A railway track disease identification method based on optical fiber distributed vibration detection

Abstract

The present application relates to the field of signal source monitoring, in particular to a railway track disease identification method based on optical fiber distributed vibration detection. The present application obtains the sound wave sensing signal of the existing communication optical cable beside the railway by connecting the existing communication optical cable, and obtains the original vibration signal S(p, q) by demodulation using the differential compression-reduction dynamic range expansion algorithm based on wavelet adaptive denoising; then the expanded demodulation signal is feature extracted, and a multi-dimensional fusion feature vector is formed by vector splicing; then the feature vector is pre-weighted based on the data sample set; finally, the feature vector is classified and identified using the trained SVM model, and the disease position is recorded, and the disease position alarm of the existing optical fiber length is output. The present application realizes the identification and classification of typical track diseases such as fish scale injury, wave abrasion injury and empty hanging injury on the railway, and real-time positioning alarm disease position, providing a long-term online monitoring means for track safety operation monitoring of railway transportation.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of signal source monitoring, and particularly relates to a railway track disease identification method based on optical fiber distributed vibration detection. BACKGROUND

[0002] At present, the railway is one of the main forces in the transportation industry in the national economic construction. With the increase of train speed and transportation capacity, the interaction between wheel and track inevitably intensifies. Severe wheel-rail dynamic interaction will cause serious wear of wheel and rail, especially when freight trains run in long and steep slope, small radius curve section and wheel damage, the wheel-rail contact relationship is very bad, the disease of the bed increases, the fatigue damage of the track such as the contact fatigue damage of the oblique crack of the track (fish scale injury), the wave-shaped wear of the top surface of the track (wave abrasion injury) and the gap between the sleeper and the lower ballast caused by serious wave abrasion (empty hanging injury) is more serious. These fatigue damages of the rail aggravate the corresponding repair work, and further increase the operation cost of railway transportation, and even seriously interfere with the normal order of railway transportation. Therefore, it is necessary to find the disease of the track in time and maintain it. However, at present, the track disease monitoring is mainly realized by ultrasonic detection, eddy current detection, acceleration measurement, video image and the like. The above-mentioned sensing methods mostly have the limitations of nonlinearity, zero drift and low tolerance to strong electromagnetic environment around the transmission line, and the installation and maintenance of the sensor are very complex and expensive.

[0003] Considering the limitations of traditional electrical sensors, optical fiber sensors have been used at home and abroad to monitor the safety of railway tracks using various methods. In 2017, Wang et al. used a railway vibration detection scheme based on phase-sensitive OTDR to identify train signals and illegal intrusions during high-speed rail operation. The scheme successfully monitored the train operation signal and calculated the train length and running speed, and identified some illegal disturbances along the railway, but did not monitor and identify railway diseases. In 2016, Buggy et al. installed seven FBG strain sensors on several rail components for testing to analyze the dynamic strain caused by the passage of trains at the sensor location. Through analysis and data collection methods, the tension changes on the bolt-torque can be classified, and any changes in the condition of the railway track components can be revealed. However, the research did not apply to real-time monitoring in actual railways, but only gave a kind of feasibility. In 2020, Xin et al. proposed a scheme for identifying surface intrusion events using signals detected by a super-weak FBG array in a subway tunnel. The scheme combines spectral subtraction and the root mean square of power spectral density to extract event signals. Then, local characteristic scale decomposition and multi-scale permutation entropy are used for feature extraction, which can improve the event recognition rate from the perspective of multi-scale analysis. Experimental verification shows that the scheme can identify four events: subway trains, passing trucks, discrete impact intrusion, and continuous impact intrusion. The scheme has completed the identification of subway-related events, but has not realized the monitoring of track diseases, and the cost of FBG layout is still much higher than that of existing optical cables. SUMMARY

[0004] In view of the above problems or deficiencies, to solve the problems of timely diagnosis and real-time accurate positioning of railway track diseases, the present application provides a railway track disease identification method based on optical fiber distributed vibration detection.

[0005] A railway track disease identification method based on optical fiber distributed vibration detection, comprising the following steps:

[0006] Step 1: The optical fiber distributed acoustic sensing system DAS transmits a detection light pulse to the sensing optical fiber to realize real-time quantitative monitoring of the vibration information along the sensing optical fiber, and obtains the original vibration signal S(p, q), where p is the number of pulses obtained, and q is the number of acoustic vibration signal acquisition segments along the optical fiber.

