Compensation capacitor fault detection method based on transient characteristics of track circuit

Abstract

The invention relates to the technical field of track monitoring, and discloses a track circuit transient characteristic-based compensation capacitor fault detection method, which comprises the following steps of: acquiring track output voltage signals corresponding to open-circuit faults of compensation capacitors at different positions and in different numbers in real time, and extracting a signal envelope; extracting a modal component enveloped by the signal by adopting a variational mode decomposition method, and adaptively selecting the modal component of the variational mode decomposition method and a secondary penalty factor by adopting a marine predator method to obtain an optimal modal component; according to the optimal modal component and the signal envelope, calculating a mutual information value, and screening and reconstructing the optimal modal component to obtain a reconstructed transient feature vector; acquiring a trained deep shrinkage residual network capable of identifying and positioning transient characteristics, synchronously acquiring a main track output voltage signal of a track circuit receiving end, and identifying and positioning a fault position of a compensation capacitor; the method improves the fault identification and positioning precision, and achieves the real-time monitoring of the state of the compensation capacitor.

Description

Technical Field

[0001] The present invention relates to the technical field of track monitoring, and in particular to a compensation capacitor fault detection method based on transient characteristics of a track circuit. Background Art

[0002] Compensation capacitors are an important part of track circuits, which can improve the transmission performance of signals in track lines and increase the effective transmission distance of signals. Because they work in complex outdoor environments, compensation capacitors are prone to open circuit failures. The failure of a single compensation capacitor will reduce the signal transmission distance. In severe cases, when multiple compensation capacitors in the same section have open circuit failures, it will cause a "red light band" or locomotive signal "dropping code" phenomenon, which may lead to serious accidents.

[0003] At present, in order to ensure the normal operation of the track circuit, the railway site generally adopts the following two combined methods to manage the compensation capacitor equipment: (1) Fault replacement: Through periodic dynamic detection and centralized monitoring of relevant data, the compensation capacitor fault is checked; (2) Expiration replacement: Combined with the ground equipment overhaul period, the compensation capacitor equipment is replaced on schedule. The open circuit fault of a single compensation capacitor has little effect on the transmission characteristics of the track circuit, and it is impossible to directly judge whether the fault has occurred by observing the centralized monitoring data; multiple compensation capacitor open circuit faults can be found through centralized monitoring data, but at this time, additional methods such as track inspection vehicles are required to locate the fault capacitor. Therefore, the existing methods cannot monitor whether the compensation capacitor open circuit fault occurs in real time and locate the fault location in time. In addition, the existing research mainly identifies the compensation capacitor status through track inspection vehicles and locomotive signals. Although the track inspection vehicle can effectively identify the compensation capacitor status and locate the fault location, it can only work during the "window period", which requires additional time and manpower costs. Although the locomotive signal can reflect the compensation capacitor fault problem, it does not have the ability to monitor a specific track section in real time, and there is a situation where safety hazards cannot be avoided in time. Summary of the invention

[0004] In view of the above-mentioned deficiencies in the prior art, the present invention provides a compensation capacitor fault detection method based on the transient characteristics of the track circuit, so as to solve the defect that the prior art cannot identify the working state of the compensation capacitor and locate the fault position in real time and accurately.

[0005] In order to achieve the above-mentioned object of the invention, the technical solution adopted by the present invention is:

[0006] The compensation capacitor fault detection method based on the transient characteristics of the track circuit includes the following steps:

[0007] S1, real-time acquisition of rail-out voltage signals corresponding to open-circuit faults of compensation capacitors at different locations and numbers, and extraction of signal envelopes through Hilbert transform;

[0008] S2, using the variational modal decomposition method to extract the modal components of the signal envelope in step S1, and using the ocean predator method to adaptively select the modal components and the quadratic penalty factor of the variational modal decomposition method to obtain the optimal modal components;

[0009] S3, calculating the mutual information value according to the optimal modal component obtained in step S2 and the signal envelope extracted in step S1, screening and reconstructing the optimal modal component based on the mutual information value, and obtaining a reconstructed transient feature vector;

[0010] S4. Using the reconstructed transient feature vector obtained in step S3, a deep shrinkage residual network capable of identifying and locating transient features is trained to obtain a trained deep shrinkage residual network capable of identifying and locating transient features, which is used to judge and locate the track circuit compensation capacitance according to the transient feature vector;

[0011] S5. Synchronously collect the main rail output voltage signal of the track circuit receiving end, and input it into the deep shrinkage residual network trained in step S4 that can identify and locate transient characteristics, so as to identify and locate the fault position of the compensation capacitor.

