Track circuit compensation capacitor fault prediction method and system

Abstract

The invention relates to the technical field of railway signal processing, and discloses a track circuit compensation capacitor fault prediction method and system. The method comprises the following steps: acquiring real-time impedance spectroscopy data by using a distributed capacitive sensor network; performing multi-scale wavelet decomposition on the data, and extracting frequency domain energy distribution and transient abnormal fluctuation characteristics; constructing a dynamic Bayesian network model, and updating hidden state node posterior probability distribution through a variational inference algorithm; generating a fault risk index sequence based on a posterior probability, and designing an adaptive threshold adjustment strategy by using a reinforcement learning framework; constructing a mixed integer dynamic programming model to output fault early warning grade data; and constructing a multi-stage decision tree which comprises a feature screening layer, a pattern matching layer and a verification feedback layer. The method and the system can accurately predict the track circuit compensation capacitor fault in real time, improve the fault prediction accuracy and the system adaptability, reduce the false alarm rate, and provide reliable basis for railway operation and maintenance.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of railway signal processing technology, in particular to a track circuit compensation capacitor fault prediction method and system. BACKGROUND

[0002] In the modern railway transportation system, track circuit is a key device to ensure the safe operation of trains and realize information interaction between trains and signal systems. Its performance stability is crucial. Track circuit compensation capacitor is one of the core components of track circuit. Its role is to improve the transmission performance of track circuit and ensure stable and efficient signal transmission.

[0003] Due to the extremely complex railway operating environment, track circuit compensation capacitors are long-term exposed to harsh conditions such as strong electromagnetic interference, mechanical vibration, drastic temperature changes, and high humidity. In such an environment, compensation capacitors are prone to failure. Once the compensation capacitor fails, it will change the transmission characteristics of the track circuit, reduce the signal transmission quality, and even cause signal interruption. This not only affects the normal operation order of the train, causing train delays, shutdowns, and other problems, but also poses a serious threat to the safe operation of the train, such as causing train collisions, derailments, and other major accidents.

[0004] For fault detection of track circuit compensation capacitors, periodic inspection is mostly used. This method has obvious limitations. On the one hand, the periodic inspection cycle is relatively fixed, making it difficult to detect problems at the first time of failure, which may cause the fault to continue to affect the normal operation of the track circuit during the period before detection. On the other hand, periodic inspection mainly relies on human experience and simple detection equipment, and the accuracy and reliability of the detection results are easily affected by human factors, making it difficult to accurately judge the potential failure of the capacitor.

[0005] In addition, some existing fault detection technologies often only focus on a single performance indicator of the capacitor, lacking comprehensive and integrated analysis of the capacitor's operating state. Since the manifestation of capacitor failure is diverse, different types of faults may be reflected in multiple performance indicators. Relying solely on a single indicator for detection can easily miss some potential fault hazards. Moreover, these technologies usually detect and repair after the fault occurs, which is a passive maintenance method that cannot predict the occurrence of faults in advance, cannot provide enough warning time for maintenance personnel, and cannot meet the growing demand for safety and reliability in railway transportation.

[0006] With the development of high speed and heavy load of railway transportation, the train operation density is increasing, and higher requirements are put forward for the stability and reliability of track circuit. The traditional track circuit compensation capacitor fault detection and maintenance method has been unable to adapt to this development trend. Therefore, developing a method and system capable of predicting track circuit compensation capacitor fault in real time and accurately has important practical significance for ensuring the stable operation of railway signal system and improving the efficiency and safety of railway transportation. SUMMARY

[0007] The purpose of the present application is to provide a track circuit compensation capacitor fault prediction method and system to solve the problems raised in the background art.