[0007] In order to effectively solve and restore the large vibration signal collected by the system, improve the signal-to-noise ratio of the collected signal, and ensure the accuracy of subsequent signal feature extraction and classification identification, the present application adopts a differential compression-recovery dynamic range expansion algorithm based on wavelet adaptive denoising to demodulate the optical fiber distributed acoustic sensing signal to obtain the original vibration signal S(p, q) according to the characteristics of the wheel-rail vibration signal, which makes the original signal well restored in the time domain, and the frequency domain peak value is high, and the signal-to-noise ratio is improved, thereby ensuring the accurate monitoring of the optical fiber along the line of the entire distributed optical fiber sensing system.

[0008] Step 2, the original vibration signal S(p, q) obtained in step 1 is used to construct an initial data set, and the data samples of normal rail signals and three kinds of disease signals (fish scale injury, wave abrasion injury and empty hanging injury) are marked. The sample data in the initial data set are respectively subjected to time-frequency domain feature extraction using the gravity frequency, standard deviation frequency, power spectrum entropy, sample entropy and normalized wavelet energy feature vector, and the construction of the multi-dimensional fusion feature vector is completed using the vector splicing method, to obtain an initial multi-dimensional fusion feature vector data set.

[0009] Step 3, select M (M≥100) groups of data of normal rail signals and three kinds of disease signals from the initial multi-dimensional fusion feature vector data set obtained in step 2, construct a multi-dimensional fusion feature vector data set, and use the average standard deviation ASD of the multi-dimensional fusion feature vector of the normal rail signals and the three kinds of disease signals as the weight of each feature value of the multi-dimensional fusion feature vector, respectively, to pre-weight the multi-dimensional fusion feature vector, to obtain a pre-weighted multi-dimensional fusion feature vector sample set.

[0010] Step 4, input the pre-weighted multi-dimensional fusion feature vector sample set obtained in step 3 into the support vector machine SVM model for classification training, use the trained SVM model to realize the classification identification of normal rail signals and three kinds of disease signals, record the position and type of the disease and output and alarm, realize real-time disease diagnosis and positioning output.

[0011] Further, in step 1, the differential compression-recovery dynamic range expansion algorithm based on wavelet adaptive denoising is used to demodulate the optical fiber distributed acoustic sensing signal, and the specific steps include:

[0012] Step 1-1, when demodulating the optical fiber distributed acoustic sensing signal, the arctangent demodulation algorithm is used, and the step of using the arctangent function to solve the phase will cause the phase value to be limited in the range of [-π, π], at this time, the true phase can be obtained by phase unwrapping The classical phase unwrapping algorithm is represented as:

[0013]

[0014] wherein, is the original wrapped signal, is the unwrapped signal. The original wrapped signal can be expressed as where k(n) is an integer sequence, the above equation can be further expressed as:

[0015]

[0016] where, is the unwrapping threshold. However, this general algorithm also has problems. If Φ>π, the unwrapped signal will still remain wrapped, thus failing to correctly retrieve the change trace. This problem is inevitable when the phase signal amplitude changes dramatically, especially for large amplitude or high frequency vibration events on optical fiber.

[0017] To address this phenomenon, the present application proposes a differential compression-recovery algorithm based on wavelet adaptive denoising. First, the original phase wrapped signal is subjected to l-order differential compression:

[0018]

[0019] where Δ l k(n) is still an integer sequence, thus an unwrapping algorithm can be used. A discrete noise N noise (n) is introduced here. Because in the original signal, the noise can be ignored, but in the high-order differential signal, the noise will rise exponentially with the order of differentiation, while the signal amplitude is compressed, and the signal-to-noise ratio drops sharply.

[0020] Step 1-2, take as the sum of N discrete sampled sinusoidal signals ; where A i is the signal amplitude, ω i is the signal frequency, and θ i is the signal initial phase). For each sinusoidal signal its l-order differential is

[0021]

[0022] At this time, the unwrapping threshold can be expressed as:

[0023]

[0024] When 2sin(ω i / (2f s ))<1, i.e. ω i <πf s / 3, is compressed.