[0012] The present invention has the following beneficial effects:

[0013] The compensation capacitor fault detection method based on transient characteristics of track circuit proposed in the present invention collects rail output voltage signals in real time and extracts transient characteristics to train a deep shrinkage residual network capable of identifying and locating transient characteristics, thereby identifying and locating compensation capacitor faults, improving the accuracy of fault identification and positioning, and realizing real-time monitoring of the compensation capacitor status. It can not only identify minor faults such as open circuits of single compensation capacitors, but also prevent multiple open circuit faults of compensation capacitors in the same section, reducing the manpower and time costs spent on locating fault locations at the railway site. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 It is a flow chart of the compensation capacitor fault detection method based on the transient characteristics of the track circuit proposed by the present invention;

[0015] Figure 2 This is a schematic diagram of the ZPW-2000A track circuit structure;

[0016] Figure 3 A schematic diagram of the structure of a deep shrinkage residual network that can identify and locate transient features;

[0017] Figure 4 It is a schematic diagram of the structure of the deep shrinkage residual module;

[0018] Figure 5 It is a schematic diagram of the envelope of the open circuit fault of the compensation capacitor in the embodiment;

[0019] Figure 6 Schematic diagram of the test set recognition results of the embodiment. DETAILED DESCRIPTION

[0020] 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.

[0021] like Figure 1 As shown, the compensation capacitor fault detection method based on the transient characteristics of the track circuit includes the following steps S1-S5:

[0022] S1. Real-time acquisition of rail-out voltage signals corresponding to open-circuit faults of compensation capacitors at different locations and numbers, and extraction of signal envelopes through Hilbert transform.

[0023] In this embodiment, the existing ground monitoring system is used to collect the rail-out voltage signals corresponding to the open circuit faults of the compensation capacitors at different positions and numbers in real time, where the collected quantity is AC voltage, the collection point is the rail-out 1 interface of the attenuator, and the collection frequency is 50kHz. The purpose of performing Hilbert transform on the rail-out voltage signal to extract the signal envelope is to transfer the frequency analysis from the very high fault frequency range to the lower fault frequency range to help simplify the signal. Figure 2 As shown in the figure, the rail output voltage signals corresponding to the open circuit faults of compensation capacitors at different positions and numbers in the ZPW-2000A track circuit structure are collected in real time as the original data, including the disconnection of a single compensation capacitor when the section is idle (C1, C3, C6, C10, C11), the disconnection of two compensation capacitors at the same time (C1C2, C1C6, C3C9, C5C7, C7C12), and the disconnection of three compensation capacitors at the same time (C1C2C3, C1C6C8, C1C10C12, C3C7C8, C 3C9C10) and the rail-out voltage signals when four compensations are disconnected at the same time (C1C2C3C4, C1C6C8C10, C1C9C10C12, C3C5C8C10, C9C10C11C12), where the corresponding fault numbers are 0-19 respectively. The purpose of obtaining this part of the raw data is to extract the transient feature vector of the raw data through subsequent steps, so as to train the deep shrinkage residual network that can identify and locate transient features, thereby identifying the fault location of the compensation capacitor.

[0024] Specifically, step S1 includes S11-S12:

[0025] S11, real-time acquisition of rail-out voltage signals corresponding to open-circuit faults of compensation capacitors at different locations and numbers.

[0026] S12, the rail voltage signal collected in real time in step S11 is subjected to Hilbert transform to obtain an analytical signal, and an envelope is extracted from the analytical signal to obtain a signal envelope, namely:

[0027]

[0028] Where t represents the current time, τ represents the offset between the rail voltage signal and the analytical signal at the current time, and π represents a constant. represents the calculus formula, h(t) represents the analytical signal at the tth moment, x(·) represents the rail output voltage signal, and s(t) represents the envelope signal at the tth moment.

[0029] S2. The modal components of the signal envelope in step S1 are extracted by using a variational modal decomposition method, and the modal components and the quadratic penalty factors of the variational modal decomposition method are adaptively selected by using an ocean predator method to obtain the optimal modal components.

[0030] In this embodiment, the modal components decomposed by the variational modal decomposition method are effective harmonic components and invalid noise components.

[0031] Specifically, step S2 specifically includes S21-S22:

[0032] In this embodiment, the variational mode decomposition can be expressed in the following form, namely:

[0033]

[0034] in, represents the minimum value, K represents the number of modal components, u k represents the kth modal component, ω k represents the center frequency of the kth modal component, δ represents the impulse response, t represents the current moment, δ t represents the impulse response at time t, j represents an imaginary unit, π represents a constant, and e {·} represents the exponential function, and s represents the signal envelope.

[0035] In order to solve the modal components, the above variational modal decomposition formula is rewritten by introducing the augmented Lagrangian equation to extract the modal components of the signal envelope.

[0036] S21, introduce the augmented Lagrangian equation to perform variational modal decomposition to extract the modal components of the signal envelope in step S1, that is:

[0037]

[0038] Where L(·) represents the augmented Lagrangian equation, K represents the number of modal components, and u k represents the kth modal component, ω k represents the center frequency of the kth modal component, λ represents the Lagrange multiplier, α represents the quadratic penalty factor, δ represents the impulse response, t represents the current time, δ t represents the impulse response at time t, j1 represents an imaginary unit, π represents a constant, and e {·} represents the exponential function, and s represents the signal envelope.

[0039] In this embodiment, the augmented Lagrangian equation can use a multiplier to replace the alternating direction multiplier method (ADMM) to calculate K modal components, namely u1(t), u2(t), ..., u K (t).

[0040] S22. The ocean predator method is used to adaptively select the modal components and quadratic penalty factors of the variational modal decomposition method to obtain the optimal modal components.