[0008] To achieve the above purpose, the present application provides the following technical solution: a track circuit compensation capacitor fault prediction method, the method comprising:

[0009] Collecting real-time impedance spectrum data of the track circuit through a distributed capacitor sensor network;

[0010] Performing multi-scale wavelet decomposition on the real-time impedance spectrum data to extract frequency energy distribution features and transient abnormal fluctuation features;

[0011] Constructing a dynamic Bayesian network model, inputting the frequency energy distribution features and transient abnormal fluctuation features as observation nodes, and updating the posterior probability distribution of the hidden state nodes through a variational inference algorithm;

[0012] Generating a fault risk indicator sequence based on the posterior probability distribution, designing an adaptive threshold adjustment strategy using a reinforcement learning framework, and taking the minimization of risk indicator volatility rate and the equalization of false alarm rate as optimization objectives;

[0013] Constructing a mixed integer dynamic programming model according to the adaptive threshold adjustment strategy, and outputting fault warning level data through the mixed integer dynamic programming model;

[0014] Constructing a multi-stage decision tree according to the fault warning level data, the multi-stage decision tree comprising a feature screening layer, a pattern matching layer and a verification feedback layer, wherein the feature screening layer iteratively removes redundant features based on information gain rate, the pattern matching layer constructs a fault pattern probability distribution graph using a kernel density estimation method, and the verification feedback layer updates the pattern library through an online incremental learning mechanism.

[0015] Preferably, the multi-scale wavelet decomposition comprises:

[0016] Performing continuous wavelet transform on the impedance spectrum data using a Morlet wavelet basis function to generate a time-frequency energy matrix;

[0017] The time-frequency energy matrix is subjected to multifractal denoising processing, and an effective frequency band and a noise frequency band are segmented by a Hurst index threshold value;

[0018] A double-channel feature extraction network is constructed, a first channel uses discrete wavelet packet transform to extract low-frequency energy entropy features, and a second channel extracts high-frequency phase mutation features through complex wavelet coherence analysis;

[0019] The low-frequency energy entropy features and the high-frequency phase mutation features are spliced through a feature fusion layer to generate a joint time-frequency feature vector.

[0020] Preferably, the dynamic Bayesian network model adopts a hierarchical hidden Markov structure, and comprises:

[0021] Three layers of hidden state nodes are defined, the first layer represents the degree of capacitor aging, the second layer represents the insulation medium loss, and the third layer reflects the contact resistance abnormality;

[0022] A bidirectional message passing mechanism is designed, a particle filtering algorithm is used for forward propagation to estimate instantaneous state probability, and a time sequence back propagation algorithm is used for backward propagation to correct historical state confidence;

[0023] An explainability constraint module is constructed, an attention mechanism is used to dynamically adjust the sparsity of the state transition matrix, and the traceability of the physical process is ensured.

[0024] Preferably, the mixed integer dynamic programming model adopts a hierarchical genetic algorithm to globally optimize the fault prediction parameters, wherein the hierarchical genetic algorithm integrates an elite reservation strategy and an adaptive crossover and mutation probability mechanism; comprising:

[0025] The fault prediction parameters are encoded into a mixed chromosome structure, including discrete fault type identification genes and continuous threshold adjustment genes;

[0026] When initializing the population, Latin hypercube sampling is used to ensure uniform coverage of the parameter space;

[0027] In the crossover operation, a multi-point crossover strategy is used for discrete genes, and a simulated binary crossover is performed for continuous genes;

[0028] A dynamic mutation operator is designed, the mutation strength is adaptively adjusted according to the population diversity index, and when the Hamming distance is lower than a set threshold, a Gaussian perturbation mutation is triggered.

[0029] Preferably, the kernel density estimation method constructs a fault mode probability distribution diagram, comprising:

[0030] An Epanechnikov kernel function is used for non-parametric density estimation of historical fault samples;

[0031] A local adaptive bandwidth selection mechanism is introduced to dynamically adjust the kernel function smoothing coefficient according to the sample space density.

[0032] A probability contour mapping model is constructed to accelerate the two-dimensional convolution operation through fast Fourier transform, and generate the probability heat map of the fault mode.

[0033] Preferably, the online incremental learning mechanism updates the mode library, including:

[0034] A drift detection module is designed to check the distribution change of the monitoring data.

[0035] When concept drift is detected, an incremental support vector machine is activated to adjust the boundary of new samples.

[0036] A memory replay buffer is constructed, and an importance sampling strategy is used to balance the weights of new and old samples to prevent catastrophic forgetting.

[0037] Preferably, the complex wavelet coherence analysis extracts high-frequency phase mutation features, including:

[0038] The analytic signal of the impedance spectrum signal is subjected to Hilbert transform to extract the instantaneous phase derivative;

[0039] A complex wavelet transform domain phase gradient detector is constructed to identify the mutation point position through the phase consistency criterion;

[0040] An adaptive threshold segmentation algorithm is used to eliminate false mutation interference and retain significant phase jump events.