[0025] When l∈[1, 5], Φ<π can be satisfied, and then the unwinding can be obtained:

[0026]

[0027] When the signal is affected by the differential noise, Φ>π, the integer sequence Δ l k(n) cannot be completely eliminated by the unwinding algorithm, leaving the misunwinding term k'(n). Directly integrating the unwinding signal by l order obtains:

[0028]

[0029] wherein K'(n) is a complex l order polynomial. Therefore, in the differential stage, the wavelet denoising algorithm can be used to remove the misunwinding term k'(n) and the noise Δ l N noise (n), and further realize correct demodulation, so that the phase demodulation of the overall signal has higher tolerance to noise. The original large range and high signal-to-noise ratio vibration signal is correctly restored by the differential compression-recovery dynamic range expansion algorithm, so as to realize accurate extraction and classification recognition of subsequent signal features.

[0030] Further, the specific steps of step 2 of normalizing the wavelet energy feature vector for time-frequency domain feature extraction are as follows:

[0031] Step 2-1, obtaining the normalized wavelet energy feature vector of the signal. The collected and restored signal is decomposed by wavelet, and each decomposition layer retains the high-frequency wavelet coefficient, and the low-frequency wavelet coefficient is further decomposed. After L-layer wavelet decomposition, the low-frequency wavelet coefficient cA L and the high-frequency wavelet coefficient cD1, cD2, … cD L ;

[0032] Step 2-2, the high-frequency wavelet coefficient and the low-frequency wavelet coefficient are arranged in order from low frequency to high frequency, and the energy E j of each wavelet coefficient is solved by formula (7) to form a wavelet energy feature vector E=[E1, E2, …, E L+1 ];

[0033]

[0034] wherein C j (m) is the decomposed wavelet coefficient, and L is the number of wavelet decomposition layers.

[0035] Step 2-3, the wavelet energy feature vector obtained in step 2-2 is normalized according to formula (8) to obtain a normalized wavelet energy feature vector E'=[E'1, E'2, …, E' L]; L+1 ​

[0036]

[0037] Step 2-4, when actually performing wavelet decomposition, the decomposition level L of wavelet decomposition and the wavelet base function are determined by the average standard deviation ASD of the normalized wavelet energy feature vector.

[0038] Wavelet decomposition test is performed on each of the 4 signals (normal rail signal and 3 disease signals) in the initial data set, different wavelet decomposition levels and wavelet functions are enumerated, wavelet decomposition is performed respectively, the normalized wavelet energy feature vector is obtained, the average standard deviation (ASD) under different decomposition levels and wavelet functions is calculated as the discrimination degree of the energy vector according to formula (9) and formula (10). The wavelet decomposition level and wavelet function with the maximum discrimination degree are determined, so that the four types of signals have the maximum discrimination degree;

[0039]

[0040]

[0041] SD j is the standard deviation, is the mean value of the jth wavelet energy component of the four types of signals, and a is the corresponding sample type.

[0042] Further, the step 3 pre-weights the multi-dimensional fusion feature vector, and the specific steps include:

[0043] Step 3-1, from the multi-dimensional fusion feature vector data set constructed in step 2, M (M≥100) groups of data of normal rail signals and 3 disease signals (fish scale injury, wave abrasion injury and empty hanging injury) are selected to construct a multi-dimensional fusion feature vector data set.

[0044] Step 3-2, each multi-dimensional fusion feature vector is composed of 4 single feature values (center of gravity frequency, frequency standard deviation, power spectrum entropy and sample entropy), and a normalized wavelet energy feature vector with a length of L+1, and the total length is L+5. Each feature value has different discrimination degrees for the four types of signals, and the signal discrimination degree corresponding to each feature value is obtained according to formula (11) as the weight w of the feature value. n The weight vector of the sample data set is [w1, w2, … w N+5 ], all the obtained multi-dimensional fusion feature vectors are weighted to construct the pre-weighted multi-dimensional fusion feature vector [w1c1, w2c2, … w N+5 c N+5 ].

[0045]

[0046] In the formula, M is the number of groups of each signal data in the data set, c s is the s-th characteristic value of the multi-dimensional fusion feature vector, r is the corresponding sample group number, and a is the corresponding sample type.