[0041] In this embodiment, since the number of modal components of variational modal decomposition and the value of the quadratic penalty factor have a great influence on the decomposition effect, the marine predator method is used to adaptively select the modal components and the quadratic penalty factor of the variational modal decomposition method, and in the process of adaptive selection, the fitness is calculated to evaluate the quality of the marine predator individuals and the elite individuals. When calculating the fitness, the envelope entropy is selected as the fitness function, wherein the value of the envelope entropy calculated by the envelope entropy calculation formula is the value of the fitness. The purpose is to use the envelope entropy as an evaluation index of the modal component to measure the size of the noise in the modal component to obtain the optimal modal component.

[0042] Specifically, step S22 specifically includes S221-S226:

[0043] S221, initialize the predator matrix P and the elite matrix E, determine the maximum value (K max ,α max ) and the minimum value (K min ,α min ), where K max , K min Respectively represent the maximum and minimum values ​​of the number of modal components, α max , α min Respectively represent the maximum and minimum values ​​of the quadratic penalty factor, and set the maximum number of iterations MaxIter, that is:

[0044]

[0045] Where n represents the number of predators, d represents the number of parameters to be optimized, and p 1,1represents the parameter value of the first parameter to be optimized of the first initialized predator, p 1,d represents the parameter value of the dth parameter to be optimized of the first initialized predator, p n,1 represents the parameter value of the first parameter to be optimized of the initialized nth predator, p n,d represents the parameter value of the dth parameter to be optimized of the initialized nth predator, p l represents the individual with the best fitness.

[0046] In this embodiment, the i-th row vector of the predator matrix P is the p-th predator p in the predator matrix P. i , that is, p i is a row vector of length d, p n,d is the value of the nth row and dth column of the corresponding predator matrix P, which is a scalar. At the same time, the elite matrix consists of the individuals p with the best fitness. l Copy n times to get.

[0047] S222, according to the current number of iterations, calculate the adaptive parameters for controlling the predator motion compensation, namely:

[0048]

[0049] Among them, CF represents the adaptive parameter controlling the predator motion compensation, and iter represents the current iteration number.

[0050] S223, updating the predator matrix according to the calculated adaptive parameters for controlling predator motion compensation and the current number of iterations to obtain an updated predator matrix, the specific process is as follows:

[0051] According to the adaptive parameters controlling the predator motion compensation, if the current iteration number is If , the predator searches in a small area around the current prey, that is:

[0052]

[0053] Among them, p i represents the i-th predator in the predator matrix P, that is, the i-th row vector of the predator matrix P, e i represents the ith elite of the elite matrix E, that is, the ith row vector of the elite matrix E, R represents a random vector that obeys the (0, 1) uniform distribution, R B represents a random vector based on Brownian motion.

[0054] According to the adaptive parameters controlling the predator motion compensation, if the current iteration number is If the predator is within the range of , the predator will be divided into two groups, one group will continue to search in a small area around the prey, and the other group will expand the search range, that is:

[0055]

[0056] Among them, R L represents a random vector based on the Levy process, p j represents the jth predator in the predator matrix P, e j represents the j-th elite of the elite matrix E.

[0057] According to the adaptive parameters controlling the predator motion compensation, if the current iteration number is Inside, the predator moves closer to the prey, that is:

[0058] p i =e i +0.5×CF×R B ×(R B ×e i -p i )].

[0059] S224. For each p in the updated predator matrix i With e i Perform a variational mode decomposition and calculate p i The fitness and e i The fitness of p i Is the fitness of less than e? i If so, use p i Cover i , update the elite matrix and record e i The minimum value of the fitness is the global optimal value, where the global optimal value is e i The modal component corresponding to the minimum value of the fitness and the quadratic penalty factor, otherwise, execute step S225.

[0060] Specifically, in step S224, for each p in the updated predator matrix i With e i Perform a variational mode decomposition and calculate p i The fitness and e i When the fitness is , the envelope entropy is selected as the fitness function and the value of the envelope entropy is calculated. The calculated value of the envelope entropy is the value of the fitness, that is:

[0061]

[0062] Among them, u i ′(j′) represents the i′th modal component at the j′th sampling point obtained by using the variational mode decomposition method, N represents the number of sampling points, and u′ i′,j′ represents the 0-1 normalized form of the modal component, es i′represents the value of the envelope entropy of the i′th modal component, and log2 represents the logarithmic function.

[0063] In this embodiment, envelope entropy is selected as the fitness function to calculate the value of envelope entropy, and the calculated envelope entropy value is used as the evaluation index of the modal component to measure the size of the noise in the modal component. Since the smaller the envelope entropy, the larger the noise component. Therefore, the parameter (i.e., the corresponding modal component) is selected according to the minimum envelope entropy value to achieve the best variational modal decomposition effect, thereby minimizing the total noise component obtained.

[0064] S225, update the updated predator matrix according to the eddy effect FADs, obtain the updated predator matrix P′, and generate a random number r∈(0,1), the specific process is:

[0065] If r≤FADs, the predator's partial position is reset, that is:

[0066] p i =p i +CF×[P min +R×(P max -P min )]×U

[0067] If r>FADs, the predator moves randomly, that is:

[0068]

[0069] Among them, P min represents the minimum value of the predator matrix, P max represents the maximum value of the predator matrix, U represents a random vector containing only 0 and 1, They represent any two different random predators in the predator matrix, and r1 and r2 represent different random values ​​that obey uniform distribution.

