n\right)=\frac{x\prime \left(n\right)-\frac{1}{M}.Math.x\prime \left(n\right)}{\sqrt{\frac{1}{M}.Math.{\left(x\prime \left(n\right)-\frac{1}{M}.Math.x\prime \left(n\right)\right)}^2}}$ wherein M denotes a length of the echo signal x′(n), wherein M is a total count of sample points in the echo signal; and n denotes the serial number of the sample point in the echo signal; through the preprocessing, a healthy baseline signal x.sub.1(n) and a defective baseline signal x.sub.2(n) are acquired; the healthy baseline signal x.sub.1(n) and the defective baseline signal x.sub.2(n) are baseline signals x(n); all different baseline signals totaling BN form the baseline signal library; and the healthy baseline signal x.sub.1(n) and the defective baseline signal x.sub.2(n) each account for half in the baseline signal library; S2: performing feature extraction on the preprocessed baseline signal x(n); S3: randomly extracting a healthy baseline signal of the switch rail in the healthy state from the baseline signal library and processing the healthy baseline signal to acquire a feature vector as a healthy feature vector F.sub.baseline:
F.sub.baseline=(SF.sub.1.sup.baseline to SF.sub.18.sup.baseline,ESEF.sub.1.sup.baseline,ESEF.sub.2.sup.baseline) wherein SF.sub.1.sup.baseline to SF.sub.18.sup.baseline denote first 18 feature parameters of the healthy feature vector F.sub.baseline, respectively, and ESEF.sub.1.sup.baseline, ESEF.sub.2.sup.baseline denote last two feature parameters of the healthy feature vector F.sub.baseline, respectively; nondimensionalizing each baseline signal x(n) in the baseline signal library:
e(F.sub.monitoring/F.sub.baseline)−e wherein e denotes a Napierian base; acquiring, by the nondimensionalization, a feature vector ssf of the guided wave signal of the switch rail:
ssf=SF.sub.1.sup.dimensionless to SF.sub.18.sup.dimensionless,ESEF.sub.1.sup.dimensionless,ESEF.sub.2.sup.dimensionless) wherein SF.sub.1.sup.dimensionless to SF.sub.18.sup.dimensionless denote first 18 feature parameters of the feature vector ssf of the guided wave signal of the switch rail, respectively, and ESEF.sub.1.sup.dimensionless, ESEF.sub.2.sup.dimensionless denote last two feature parameters of the feature vector ssf of the guided wave signal of the switch rail, respectively; S4: selecting a defect-sensitive feature from the feature vector ssf of the guided wave signal of the switch rail and acquiring, by a binary particle swarm optimization (BPSO) algorithm, a best feature subset by: calculating a fitness function of a population after each iteration, updating a global best solution gbest, and taking a final global best solution gbest as the best feature subset SSF.sub.gbest, SSF.sub.gbest={ssf.sub.1, ssf.sub.2, . . . , ssf.sub.BN}, BN denoting the count of the baseline signals in the baseline signal library; S5: acquiring an automatic online defect identification model of the switch rail based on a least squares support vector machine (LS-SVM): taking the best feature subset SSF.sub.gbest as a training set of the LS-SVM and training by a cross-validation method to acquire the automatic online defect identification model of the switch rail based on the LS-SVM; and S6: monitoring the switch rail by the automatic online defect identification model of the switch rail based on the LS-SVM: processing, by steps S1 to S3, a target echo signal MS(n) of the switch rail acquired in real time to acquire a feature vector of a guided wave signal of the switch rail using ssf.sub.ms=(SF.sub.1.sup.dimensionless to SF.sub.18.sup.dimensionless, ESEF.sub.1.sup.dimensionless, ESEF.sub.2.sup.dimensionless); performing, by step S4, feature selection on the feature vector ssf.sub.ms of the guided wave signal of the switch rail based on a serial number NF of a feature to acquire a feature vector ssf.sub.ms.sup.new of the guided wave signal of the switch rail; inputting the feature vector ssf.sub.ms.sup.new of the guided wave signal of the switch rail into the automatic online defect identification model of the switch rail based on the LS-SVM acquired in step S5; and outputting a current service state of the switch rail, wherein the current service state of the switch rail is a healthy state or a defective state.