[0047] Further, the step 4 specifically comprises:

[0048] Step 4-1, using the pre-weighted multi-dimensional fusion feature vector sample set obtained in step 3 to train the support vector machine, to obtain the trained support vector machine SVM.

[0049] Step 4-2, engineering deployment of the trained support vector machine SVM, using the method of steps 2 and 3 to extract and pre-weight the original vibration signal, taking the pre-weighted multi-dimensional fusion feature vector as the input vector of the support vector machine, to complete the classification and recognition.

[0050] Step 4-3, recording the recognition result of the support vector machine SVM, and outputting the position and type of the disease, to realize real-time disease diagnosis and positioning output.

[0051] The application of the optical fiber distributed vibration detection method to railway track disease monitoring is proposed, and a railway track disease recognition method based on optical fiber distributed vibration detection is proposed. The distributed optical fiber sound wave sensing system with dynamic range expansion is used to monitor and record the vibration information generated by the train running on the railway by connecting the existing communication optical cable beside the railway, and the dynamic range expansion algorithm is used to realize the waveform restoration and signal-to-noise ratio improvement of large amplitude vibration signals under the action of wheel-rail vibration, solve the problem of high-fidelity restoration of large amplitude vibration signals by the system, and extract single characteristic value and wavelet decomposition from the restored vibration information, construct a multi-dimensional fusion feature vector in the form of vector splicing, and complete the recognition and classification of various track disease types. The application can well solve the problems of timely diagnosis and real-time positioning of railway track diseases, and realize real-time discrimination and positioning of track diseases based on distributed optical fiber sound sensing.

[0052] In summary, the distributed optical fiber sound wave sensing system realized by the differential compression-restoration dynamic range expansion algorithm based on wavelet adaptive denoising is combined with multi-dimensional feature extraction and SVM classification to realize effective extraction and recognition of track diseases along the railway, solve the problems of effective extraction and positioning of track diseases, and realize timely diagnosis and real-time positioning of track diseases. BRIEF DESCRIPTION OF DRAWINGS

[0053] Figure 1 is a flowchart of the application;

[0054] Figure 2 is a flowchart of the signal differential compression-restoration dynamic range expansion algorithm of the application;

[0055] Figure 3 is a signal normalization wavelet energy feature solving flowchart of the embodiment of the present application;

[0056] Figure 4 is a signal restoration condition comparison chart of the embodiment of the present application and a traditional unwinding algorithm;

[0057] Figure 5 is a signal-to-noise ratio comparison chart of the embodiment of the present application and a traditional winding algorithm;

[0058] Figure 6 is a center of gravity frequency, frequency variance, power spectrum entropy and sample entropy comparison schematic diagram of four kinds of signals of the embodiment;

[0059] Figure 7 is a normalized wavelet energy feature vector comparison schematic diagram of four kinds of signals of the embodiment;

[0060] Figure 8 is a single-trip train multi-dimensional fusion feature vector SVM recognition result schematic diagram of the embodiment. DETAILED DESCRIPTION

[0061] The present application will be further described in detail below in combination with the drawings and embodiments.

[0062] A railway track disease identification method based on optical fiber distributed vibration detection (as shown in Figure 1 The specific implementation steps are as follows:

[0063] Step 1: Adopting an optical fiber distributed sensing system to emit a detection light pulse to a sensing optical fiber, adopting a differential compression-reduction dynamic range expansion algorithm based on wavelet adaptive denoising to complete demodulation of original large amplitude vibration signals, and performing real-time quantitative monitoring on vibration information along the sensing optical fiber, and obtaining the demodulated original vibration signal S(4000, 900).

[0064] Step 2: Constructing an initial data set for the demodulated original vibration signal S(4000, 900) obtained in step 1, marking data samples of normal track signals and three kinds of disease signals (fish scale injury, wave abrasion injury and empty hanging injury) therein. Four single characteristic values of center of gravity frequency, standard deviation frequency, power spectrum entropy and sample entropy are used for feature extraction of the signals, and then wavelet decomposition is tested on one group of data of each of the four signals in the initial data set, and it is determined that when the decomposition layer number L = 5 and the wavelet function db9 are used, the four signals have the largest distinguishability.