[0070] S226. Determine whether the predator matrix P′ updated again has reached the maximum number of iterations. If so, terminate the iteration to obtain the optimal modal component and quadratic penalty factor. Otherwise, return to step S222 and continue to iterate until the maximum number of iterations is reached. Stop the iteration to obtain the final updated predator matrix, that is, the optimal modal component and quadratic penalty factor.

[0071] S3. Calculate the mutual information value according to the optimal modal component obtained in step S2 and the signal envelope extracted in step S1, and screen and reconstruct the optimal modal component based on the mutual information value to obtain a reconstructed transient feature vector.

[0072] In this embodiment, since the variational modal decomposition method decomposes the signal envelope into multiple modal components, the modal components at this time can be regarded as effective harmonic components and invalid noise components. Since the decomposed modal components have been adaptively selected in step S2, the optimal modal components u1′(t), u2′(t), ..., u K ′(t). However, the optimal modal components still contain harmonic components and noise components, so it is necessary to calculate the mutual information value to further filter and remove the noise components. The smaller the mutual information value between the modal component and the signal envelope, the closer the modal component is to the noise. Therefore, in this embodiment, the modal components with mutual information values ​​less than 0.3 are regarded as noise components and removed, and the remaining modal components (harmonic components) are reconstructed to obtain the reconstructed transient feature vector.

[0073] Specifically, S3 specifically includes S31-S32:

[0074] S31, according to the optimal modal component obtained in step S2 and the signal envelope extracted in step S1, the mutual information value of the optimal modal component and the signal envelope is calculated using the k-nearest neighbor estimation method, that is:

[0075]

[0076] Among them, u k ′ represents the optimal kth modal component, s represents the signal envelope, I(u k ′; s) represents the optimal kth modal component u k ′ is the mutual information value of the signal envelope s, ψ(·) represents the double gamma function, k′ represents the number of selected neighboring points, <·> represents the average value, and N represents the total logarithm of the optimal modal component and signal envelope at the same time. n s They represent the optimal kth modal component u k ′ and the value of the signal envelope at the same time.

[0077] S32, using the mutual information value calculated in step S31 to filter out the noise component in the optimal modal component and reconstruct the remaining harmonic components to obtain a reconstructed transient feature vector.

[0078] S4. Use the reconstructed transient feature vector obtained in step S3 to train a deep shrinkage residual network capable of identifying and locating transient features, so as to obtain a trained deep shrinkage residual network capable of identifying and locating transient features, which is used to judge and locate the compensation capacitance of the track circuit according to the transient feature vector.

[0079] Specifically, the deep shrinkage residual network capable of identifying and locating transient features includes a one-dimensional convolutional layer, four deep shrinkage residual modules, a flattening layer, a fully connected layer and a normalization layer; wherein the deep shrinkage residual module includes a first one-dimensional convolutional layer, a first batch of normalization layers, a linear rectification function activation layer, a second one-dimensional convolutional layer, a second batch of normalization layers, a global average pooling layer, a fully connected layer, a third batch of normalization layers and a Sigmoid function activation layer.

[0080] like Figure 3 As shown, Figure 3 The structure of a deep contraction residual network capable of identifying and locating transient features is shown in Figure 3 The medium-depth shrinkage residual network includes a one-dimensional convolutional layer, four deep shrinkage residual modules, a flattening layer, a fully connected layer and a normalization layer. The one-dimensional convolutional layer and the four deep shrinkage residual modules are used to further extract the multi-channel features of the input signal, that is, the multi-channel features of the input reconstructed transient feature vector. The flattening layer is used to expand the multi-channel features into one dimension. The fully connected layer acts as a classifier, that is, outputs the probability of the category to which the data belongs according to the high-dimensional features learned previously. The normalization layer is used to normalize the category probability and output the state corresponding to the maximum value in the result.

[0081] Specifically, step S4 specifically includes S41-S46:

[0082] S41, dividing the reconstructed transient feature vector obtained in step S3 into a training set, a validation set and a test set according to a certain ratio.

[0083] S42, input the training set into the one-dimensional convolutional layer to extract the shallow multi-channel features of the reconstructed transient feature vector, that is:

[0084]

[0085] Among them, t represents the current time, represents the time point of the reconstructed transient feature vector, x′(t) represents the shallow multi-channel features of the reconstructed transient feature vector at the tth moment, Indicates The transient feature vector reconstructed at time t, g(·) represents the weight function.

[0086] S43, inputting the shallow multi-channel features extracted in step S42 into four deep shrinkage residual modules to extract deep multi-channel features to obtain deep multi-channel features.

[0087] Specifically, step S43 specifically includes S431-S437:

[0088] S431. Input the shallow multi-channel features into the first one-dimensional convolution layer, the first batch of standardization layers and the linear rectification function activation layer in sequence to perform one-dimensional convolution operation, batch normalization operation and nonlinear function transformation respectively to obtain single-layer convolution multi-channel features.