[0041] Preferably, the attention mechanism dynamically adjusts the sparsity of the state transition matrix, including:

[0042] A state correlation degree index is defined to measure the difference between adjacent state distributions based on KL divergence;

[0043] A sparsity control coefficient is generated through a gated recurrent unit to constrain the proportion of non-zero elements in the state transition matrix;

[0044] A projection gradient descent method is used to solve the constrained matrix optimization problem to ensure the normalization characteristics of the transition probability.

[0045] Preferably, the dynamic mutation operator is designed using a multi-objective optimization framework, including:

[0046] A mutation intensity evaluation function is constructed, which includes the weighted combination of population convergence index and distribution index;

[0047] The NSGA-II algorithm is used to perform Pareto front search on the mutation parameters;

[0048] The fuzzy logic controller dynamically adjusts the multi-objective weight coefficient according to the iteration number and convergence speed.

[0049] Preferably, the present application also includes a track circuit compensation capacitor fault prediction system, the system comprising:

[0050] A data acquisition module: for acquiring real-time impedance spectrum data of the track circuit through a distributed capacitor sensor network;

[0051] A feature extraction module: for multi-scale wavelet decomposition of the real-time impedance spectrum data, extracting frequency energy distribution features and transient abnormal fluctuation features;

[0052] A probability updating module: for constructing a dynamic Bayesian network model, inputting the frequency energy distribution features and transient abnormal fluctuation features as observation nodes, and updating the posterior probability distribution of hidden state nodes through a variational inference algorithm;

[0053] A threshold optimization module: for generating a fault risk indicator sequence based on the posterior probability distribution, and designing an adaptive threshold adjustment strategy using a reinforcement learning framework, the reinforcement learning framework taking risk indicator volatility minimization and false alarm rate equalization as optimization objectives;

[0054] An early warning decision module: for constructing a mixed integer dynamic programming model according to the adaptive threshold adjustment strategy, and outputting fault warning level data through the mixed integer dynamic programming model;

[0055] A decision tree construction module: for constructing a multi-stage decision tree according to the fault warning level data, the multi-stage decision tree including a feature screening layer, a pattern matching layer, and a verification feedback layer, wherein the feature screening layer iteratively eliminates redundant features based on information gain rate, the pattern matching layer constructs a fault pattern probability distribution map using a kernel density estimation method, and the verification feedback layer updates a pattern library through an online incremental learning mechanism.

[0056] Compared with the prior art, the present application has the following advantages:

[0057] At the data acquisition and feature extraction level, real-time impedance spectrum data is acquired using a distributed capacitor sensor network, which can comprehensively and accurately obtain the operating information of the track circuit compensation capacitor. By extracting frequency energy distribution features and transient abnormal fluctuation features through multi-scale wavelet decomposition, key information about the health status of the capacitor can be deeply mined from the data. Traditional methods may only be able to obtain limited capacitor parameters, making it difficult to detect subtle performance changes, while the present application can sensitively capture these feature changes, providing a solid data foundation for subsequent fault prediction, greatly improving the accuracy and timeliness of fault prediction.

[0058] The construction of the dynamic Bayesian network model is a highlight. The extracted features are input into the model as observation nodes, and the posterior probability distribution of the hidden state nodes is updated by means of the variational inference algorithm, so that the model can dynamically simulate the state change of the capacitor under different operating conditions. The model not only considers the current observation data, but also combines historical state information, compared with the traditional static model, can more accurately reflect the actual operating condition of the capacitor, effectively reduce the false alarm rate and the missing alarm rate, and provide more reliable fault prediction results for railway operation and maintenance personnel.

[0059] The adaptive threshold adjustment strategy adopts a reinforcement learning framework, with the optimization objectives of minimizing the risk indicator volatility and balancing the false alarm rate, and dynamically adjusts the fault judgment threshold according to the real-time operation of the track circuit. The traditional fixed threshold method cannot adapt to the complex and variable railway operating environment, and is prone to misjudgment. The adaptive strategy of the present application can automatically optimize the threshold to ensure accurate prediction of faults under different operating conditions, improving the adaptability and robustness of the fault prediction system.