[0065] Wavelet decomposition is performed on the signals, and the normalized wavelet energy feature vector of the signals is calculated. The single characteristic values of the four signals and the normalized wavelet energy feature vector are constructed into a multi-dimensional fusion feature vector using a vector splicing method, and the initial data set is obtained.

[0066] Step 3, 150 groups of data of normal rail signals and 3 disease signals (fish scale injury, wave abrasion injury and empty hanging injury) are screened out from the initial data set constructed in step 2, and a multi-dimensional fusion feature vector sample set of each group of data is constructed, the average standard deviation of the multi-dimensional fusion feature vector of each disease type data is taken as a weight, the weight vector of the sample data set is [7.26, 606.42, 0.18, 0.31, 0.20, 0.11, 0.07, 0.03, 0.0062, 0.0025], and the multi-dimensional fusion feature vector is pre-weighted to obtain a pre-weighted multi-dimensional fusion feature vector sample set.

[0067] Step 4, the pre-weighted multi-dimensional fusion feature vector sample set obtained in step 3 is used to train and test the support vector machine SVM, the kernel function of the SVM is selected as a Gaussian kernel function, the penalty coefficient c is 131, and the kernel parameter g is 3.85. The pre-weighted multi-dimensional fusion feature vector of the demodulated vibration signal obtained by actual sampling is input into the support vector machine SVM model, classification and recognition of normal rail signals and 3 disease signals are realized, the position and type of the disease are recorded, and the position and type of the disease are output, so that real-time disease diagnosis and positioning output are realized.

[0068] In the embodiment, the data sample used is the vibration signal data of multiple trains of a certain heavy load railway. Figure 2 It is a flowchart of the differential compression-reduction dynamic range expansion algorithm. Figure 3 It is a flowchart of the normalized wavelet energy feature vector algorithm with the decomposition layer number N = 3 and the wavelet function db2 as an example. Figure 4 It is a comparison of the signal restoration of the differential compression-reduction dynamic range expansion algorithm and the traditional unwrapping algorithm in the embodiment, Figure 5 It is a comparison of the signal-to-noise ratio of the differential compression-reduction dynamic range expansion algorithm and the traditional unwrapping algorithm in the embodiment.

[0069] In the embodiment, the original signal after the differential compression-reduction dynamic range expansion in step 1 is well restored in the time domain and the frequency domain, and the introduction of the wavelet adaptive denoising algorithm further improves the signal-to-noise ratio of the original signal. Figure 4 In the embodiment, the dashed line is the incorrect restoration of the traditional unwrapping algorithm for large amplitude signals, and the solid line is the demodulation result of the embodiment. It can be seen from the figure that the dynamic range expansion algorithm proposed in the embodiment can effectively restore the original large range signal, and the traditional unwrapping algorithm has the situation of incorrect restoration.

[0070] Figure 5For the signal-to-noise ratio comparison between the traditional winding algorithm and the differential compression-recovery dynamic range expansion algorithm based on wavelet adaptive denoising of the present embodiment, both methods correctly recover the original waveform of the signal, but the dynamic range expansion algorithm of the present embodiment has a higher signal-to-noise ratio.

[0071] In the above method step 2, the barycentric frequency, frequency variance, spectral entropy, sample entropy and normalized wavelet energy feature vector are time-frequency domain related feature values, and the differential compression-recovery dynamic range expansion algorithm based on wavelet adaptive denoising guarantees the accuracy of original signal feature extraction and further improves the accuracy of final classification recognition through effective recovery and signal-to-noise ratio improvement of the signal time-frequency domain. Figure 6 As shown in the figure, I1, I2, I3 and I4 are normal track signals, fish scale damage, wave abrasion damage and empty hanging damage, respectively, and the feature value differences of the four signal samples can be clearly seen, and the four signal samples can be distinguished. Figure 7 As shown in the figure, the differences between the four signal samples can be clearly seen.

[0072] The extracted feature values and normalized wavelet energy feature vectors are further spliced into a multi-dimensional feature vector, pre-weighted, and input into the trained support vector machine SVM for classification and recognition. Figure 8 As shown in the figure, I1, I2, I3 and I4 are normal track signals, fish scale damage, wave abrasion damage and empty hanging damage, respectively. As can be seen from the figure, the above method can realize the monitoring and early warning of track diseases along the heavy haul railway with a length of about 9km, and the early warning results can correspond to the actual track disease positions, fully demonstrating the effectiveness of the present application.