[0089] S432. Input the single-layer convolution multi-channel features into the second one-dimensional convolution layer and the second batch normalization layer in sequence to perform one-dimensional convolution operation and batch normalization operation respectively, to obtain a double-layer convolution multi-channel features.

[0090] S433, input the double-layer convolution multi-channel features into the global average pooling layer to obtain the average value of the channel.

[0091] S434, input the average value of the channel into the fully connected layer and the third batch of normalized layers in sequence for feature fusion to obtain the weight attenuation coefficient.

[0092] S435, input the weight attenuation coefficient into the Sigmoid function activation layer for threshold mapping to obtain a normalized attenuation coefficient.

[0093] S436 , multiplying the normalized attenuation coefficient by the average value of the channel to obtain the threshold of the channel.

[0094] S437. Perform a soft threshold operation on the double-layer convolution multi-channel feature according to the channel threshold to remove noise, obtain a denoised double-layer convolution multi-channel feature, and add the shallow multi-channel feature to the denoised double-layer convolution multi-channel feature to obtain a deep multi-channel feature.

[0095] Specifically, the deep multi-channel features in step S437 are:

[0096]

[0097] Among them, y′ represents the deep multi-channel feature, x′ represents the shallow multi-channel feature, c represents the denoised double-layer convolutional multi-channel feature, f2 represents the double-layer convolutional multi-channel feature, and τ′ represents the channel threshold.

[0098] like Figure 4 As shown, Figure 4 The structure of the deep shrinkage residual module is shown, Figure 4 The structure of the medium-depth shrinkage residual module includes the first one-dimensional convolution layer, the first batch of standardization layers, the linear rectification function activation layer, the second one-dimensional convolution layer, the second batch of standardization layers, the global average pooling layer, the fully connected layer, the third batch of standardization layers, and the Sigmoid function activation layer, where x′ represents the shallow multi-channel feature, f1 represents the single-layer convolution multi-channel feature, and f2 represents the double-layer convolution multi-channel feature. Represents the average value of the channel, z represents the weight attenuation coefficient, α′ represents the normalized attenuation coefficient, τ′ represents the threshold of the channel, and c represents the denoised double-layer convolutional multi-channel feature. Among them, the deep multi-channel feature y′ is the feature finally extracted by the deep residual module, and the soft threshold operation can force unimportant features to be 0, so as to make the learned high-level features more discriminative. At the same time, the partial derivative of the soft threshold function is 0 or 1, so it can prevent the problem of gradient disappearance and gradient explosion.

[0099] S44. Input the deep multi-channel features into the flattening layer, expand the deep multi-channel features into one dimension, obtain the one-dimensional features and input them into the fully connected layer for classification, obtain the probabilities of different categories and input them into the normalization layer for probability normalization, and obtain a converged deep shrinkage residual network that can identify and locate transient features.

[0100] S45. Input the validation set into the converged deep shrinkage residual network capable of identifying and locating transient features obtained in step S44 for performance evaluation, and obtain the deep shrinkage residual network capable of identifying and locating transient features with the best performance.

[0101] In this embodiment, after the deep shrinkage residual network model capable of identifying and locating transient features converges, it is necessary to use a validation set to evaluate the performance of the deep shrinkage residual network model capable of identifying and locating transient features. By evaluating the performance of different network structures or hyperparameter models on the validation set, the best performing deep shrinkage residual network model capable of identifying and locating transient features is selected as the final deep shrinkage residual network model capable of identifying and locating transient features.

[0102] S46. Input the test set into the deep shrinkage residual network with the best performance that can identify and locate transient features obtained in step S45 to perform compensation capacitor fault identification and location, obtain a trained deep shrinkage residual network that can identify transient features, and output the fault numbers corresponding to the open circuit faults of compensation capacitors at different locations and numbers, and obtain the location of the compensation capacitor fault.

[0103] In this embodiment, after obtaining the deep shrinkage residual network model with optimal performance that can identify and locate transient features, the test set is input into the deep shrinkage residual network with optimal performance that can identify and locate transient features to perform compensation capacitor fault identification and location, thereby ending the training process.

[0104] S5. Synchronously collect the main rail output voltage signal of the track circuit receiving end, and input it into the deep shrinkage residual network trained in step S4 that can identify and locate transient characteristics, so as to identify and locate the fault position of the compensation capacitor.

[0105] In this embodiment, the main rail output voltage signal of the track circuit receiving end is synchronously collected, the collected amount is AC voltage, the collection point is the attenuator rail output 1 interface, the sampling frequency is 50kHz, and the transient feature vector of the main rail output voltage signal is extracted. The transient feature vector is input into a trained deep shrinkage residual network that can identify and locate transient features for fault identification and location, and the fault number and position of the open circuit fault of the compensation capacitor are obtained.