2. The method for monitoring the service state of the switch rail based on the feature fusion according to claim 1, wherein step S2 specifically comprises: S2.1: extracting a time-domain statistical feature and a power spectrum-domain statistical feature of the preprocessed baseline signal x(n), wherein the time-domain statistical feature comprises skewness (SF.sub.1), kurtosis (SF.sub.2), peak-to-peak value (SF.sub.3), kurtosis factor (SF.sub.4), root mean square (RMS) (SF.sub.5), standard deviation (SF.sub.6), crest factor (SF.sub.7), shape factor (SF.sub.8), pulse factor (SF.sub.9), maximum (SF.sub.15), variance (SF.sub.16), minimum (SF.sub.17), root mean square amplitude (RMSA) (SF.sub.18), and margin factor (SF.sub.10); and the power spectrum-domain statistical feature comprises RMS (SF.sub.11), standard deviation (SF.sub.12), and centroid (SF.sub.13); S2.2: performing feature extraction on the preprocessed baseline signal x(n) by a time-frequency analysis: performing complementary ensemble empirical mode decomposition (CEEMD) on the baseline signal x(n) to acquire a set of intrinsic mode functions (IMFs) with different frequency bands from high to low, IMFs={c.sub.1(n), c.sub.2(n), . . . , c.sub.N(n)}, n∈[1,M]:
x(n)=Σ.sub.t=1.sup.Nc.sub.t(n)+r.sub.n(n) wherein c.sub.1(n) denotes a t-th IMF obtained after the CEEMD; N denotes a total count of the IMFs; and r.sub.n(n) denotes an n-th residual component; S2.3: reconstructing the preprocessed baseline signal x(n) according to the IMF to acquire a reconstructed signal:
y(n)=Σ.sub.q=1.sup.Nc.sub.q(n) wherein y(n) denotes the reconstructed signal of the baseline signal x(n) based on the CEEMD; S2.4: extracting an average energy feature SF.sub.14 of the baseline signal x(n) according to the reconstructed signal y(n) based on the CEEMD: ${\mathrm{SF}}_{14}=\frac{1}{M}{.Math.}_{n=1}^M{y}^2\left(n\right)$ wherein y(n) denotes the reconstructed signal of the baseline signal x(n) based on the CEEMD; and S2.5: extracting a spectral entropy feature of the baseline signal x(n) according to the reconstructed signal y(n) based on the CEEMD: extracting the spectral entropy feature of the baseline signal x(n) based on the CEEMD, wherein the spectral entropy feature comprises a local spectral entropy feature ESEF.sub.1 of first two IMFs and a global spectral entropy feature ESEF.sub.2 of all the IMFs; the local spectral entropy feature ESEF.sub.1 is expressed as: ${\mathrm{ESEF}}_1={.Math.}_{i=1}^M{p}_i\left(e^{\left(1-p_i\right)}-1\right)$ $p_i=\frac{{\psi \left(i\right)}^2}{{.Math.}_{i=1}^M{\psi \left(i\right)}^2},$ wherein p.sub.i denotes an energy coefficient of an i-th sample point in the baseline signal x(n); and ψ(i) denotes a locally reconstructed signal acquired by reconstruction using the first two IMFs; the global spectral entropy feature ESEF.sub.2 is expressed as: ${\mathrm{ESEF}}_2={.Math.}_{j=1}^N{.Math.}_{i=1}^M{p}_{\mathrm{ij}}\left(e^{\left(1-p_{\mathrm{ij}}\right)}-1\right)$ $p_{\mathrm{ij}}=\frac{c_{\mathrm{ij}}^2}{{.Math.}_{i=1}^M{y\left(n\right)}^2}$ wherein p.sub.ij denotes an energy coefficient of an i-th sample point of a j-th IMF in the baseline signal x(n); e denotes a Napierian base; and c.sub.ij denotes a value of the i-th sample point of the j-th IMF; and through the above process, the feature of each baseline signal in the baseline signal library is extracted, and all feature extraction results form a feature vector F.sub.monitoring of the baseline signal, F.sub.monitoring=(SF.sub.1 to SF.sub.18, ESEF.sub.1, ESEF.sub.2).