[0073] In summary, the present application obtains the acoustic wave sensing signal by connecting the existing communication optical cable beside the railway, demodulates the original vibration signal S(p, q) by using the differential compression-recovery dynamic range expansion algorithm based on wavelet adaptive denoising, then extracts features from the expanded demodulated signal, splices the features into a multi-dimensional fusion feature vector, pre-weights the feature vector based on the data sample set, finally classifies and recognizes the feature vector using the trained SVM model, records the disease position, and outputs the disease position alarm of the optical fiber length. The present application realizes the recognition and classification of typical track diseases such as fish scale damage, wave abrasion damage and empty hanging damage on the railway, and real-time positioning and alarm of the disease position, providing a long-term online monitoring method for track safety operation and maintenance of railway transportation.

Claims

1. A method for identifying railway track defects based on fiber optic distributed vibration detection, characterized in that, Includes the following steps: Step 1: The fiber optic distributed acoustic wave sensing system (DAS) emits probe light pulses to the sensing fiber to achieve real-time quantitative monitoring of vibration information along the sensing fiber and acquire the original vibration signal S(p,q), where p is the number of acquired pulses and q is the number of segments of acoustic vibration signal acquired along the fiber. The original vibration signal S(p,q) is obtained by demodulating the fiber optic distributed acoustic wave sensing signal using a wavelet adaptive denoising differential compression-restore dynamic range extension algorithm. The specific steps of the wavelet adaptive denoising differential compression-restore dynamic range extension algorithm for demodulating fiber optic distributed acoustic wave sensing signals include: Step 1-1: When demodulating the fiber optic distributed acoustic wave sensing signal, the arctangent demodulation algorithm is used. The step of solving for the phase using the arctangent function will result in the phase value being limited to the range of [-π, π]. At this time, phase expansion is required to obtain the true phase. The classic phase unpacking algorithm is represented as: in, It is the original wound signal. It is the unwound signal; the original wound signal can be represented as Where k(n) is a sequence of integers, the above formula can be further expressed as: in, It is the threshold for unwinding; L-order differential compression of the original phase-wound signal: Where Δ l k(n) is a sequence of integers, so an unwinding algorithm can be used to introduce discrete noise N. noise (n); Step 1-2: Put Consider it as a sinusoidal signal with N discrete samples The sum of, where A i It is the signal amplitude, ω i It is the signal frequency, θ i It is the initial phase of the signal; for each sinusoidal signal Its l-th derivative has a i =A i [2sin(ω i / (2f s ))] l ,b i =θ i +lω i / (2f s ) At this point, the unwinding threshold can be expressed as: When 2sin(ω) i / (2f s ))<1, that is, ω i <πf s / 3 o'clock, Compressed; When l∈[1,5], Φ<π can be satisfied, and then the unwinding can be obtained: When the signal is affected by noise after differentiation, causing Φ to exceed π, the integer sequence Δ l k(n) cannot be completely eliminated by the dewinding algorithm, leaving the misunderstanding term k′(n); directly performing an l-th order integral on the dewinding signal yields: Here, K′(n) is a complex l-th order polynomial, so wavelet denoising algorithm can be used in the differentiation stage to remove the misunderstanding convolution term k′(n) and noise Δ. l N noise (n), thereby achieving correct demodulation; Step 2: Construct an initial dataset from the original vibration signal S(p,q) obtained in Step 1, and label the data samples of normal rail signals and three types of damage signals. Extract time-frequency domain features from the sample data in the initial dataset using centroid frequency, standard deviation frequency, power spectral entropy, sample entropy, and normalized wavelet energy feature vectors, and construct a multi-dimensional fusion feature vector using vector concatenation to obtain the initial multi-dimensional fusion feature vector dataset. The three types of damage signals are fish scale damage, wave abrasion damage, and empty hanging damage. Step 3: Select M sets of data for normal rail signals and three types of defect signals from the initial multidimensional fusion feature vector dataset obtained in Step 2, M≥100, to construct a multidimensional fusion feature vector dataset. Then, pre-weight the multidimensional fusion feature vectors based on the average standard deviation (ASD) of the multidimensional fusion feature vectors of normal rail signals and three types of defect signals as the weights of each feature value of the multidimensional fusion feature vectors to obtain a pre-weighted multidimensional fusion feature vector sample set. Step 4: Input the pre-weighted multidimensional fusion feature vector sample set obtained in Step 3 into the Support Vector Machine (SVM) model for classification training. Use the trained SVM model to classify and identify normal rail signals and three types of defect signals, record the location and type of defects and issue alarms, and realize real-time defect diagnosis and location output.