[0106] In order to verify the effectiveness of the compensation capacitor fault detection method based on the transient characteristics of the track circuit proposed in the present invention, a physical simulation platform is built in a laboratory environment to simulate the simultaneous disconnection process of a single or multiple compensation capacitors at different positions when the section is idle, and the real track output voltage data of the receiving end is collected for verification. Figure 5 As shown, Figure 5 The envelopment diagram of some of the collected data is shown. Figure 5 It can be seen from the figure that when the compensation capacitor is open, the voltage first drops rapidly to the minimum value, then rises slightly, and then remains stable. This process is called a transient process. Among them, the disconnection of C1C2 and C1C2C3 has a similar impact on the rail voltage. It is impossible to distinguish the two situations directly by the rail voltage. Combining the transient process can improve the recognition accuracy. For the 12 compensation capacitors in the same section, the number of compensation capacitor fault types within 4 is huge. 20 compensation capacitor fault combinations are randomly selected for analysis. The distribution of experimental data is shown in Table 1:

[0107] Table 1 Dataset distribution table

[0108] Fault number Failure Point Train / Validate / Test Fault number Failure Point Train / Validate / Test 0 C1 800 / 100 / 100 10 C1C2C3 800 / 100 / 100 1 C3 800 / 100 / 100 11 C1C6C8 800 / 100 / 100 2 C6 800 / 100 / 100 12 C1C10C12 800 / 100 / 100 3 C10 800 / 100 / 100 13 C3C7C8 800 / 100 / 100 4 C11 800 / 100 / 100 14 C3C9C10 800 / 100 / 100 5 C1C2 800 / 100 / 100 15 C1C2C3C4 800 / 100 / 100 6 C1C6 800 / 100 / 100 16 C1C6C8C10 800 / 100 / 100 7 C3C9 800 / 100 / 100 17 C1C9C10C12 800 / 100 / 100 8 C5C7 800 / 100 / 100 18 C3C5C8C10 800 / 100 / 100 9 C7C12 800 / 100 / 100 19 C9C10C11C12 800 / 100 / 100

[0109] Table 1 is divided into three parts: training set, validation set and test set. The training set is used to train the model, the validation set is used to select the model with the best hyperparameters, and finally the test set is used to evaluate the performance of the model. The experimental results are shown in Figure 6 As shown, from Figure 6 It can be seen that the intelligent recognition method for compensation capacitor faults based on transient characteristics of track circuits proposed by the present invention has 37 recognition errors, and the transient recognition accuracy is 98.15%, which can accurately identify the compensation capacitor fault number and locate the fault location. And the precision of the model proposed by this method is 97.1%, the recall rate is 97.3%, and the F value is 0.97, indicating that the trained deep shrinkage residual network model that can recognize and locate transient characteristics has excellent performance in multi-compensation capacitor fault location.

[0110] In summary, the compensation capacitor fault detection method based on the transient characteristics of the track circuit proposed in the present invention collects the rail output voltage signal in real time and extracts the transient characteristics to train a deep shrinkage residual network that can identify and locate the transient characteristics, thereby identifying and locating the compensation capacitor fault, improving the fault identification and positioning accuracy, and realizing real-time monitoring of the compensation capacitor status. It can not only identify the minor faults of a single compensation capacitor open circuit, but also prevent multiple compensation capacitor open circuit faults in the same section, reducing the manpower and time costs spent on locating the fault position at the railway site.

[0111] The present invention uses specific embodiments to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only used to help understand the method of the present invention and its core idea. At the same time, for those skilled in the art, according to the idea of ​​the present invention, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as a limitation on the present invention.

[0112] Those skilled in the art will appreciate that the embodiments described herein are intended to help readers understand the principles of the present invention, and should be understood that the protection scope of the present invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific variations and combinations that do not deviate from the essence of the present invention based on the technical revelations disclosed by the present invention, and these variations and combinations are still within the protection scope of the present invention.

Claims

1. A compensation capacitor fault detection method based on transient characteristics of track circuits, characterized in that: The following steps are involved: S1, real-time acquisition of rail-out voltage signals corresponding to open-circuit faults of compensation capacitors at different locations and numbers, and extraction of signal envelopes through Hilbert transform; S2, using the variational modal decomposition method to extract the modal components of the signal envelope in step S1, and using the ocean predator method to adaptively select the modal components and the quadratic penalty factor of the variational modal decomposition method to obtain the optimal modal components; S3, calculating the mutual information value according to the optimal modal component obtained in step S2 and the signal envelope extracted in step S1, screening and reconstructing the optimal modal component based on the mutual information value, and obtaining a reconstructed transient feature vector; S4. Using the reconstructed transient feature vector obtained in step S3, a deep shrinkage residual network capable of identifying and locating transient features is trained to obtain a trained deep shrinkage residual network capable of identifying and locating transient features, which is used to judge and locate the track circuit compensation capacitance according to the transient feature vector; S5. Synchronously collect the main rail output voltage signal of the track circuit receiving end, and input it into the deep shrinkage residual network trained in step S4 that can identify and locate transient characteristics, so as to identify and locate the fault position of the compensation capacitor.

2. The compensation capacitor fault detection method based on track circuit transient characteristics according to claim 1 is characterized in that: Step S1 specifically includes: S11, real-time acquisition of rail output voltage signals corresponding to open circuit faults of compensation capacitors at different positions and numbers; S12, the rail voltage signal collected in real time in step S11 is subjected to Hilbert transform to obtain an analytical signal, and an envelope is extracted from the analytical signal to obtain a signal envelope, namely: Where t represents the current time, τ represents the offset between the rail voltage signal and the analytical signal at the current time, and π represents a constant. represents the calculus formula, h(t) represents the analytical signal at the tth moment, x(·) represents the rail output voltage signal, and s(t) represents the envelope signal at the tth moment.