3. The method for monitoring the service state of the switch rail based on the feature fusion according to claim 1, wherein the healthy echo signal of the switch rail in the healthy state comprises echo signals acquired from different positions of a base, a web, and a head of the switch rail in the healthy state, and comprises echo signals of the switch rail in three states of closed, open and moving, each accounting for one-third of all echo signals of the switch rail in the healthy state; and the defective echo signal of the switch rail in the simulated defective state comprises echo signals acquired from different positions of the base, the web, and the head of the switch rail in the simulated defective state, and comprises echo signals of the switch rail in the three states of closed, open and moving, each accounting for one-third of all echo signals of the switch rail in the simulated defective state.

4. The method for monitoring the service state of the switch rail based on the feature fusion according to claim 1, wherein the LS-SVM uses a Gaussian radial basis function as a kernel function, and the cross-validation method is a leave-one-out method.

5. The method for monitoring the service state of the switch rail based on the feature fusion according to claim 1, wherein the fitness function is calculated as a ratio of an intra-class Mahalanobis distance to an inter-class Mahalanobis distance by the following steps: calculating a Mahalanobis distance of a j-th class: $d_j=\frac{1}{n_j}{.Math.}_{k=1}^{n_j}{.Math.}_{i=1}^{\gamma }\sqrt{{\left({\mathrm{ssf}}_{\mathrm{ik}}-v_j\right)}^T{\psi }_j^{-1}\left({\mathrm{ssf}}_{\mathrm{ik}}-v_j\right)}$ $\left(j=1,2,.Math.,p\right)$ $v_j=\frac{1}{2}{.Math.}_{i=1}^{n_j}{\mathrm{ssf}}_{\mathrm{ik}}$ wherein v.sub.j denotes a centroid vector of the j-th class; p denotes a count of classes, wherein p is 2 for the healthy state and the defective state; ψj: denotes a covariance matrix of a feature set SSF.sub.j, SSF.sub.j={ssf.sub.1, ssf.sub.2, . . . , ssf.sub.nj}; SSF.sub.j is a set of feature vectors ssf of the guided wave signal of the switch rail based on the baseline signals of the j-th class; ssf.sub.nj denotes a feature vector of the guided wave signal of the switch rail based on an n.sub.j-th sample; n.sub.j denotes a count of the baseline signals of the j-th class; γ denotes a dimension of the feature vector ssf.sub.i; T denotes a matrix transposition; and ssf.sub.ik denotes an i-th feature in a k-th feature vector of the guided wave signal of the switch rail; calculating the intra-class Mahalanobis distance d and the inter-class Mahalanobis distance dd by the Mahalanobis distance of the j-th class:
d=1/pΣd.sub.j
dd=√{square root over (v.sub.1−v.sub.2).sup.Tζ(v.sub.1−v.sub.2).)} wherein ζ denotes a variance of a vector v.sub.1; the vector v.sub.1 denotes a centroid vector of a 1.sup.st class; and a vector v.sub.2 denotes a centroid vector of a 2.sub.nd class; and calculating the fitness function of the BPSO algorithm:
fitness=e.sup.d/(dd+τ) wherein τ denotes a non-zero parameter of the fitness function.

6. The method for monitoring the service state of the switch rail based on the feature fusion according to claim 1, wherein an output result of the automatic online defect identification model of the switch rail based on the LS-SVM is that the switch rail is healthy, or the switch rail is defective.

7. The method for monitoring the service state of the switch rail based on the feature fusion according to claim 2, wherein step S2.2 further comprises: calculating the count N of the IMFs with different frequency bands based on the length M of the input baseline signal x(n):
N=log.sub.2(M) wherein log.sub.2(M) denotes a maximum integer less than or equal to log.sub.2(M).