[0060] The mixed integer dynamic programming model optimizes the fault prediction parameters globally through the hierarchical genetic algorithm, effectively improving the prediction accuracy and efficiency of the model. The hierarchical genetic algorithm integrates the elite reservation strategy and the adaptive crossover and mutation probability mechanism, which can avoid the algorithm falling into local optimal solution and quickly search for the optimal combination of fault prediction parameters. This optimization method enables the model to more accurately output fault warning level data, providing a more scientific basis for railway operation and maintenance decision-making.

[0061] The construction of the multi-stage decision tree further enhances the reliability and interpretability of fault prediction. The feature selection layer iteratively removes redundant features based on information gain rate, reducing noise and interference in the data and improving the running efficiency and prediction accuracy of the decision tree. The pattern matching layer uses kernel density estimation method to construct the fault mode probability distribution graph, which intuitively shows the probability of occurrence of different fault modes, facilitating the understanding and analysis of operation and maintenance personnel. The verification feedback layer updates the pattern library through online incremental learning mechanism, which can continuously adapt to new fault modes and operating data, and continuously improve the performance of fault prediction.

[0062] In summary, the track circuit compensation capacitor fault prediction method and system of the present application can accurately predict faults in advance, reduce the delay and loss caused by train operation faults, improve the safety and reliability of railway transportation, and reduce the operation and maintenance cost, with significant economic and social benefits. BRIEF DESCRIPTION OF DRAWINGS

[0063] Figure 1 The working principle diagram of the track circuit compensation capacitor fault prediction method described in the present application;

[0064] Figure 2 The flowchart of multi-scale wavelet decomposition;

[0065] Figure 3 Workflow diagram for constructing failure mode probability distribution for kernel density estimation

[0066] Figure 4 Principle diagram designed for dynamic variation operator DETAILED DESCRIPTION

[0067] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0068] Please refer to Figures 1-4 The present application provides a technical solution: a track circuit compensation capacitor fault prediction method, the method comprising:

[0069] Real-time impedance spectrum data of the track circuit is collected by using a distributed capacitor sensor network. These sensors are reasonably distributed at key positions of the track circuit, and can accurately obtain impedance spectrum information of the track circuit during operation in real time, providing a raw data basis for subsequent analysis.

[0070] The collected real-time impedance spectrum data is subjected to multi-scale wavelet decomposition. Through this operation, frequency energy distribution features and transient abnormal fluctuation features are extracted. These features can reflect the energy changes of the track circuit compensation capacitor at different frequencies and abnormal fluctuations during operation, providing key evidence for subsequent judgment of whether the capacitor is faulty.

[0071] A dynamic Bayesian network model is constructed, and the extracted frequency energy distribution features and transient abnormal fluctuation features are input into the model as observation nodes. With the aid of variational inference algorithm, the posterior probability distribution of the hidden state node is updated. The model can simulate the change of the running state of the track circuit compensation capacitor with time, and more accurately evaluate the health status of the capacitor through the update of the posterior probability distribution.

[0072] Based on the updated posterior probability distribution, a fault risk indicator sequence is generated. An adaptive threshold adjustment strategy is designed using a reinforcement learning framework, with the optimization objectives of minimizing the risk indicator volatility rate and equalizing the false alarm rate. This strategy can dynamically adjust the threshold value of fault judgment according to the actual operation of the track circuit, improving the accuracy of fault prediction.

[0073] According to the adaptive threshold adjustment strategy, a mixed integer dynamic programming model is constructed. Through the model output fault warning level data, the fault risk of the capacitor is divided into different grades, which is convenient for the staff to understand the fault degree of the capacitor in time and take corresponding measures.

[0074] According to the fault warning level data, a multi-stage decision tree is constructed, which includes a feature screening layer, a pattern matching layer and a verification feedback layer. The feature screening layer iteratively removes redundant features based on information gain rate, removing unnecessary interference information; the pattern matching layer uses kernel density estimation method to construct fault mode probability distribution graph, which intuitively shows the probability distribution of fault occurrence; the verification feedback layer updates the pattern library through online incremental learning mechanism, continuously improves the accuracy and adaptability of the decision tree.