3. The compensation capacitor fault detection method based on track circuit transient characteristics according to claim 1 is characterized in that: Step S2 specifically includes: S21, introduce the augmented Lagrangian equation to perform variational modal decomposition to extract the modal components of the signal envelope in step S1, that is: Where L(·) represents the augmented Lagrangian equation, K represents the number of modal components, and u k represents the kth modal component, ω k represents the center frequency of the kth modal component, λ represents the Lagrange multiplier, α represents the quadratic penalty factor, δ represents the impulse response, t represents the current time, δ t represents the impulse response at time t, j1 represents an imaginary unit, π represents a constant, and e {·} represents the exponential function, s represents the signal envelope; S22. The ocean predator method is used to adaptively select the modal components and quadratic penalty factors of the variational modal decomposition method to obtain the optimal modal components.

4. The compensation capacitor fault detection method based on track circuit transient characteristics according to claim 3 is characterized in that: Step S22 specifically includes: S221, initialize the predator matrix P and the elite matrix E, determine the maximum value (K) of the parameter to be optimized (K, α) max , α max ) and the minimum value (K min , α min ), where K max , K min Respectively represent the maximum and minimum values ​​of the number of modal components, α max , α min Respectively represent the maximum and minimum values ​​of the quadratic penalty factor, and set the maximum number of iterations MaxIter, that is: Where n represents the number of predators, d represents the number of parameters to be optimized, and p 1，1 represents the parameter value of the first parameter to be optimized of the first initialized predator, p 1，d represents the parameter value of the dth parameter to be optimized of the first initialized predator, p n，1 represents the parameter value of the first parameter to be optimized of the initialized nth predator, p n，d represents the parameter value of the dth parameter to be optimized of the initialized nth predator, p l represents the individual with the best fitness; S222, according to the current number of iterations, calculate the adaptive parameters for controlling the predator motion compensation, namely: Among them, CF represents the adaptive parameter controlling the predator motion compensation, and iter represents the current iteration number; S223, updating the predator matrix according to the calculated adaptive parameters for controlling predator motion compensation and the current number of iterations to obtain an updated predator matrix, the specific process is as follows: According to the adaptive parameters controlling the predator motion compensation, if the current iteration number is If , the predator searches in a small area around the current prey, that is: Among them, p i represents the i-th predator in the predator matrix P, that is, the i-th row vector of the predator matrix P, e i represents the ith elite of the elite matrix E, that is, the ith row vector of the elite matrix E, R represents a random vector that obeys the (0, 1) uniform distribution, R B represents a random vector based on Brownian motion; According to the adaptive parameters controlling the predator motion compensation, if the current iteration number is If the predator is within the range of , the predator will be divided into two groups, one group will continue to search in a small area around the prey, and the other group will expand the search range, that is: Among them, R L represents a random vector based on the Levy process, p j represents the jth predator in the predator matrix P, e j represents the j-th elite in the elite matrix E; According to the adaptive parameters controlling the predator motion compensation, if the current iteration number is Inside, the predator moves closer to the prey, that is: p i =e i +0.5×CF×[R B ×(R B ×e i -p i )]; S224. For each p in the updated predator matrix i With e i Perform a variational mode decomposition and calculate p i The fitness and e i The fitness of p i Is the fitness of less than e? i If so, use p i Cover i , update the elite matrix and record e i The minimum value of the fitness is the global optimal value, where the global optimal value is e i The modal component and the quadratic penalty factor corresponding to the minimum value of the fitness, otherwise, execute step S225; S225, updating the updated predator matrix according to the eddy current effect FADs, obtaining a re-updated predator matrix P′, and generating a random number r∈(0,1), the specific process is: If r≤FADs, the predator's partial position is reset, that is: p i =p i +CF×[P min +R×(P max -P min )]×U If r>FADs, the predator moves randomly, that is: Among them, P mm represents the minimum value of the predator matrix, P max represents the maximum value of the predator matrix, U represents a random vector containing only 0 and 1, They represent any two different predators in the predator matrix, r1 and r2 represent different random values ​​that obey uniform distribution; S226. Determine whether the predator matrix P′ updated again has reached the maximum number of iterations. If so, terminate the iteration to obtain the optimal modal component and quadratic penalty factor. Otherwise, return to step S222 and continue to iterate until the maximum number of iterations is reached. Stop the iteration to obtain the final updated predator matrix, that is, the optimal modal component and quadratic penalty factor.

5. The compensation capacitor fault detection method based on track circuit transient characteristics according to claim 4 is characterized in that: In step S224, for each p in the updated predator matrix i With e i Perform a variational mode decomposition and calculate p i The fitness and e i When the fitness is , the envelope entropy is selected as the fitness function and the value of the envelope entropy is calculated. The calculated value of the envelope entropy is the value of the fitness, that is: Among them, u i′ (j′) represents the i′th modal component at the j′th sampling point obtained by using the variational mode decomposition method, N represents the number of sampling points, and u′ i′，j′ represents the 0-1 normalized form of the modal component, es i′ represents the value of the envelope entropy of the i′th modal component, and log2 represents the logarithmic function.