Description

BRIEF DESCRIPTION OF THE DRAWINGS

[0076] FIG. 1 is a flowchart of a monitoring method according to the present disclosure.

[0077] FIG. 2 is a diagram of a time-domain guided wave signal according to an embodiment of the present disclosure.

[0078] FIGS. 3A and 3B show diagrams of a baseline signal extracted from a baseline signal library according to an embodiment of the present disclosure, where FIG. 3A is a three-dimensional diagram of an original feature and FIG. 3B is a three-dimensional diagram of a dimensionless feature.

[0079] FIG. 4 is an iterative convergence graph of a binary particle swarm optimization (BPSO) algorithm according to an embodiment of the present disclosure.

[0080] FIG. 5 is a three-dimensional diagram of a best feature subset of the baseline signal extracted from the baseline signal library acquired by an iteration of the BPSO algorithm according to an embodiment of the present disclosure.

[0081] FIG. 6 is a three-dimensional diagram of a feature extracted from a target guided wave signal according to an embodiment of the present disclosure.

[0082] FIG. 7 is a three-dimensional diagram of a best feature subset of the target guided wave signal acquired by an iteration of the BPSO algorithm according to an embodiment of the present disclosure.

[0083] FIG. 8 is an identification result of an actual monitoring signal of a switch rail output by an automatic online defect identification model based on the least squares support vector machine (LS-SVM) according to an embodiment of the present disclosure.

DETAILED DESCRIPTION OF THE EMBODIMENTS

[0084] The present disclosure will be further described in detail below in conjunction with the drawings and embodiments.

[0085] As shown in FIG. 1, an embodiment implemented according to the method of the present disclosure is as follows:

[0086] As shown in FIG. 1 to FIG. 8, the present disclosure performs monitoring processing and analysis according to a guided wave signal acquired by monitoring.

[0087] A missing report rate test was performed on the base of a switch rail to validate the present disclosure. According to step S1, during the monitoring process, there was no defect on the switch rail, and the switch rail was in three states: open, closed and moving. 437 healthy echo signals x′.sub.1(n) in a healthy state were acquired. Simulation metal block were attached, and the switch rail was in three states of open, closed and moving. 437 defective echo signals x′.sub.2(n) in a simulated defective state were further acquired. FIG. 2 shows a typical time-domain guided wave signal of the switch rail in the three working states. FIG. 2 shows, from top to bottom, the healthy guided wave signals in the three states of open, closed and moving.

[0088] The guided wave signals were preprocessed according to step S1.1 to improve the data quality. After preprocessing, a healthy baseline signal x.sub.1(n) and a defective baseline signal x.sub.2(n) were acquired. The healthy baseline signal x.sub.1(n) and the defective baseline signal x.sub.2(n) were baseline signals x(n). A baseline signal library was formed by all different baseline signals. The number of baseline signals in the baseline signal library was BN=874. Further, feature extraction was performed on the preprocessed baseline signal x(n) according to step S2. First, a time-domain statistical feature was extracted, including skewness (SF1), kurtosis (SF2), peak-to-peak value (SF3), kurtosis factor (SF4), root mean square (RMS) (SF5), standard deviation (SF6), crest factor (SF7), shape factor (SF8), pulse factor (SF9), average energy feature (SF14), maximum (SF15), variance (SF16), minimum (SF17), root mean square amplitude (RMSA) (SF18), and margin factor (SF10). Then, a power spectrum-domain statistical feature was extracted, including RMS (SF11), standard deviation (SF12), and centroid (SF13). Finally, a time-frequency analysis feature was extracted, including ESEF.sub.1 and ESEF.sub.2. FIG. 3A shows a three-dimensional feature diagram of the 874 baseline signals. Due to the different dimensions of these original features, the size range of the features is quite different, which is not conducive to subsequent machine learning. To this end, according to step S3, the feature vector of the baseline signal was subjected to nondimensionalization, and the three-dimensional feature diagram of 874 samples was acquired. As shown in FIG. 3B, these features were all in a similar size range. According to step S4, damage-sensitive feature selection was performed on the nondimensionalized feature vector of the switch rail to acquire a best feature subset. The initialization iteration parameters include the number of particles in the population, Num=20; the maximum number of iterations, Die=100; the maximum iteration speed, VV.sub.max=6; the minimum iteration speed, VV.sub.min=−6; and the learning factors, c1=c2=2. After 100 iterations, a final damage-sensitive best feature subset SSF.sub.gbest, was acquired, as shown in FIG. 5. The serial numbers of the selected features are NF={2 3 5 7 12 19 20}. Through a fitness function of the binary particle swarm optimization (BPWO) algorithm, a convergence curve was acquired after the iteration, as shown in FIG. 4. A 7-dimensional damage-sensitive feature subset of CH1 and CH2 was selected. According to step S5, the final damage-sensitive best feature subset SSF.sub.gbest, was taken as a training set of the least squares vector machine (LS-SVM). An automatic online defect identification model of the switch rail based on the LS-SVM was acquired by training through a leave-one-out method.