[0075] The application will be further described below in conjunction with examples 1 to 6:

[0076] Example 1:

[0077] When performing multi-scale wavelet decomposition on the real-time impedance spectrum data of the track circuit, Morlet wavelet basis function is used for continuous wavelet transform of the impedance spectrum data. Morlet wavelet basis function is a commonly used wavelet function, which has good time-frequency localization characteristics and can effectively convert impedance spectrum data from time domain to time-frequency domain to generate time-frequency energy matrix. In this time-frequency energy matrix, each element represents the energy distribution at a specific time and frequency.

[0078] The generated time-frequency energy matrix is subjected to multi-fractal denoising processing. The Hurst index is calculated to threshold segment the effective frequency band and the noise frequency band. The Hurst index is an index for measuring the long-term correlation of time series, which can determine in this embodiment which frequency band is the effective frequency band reflecting the running state of the track circuit and which is the noise frequency band. When the Hurst index is greater than a certain pre-set threshold, the corresponding frequency band is considered as the effective frequency band, otherwise it is the noise frequency band. In this way, the interference of noise can be removed and the accuracy of subsequent analysis can be improved.

[0079] A double-channel feature extraction network is constructed. The first channel uses discrete wavelet packet transform to extract low-frequency energy entropy features. Discrete wavelet packet transform is an extension of discrete wavelet transform, which can divide the frequency band of the signal more finely. The low-frequency energy entropy feature reflects the uncertainty of the energy distribution of the track circuit compensation capacitor in the low-frequency band. The greater the entropy value, the more dispersed the energy distribution, and the more unstable the state of the capacitor.

[0080] The second channel extracts the high-frequency phase mutation feature through complex wavelet coherence analysis. The specific process is as follows: first, the analytic signal of the impedance spectrum signal is subjected to Hilbert transform to extract the instantaneous phase derivative. Hilbert transform can convert real signals into analytic signals, thereby facilitating the extraction of instantaneous phase information. Then, a complex wavelet transform domain phase gradient detector is constructed to identify the mutation point position through the phase consistency criterion. Phase consistency is a method for detecting feature points in images or signals, which is used in this embodiment to determine the position of high-frequency phase mutation. Finally, an adaptive threshold segmentation algorithm is used to eliminate false mutation interference and retain significant phase jump events. In this way, the high-frequency phase mutation feature can be accurately extracted, which is of great significance for judging the sudden failure of the capacitor.

[0081] The low-frequency energy entropy feature and the high-frequency phase mutation feature are spliced in the tensor dimension through the feature fusion layer to generate a joint time-frequency feature vector. Tensor splicing is an operation of merging different features in the tensor dimension, and the joint time-frequency feature vector generated in this way comprehensively reflects the low-frequency and high-frequency information, and more comprehensively reflects the running state of the track circuit compensation capacitor.

[0082] Embodiment 2:

[0083] The dynamic Bayesian network model adopts a hierarchical hidden Markov structure. Three layers of hidden state nodes are defined, the first layer representing the aging degree of the capacitor. The internal medium of the capacitor will gradually age during long-term use, leading to a decline in capacitor performance. The aging degree of the capacitor is quantified through this layer of hidden state nodes, for example, a numerical value can be used to represent the percentage of aging, and the larger the numerical value, the more serious the aging. The second layer represents the insulation medium loss. Insulation medium loss will affect the insulation performance of the capacitor, and thus affect the normal operation of the track circuit. The insulation medium loss situation can be reflected by measuring parameters such as the loss tangent of the capacitor, and the second layer of hidden state nodes is an abstract representation of these parameters. The third layer reflects the contact resistance anomaly. Contact resistance anomalies can cause unstable signal transmission, and the change in contact resistance is reflected in the third layer of hidden state nodes through monitoring.

[0084] A bidirectional message passing mechanism is designed. The forward propagation estimates the instantaneous state probability using a particle filter algorithm. The particle filter algorithm is a filtering algorithm based on the Monte Carlo method, which represents the probability distribution of the state by randomly sampling a large number of particles in the state space. In this embodiment, the particle filter algorithm is used to predict the instantaneous state probability at the current time based on the current observation data and the state estimation at the previous time. The backward propagation corrects the historical state confidence by a time sequence back propagation algorithm. The time sequence back propagation algorithm is an algorithm for training recurrent neural networks, which can correct the state confidence at the previous time according to the subsequent observation data in the dynamic Bayesian network, so that the estimation of the historical state is more accurate.