2. The railway track defect identification method based on fiber optic distributed vibration detection as described in claim 1, characterized in that; The specific steps for extracting time-frequency domain features from the normalized wavelet energy feature vector in step 2 are as follows: Step 2-1: Obtain the normalized wavelet energy eigenvector of the signal; perform wavelet decomposition on the acquired and reconstructed signal, retaining high-frequency wavelet coefficients at each decomposition level, and further decompose the low-frequency wavelet coefficients. After L levels of wavelet decomposition, the low-frequency wavelet coefficients cA can be obtained. L With high-frequency wavelet coefficients cD1, cD2, ... cD L ; Step 2-2: Arrange the high-frequency and low-frequency wavelet coefficients in order from low to high frequency, and solve for the energy E of each wavelet coefficient using equation (7). j The wavelet energy eigenvectors are composed of E = [E1, E2, ..., E...]. L+1 ]; Among them, C j (m) are the decomposed wavelet coefficients, and L is the number of wavelet decomposition levels; Step 2-3: Normalize the wavelet energy feature vector obtained in Step 2-2 according to formula (8) to obtain the normalized wavelet energy feature vector E′=[E′1,E′2,…,E′ L+1 ]; Steps 2-4: Select one set of data for each of the four types of signals in the initial dataset: normal rail signal and three types of defect signals. Perform wavelet decomposition tests, enumerate different wavelet decomposition levels and wavelet functions, and perform wavelet decomposition to obtain normalized wavelet energy feature vectors. Calculate the average standard deviation (ASD) under different decomposition levels and wavelet functions according to formulas (9) and (10) as the distinguishability of the energy vectors. Determine the wavelet decomposition level and wavelet function with the maximum distinguishability so that the four types of signals have the maximum distinguishability. SD j Standard deviation, Let be the mean of the j-th wavelet energy component of the four types of signals, and let a be the corresponding sample type.

3. The railway track defect identification method based on fiber optic distributed vibration detection as described in claim 1, characterized in that: The specific steps for pre-weighting the multi-dimensional fused feature vector in step 3 include: Step 3-1: From the initial dataset constructed in Step 2, select M sets of data for each of the normal rail signal and the three types of defect signals, where M≥100, and construct a multidimensional fusion feature vector dataset. Step 3-2: For each multidimensional fused feature vector, it is composed of four single feature values: centroid frequency, frequency standard deviation, power spectral entropy, and sample entropy, as well as a normalized wavelet energy feature vector of length L+1, for a total length of L+5. Each feature value has a different distinguishing degree for the four types of signals. The signal distinguishing degree corresponding to each feature value is obtained according to formula (11) and used as the weight w of that feature value. n The weight vector for constructing the multidimensional fusion feature vector dataset is [w1, w2, ... w N+5 The obtained multidimensional fusion feature vectors are weighted to construct a pre-weighted multidimensional fusion feature vector [w1c1, w2c2, ... w N+5 c N+5 ]; In the formula, M is the number of sets of each type of signal data in the dataset, and c s Let be the s-th feature value of the multidimensional fusion feature vector, r be the corresponding number of sample groups, and a be the corresponding sample type.

4. The railway track defect identification method based on fiber optic distributed vibration detection as described in claim 1, characterized in that, Step 4 specifically includes: Step 4-1: Use the pre-weighted multidimensional fusion feature vector sample set obtained in Step 3 to train the support vector machine, and obtain the trained support vector machine SVM; Step 4-2: Deploy the trained Support Vector Machine (SVM) in the project. Use the methods in Step 2 and Step 3 to extract features and pre-weight the original vibration signal. Use the pre-weighted multi-dimensional fused feature vector as the input vector of the support vector machine to complete the classification and recognition. Step 4-3: Record the recognition results of the Support Vector Machine (SVM) and output the location and type of the disease to achieve real-time disease diagnosis and location output.

Images