6. The compensation capacitor fault detection method based on track circuit transient characteristics according to claim 1 is characterized in that: S3 specifically includes: S31, according to the optimal modal component obtained in step S2 and the signal envelope extracted in step S1, the mutual information value of the optimal modal component and the signal envelope is calculated using the k-nearest neighbor estimation method, that is: Among them, u k ′ represents the optimal kth modal component, s represents the signal envelope, I(u k ′; s) represents the optimal kth modal component u k ′ is the mutual information value of the signal envelope s, ψ(·) represents the double gamma function, k′ represents the number of selected neighboring points, <·> represents the average value, and N represents the total logarithm of the optimal modal component and signal envelope at the same time. n s Respectively represent the optimal kth modal component u k ′ is the value taken at the same time as the signal envelope; S32, using the mutual information value calculated in step S31 to filter out the noise component in the optimal modal component and reconstruct the remaining harmonic components to obtain a reconstructed transient feature vector.

7. The compensation capacitor fault detection method based on track circuit transient characteristics according to claim 1 is characterized in that: The deep shrinkage residual network capable of identifying and locating transient features includes a one-dimensional convolutional layer, four deep shrinkage residual modules, a flattening layer, a fully connected layer, and a normalization layer; Among them, the deep shrinkage residual module includes the first one-dimensional convolution layer, the first batch of standardization layers, the linear rectification function activation layer, the second one-dimensional convolution layer, the second batch of standardization layers, the global average pooling layer, the fully connected layer, the third batch of standardization layers and the Sigmoid function activation layer.

8. The compensation capacitor fault detection method based on track circuit transient characteristics according to claim 7 is characterized in that: Step S4 specifically includes: S41, dividing the reconstructed transient feature vector obtained in step S3 into a training set, a validation set and a test set according to a certain ratio; S42, input the training set into a one-dimensional convolutional layer to extract the shallow multi-channel features of the reconstructed transient feature vector, that is: Among them, t represents the current time, represents the time point of the reconstructed transient feature vector, x′(t) represents the shallow multi-channel features of the reconstructed transient feature vector at the tth moment, Indicates The transient feature vector reconstructed at time instant, g(·) represents the weight function; S43, inputting the shallow multi-channel features extracted in step S42 into four deep shrinkage residual modules to extract deep multi-channel features, thereby obtaining deep multi-channel features; S44, inputting the deep multi-channel features into the flattening layer, expanding the deep multi-channel features into one dimension, obtaining the one-dimensional features and inputting them into the fully connected layer for classification, obtaining the probabilities of different categories and inputting them into the normalization layer for probability normalization, and obtaining a converged deep shrinkage residual network capable of identifying and locating transient features; S45, inputting the validation set into the converged deep shrinkage residual network capable of identifying and locating transient features obtained in step S44 for performance evaluation, and obtaining the deep shrinkage residual network capable of identifying and locating transient features with the best performance; S46. Input the test set into the deep shrinkage residual network with the best performance that can identify and locate transient features obtained in step S45 to perform compensation capacitor fault identification and location, obtain a trained deep shrinkage residual network that can identify transient features, and output the fault numbers corresponding to the open circuit faults of compensation capacitors at different locations and numbers, and obtain the location of the compensation capacitor fault.

9. The compensation capacitor fault detection method based on track circuit transient characteristics according to claim 8 is characterized in that: Step S43 specifically includes: S431, inputting the shallow multi-channel features into the first one-dimensional convolution layer, the first batch of standardization layers and the linear rectification function activation layer in sequence, respectively performing one-dimensional convolution operation, batch normalization operation and nonlinear function transformation, to obtain single-layer convolution multi-channel features; S432, inputting the single-layer convolution multi-channel features into the second one-dimensional convolution layer and the second batch normalization layer in sequence to perform one-dimensional convolution operation and batch normalization operation respectively, to obtain a double-layer convolution multi-channel features; S433, inputting the double-layer convolution multi-channel features into the global average pooling layer to obtain the average value of the channel; S434, inputting the average value of the channel into the fully connected layer and the third batch of normalized layers in sequence for feature fusion to obtain a weight attenuation coefficient; S435, inputting the weight attenuation coefficient into the Sigmoid function activation layer for threshold mapping to obtain a normalized attenuation coefficient; S436, multiplying the normalized attenuation coefficient by the average value of the channel to obtain a threshold value of the channel; S437. Perform a soft threshold operation on the double-layer convolution multi-channel feature according to the channel threshold to remove noise, obtain a denoised double-layer convolution multi-channel feature, and add the shallow multi-channel feature to the denoised double-layer convolution multi-channel feature to obtain a deep multi-channel feature.

10. The compensation capacitor fault detection method based on track circuit transient characteristics according to claim 9 is characterized in that: The deep multi-channel features in step S437 are: Among them, y′ represents the deep multi-channel feature, x′ represents the shallow multi-channel feature, c represents the denoised double-layer convolutional multi-channel feature, f2 represents the double-layer convolutional multi-channel feature, and τ′ represents the channel threshold.

Images