[0089] A target guided wave signal of the switch rail acquired in real time was denoted MS(n) with a total of 6,172 samples. Through steps S1.1, S2, and S3, the feature vector of the target guided wave signal MS(n) was extracted, namely, ssf.sub.ms=(SF.sub.1.sup.dimensionless to SF.sub.18.sup.dimensionless, ESEF.sub.1.sup.dimensionless, ESEF.sub.2.sup.dimensionless), as shown in FIG. 6.

[0090] Feature selection was performed on ssf.sub.ms according to the serial numbers of the features NF={2 3 5 7 12 19 20} in step S4 to acquire selected features ssf.sub.ms.sup.new, as shown in FIG. 7. According to step S6, the selected features ssf.sub.ms.sup.new of the 6,172 samples were input into the automatic online defect identification model of the switch rail based on the LS-SVM acquired in step S5 to acquire an identification result of the current service state of the switch rail, as shown in FIG. 8. In FIG. 8, the target guided wave signals were acquired by attaching and removing metal block from different positions on the base of the switch rail. The automatic online defect identification model of the switch rail based on the LS-SVM might output the number 5 to indicate that the switch rail is defective and the number 1 to indicate that the switch rail is healthy. An identification result of 99.45% (6,138/6,172) indicated that there were 6,172 monitoring signals to be analyzed, and the automatic online defect identification model of the switch rail based on the LS-SVM had correctly identified 6,138 monitoring signals with an identification accuracy of 99.45%, thus satisfying the engineering monitoring requirements. In the traditional monitoring method based on independent component analysis (ICA), a defect threshold is set in order to realize automatic defect identification. The setting of the defect threshold is related to the count of the baseline signals. When the count of the baseline signals is different, the identification accuracy of the switch rail is different. This will affect the automation degree of switch rail defect identification and greatly reduce the stability of identification. The method of the present disclosure does not need to set a threshold, and the trained model automatically identifies the defects.

[0091] In practical application, the present disclosure utilizes the ultrasonic guided wave monitoring apparatus to acquire the baseline signals of the switch rail, including the defective echo signal x′.sub.2(n) and the healthy echo signal x′.sub.1(n). The defective echo signal is acquired by attaching metal block to simulate defects. The percentage of the cross-sectional area of the switch rail with metal block used in the simulation is comparable to that of the cross-sectional area of the switch rail with the simulated defects. The metal block are attached from the base to the head of the switch rail, and the metal block are spaced one meter apart. The present disclosure trains the baseline signals to acquire automatic online defect identification model of the switch rail based on the LS-SVM with desired generalization ability to different defects, and this model will monitor the service state of the switch rail in the follow-up monitoring.

[0092] Compared with the traditional method, the present disclosure can carry out defect identification under complex conditions and realize online defect monitoring of the switch rail, which has important practical significance and engineering value.

[0093] The above specific implementations are used to explain the present disclosure, rather than to limit the present disclosure. Within the spirit of the present disclosure and the protection scope of the claims, any modification and change to the present disclosure should fall within the protection scope of the present disclosure.