[0085] An explainability constraint module is constructed, and an attention mechanism is used to dynamically adjust the sparsity of the state transition matrix. A state correlation degree index is defined, and the difference between the state distributions at adjacent times is measured based on KL divergence. The KL divergence is an index for measuring the difference between two probability distributions. Here, the KL divergence between the state distributions at adjacent times is calculated to determine the degree of correlation between the states. A sparsity control coefficient is generated by a gated recurrent unit to constrain the proportion of non-zero elements in the state transition matrix. The gated recurrent unit is a special recurrent neural network structure that can effectively process time series data. The sparsity control coefficient generated by the gated recurrent unit can adjust which elements in the state transition matrix are important and which elements can be ignored. A projection gradient descent method is used to solve the constrained matrix optimization problem to ensure the normalization characteristics of the transition probabilities. The projection gradient descent method is a method for solving optimization problems under constraints, which can ensure that the sum of the transition probabilities of the state transition matrix is 1, which meets the basic definition of probability.

[0086] Embodiment 3:

[0087] The mixed integer dynamic programming model uses a hierarchical genetic algorithm to globally optimize the fault prediction parameters. The hierarchical genetic algorithm integrates an elite reservation strategy and an adaptive crossover and mutation probability mechanism. The fault prediction parameters are encoded into a mixed chromosome structure, including discrete fault type identification genes and continuous threshold adjustment genes. The discrete fault type identification genes can represent different fault types with different numbers or symbols, such as 0 for normal, 1 for capacitor aging fault, 2 for insulation medium loss fault, etc. The continuous threshold adjustment gene is used to dynamically adjust the threshold for fault judgment, which is a continuously changing value.

[0088] Latin hypercube sampling is used to initialize the population to ensure uniform coverage of the parameter space. Latin hypercube sampling is an efficient sampling method that can uniformly select sample points in the parameter space, so that the initial population can better cover the entire parameter space and improve the search efficiency of the algorithm.

[0089] In the crossover operation, a multi-point crossover strategy is adopted for discrete genes. Multi-point crossover refers to randomly selecting multiple crossover points on the chromosome and then exchanging gene segments of the two parent chromosomes at these crossover points. This can increase the diversity of genes and improve the search ability of the algorithm. Simulated binary crossover is implemented for continuous genes. Simulated binary crossover is a crossover operation method for continuous variables, which simulates the binary number crossover process. It generates offspring genes by linear combination of parent genes, so that offspring genes can inherit the advantages of parent genes.

[0090] A dynamic mutation operator is designed using a multi-objective optimization framework. A mutation strength evaluation function is constructed, which includes a weighted combination of population convergence indicators and distribution indicators. The population convergence indicator can be measured by the similarity between individuals in the population. The higher the similarity, the more convergent the population. The distribution indicator can be measured by the uniformity of the population distribution in the parameter space. By weighting the combination of these two indicators, a comprehensive mutation strength evaluation function is obtained. The NSGA-II algorithm is used to search the Pareto front of the mutation parameters. NSGA-II is a commonly used multi-objective optimization algorithm that can find the optimal balance solution between multiple objectives, i.e. the Pareto front. The fuzzy logic controller dynamically adjusts the multi-objective weight coefficients according to the iteration number and convergence speed. The fuzzy logic controller is a controller based on fuzzy rules, which can dynamically adjust the weight coefficients of population convergence indicators and distribution indicators according to the changes of iteration number and convergence speed, so that the mutation operator can better adapt to different optimization stages. Gaussian perturbation mutation is triggered when the Hamming distance is below a certain threshold. Hamming distance is an indicator used to measure the difference between two strings, which is used here to measure the difference between individuals in the population. When the Hamming distance is too low, it means that the population may have fallen into a local optimal solution, so the Gaussian perturbation mutation is triggered to add Gaussian noise to the genes, increase the diversity of genes, and help the algorithm escape from the local optimal solution.

[0091] Example 4:

[0092] Kernel density estimation method is used to construct the fault mode probability distribution graph in the pattern matching layer. Epanechnikov kernel function is used for non-parametric density estimation of historical fault samples. Epanechnikov kernel function is a commonly used kernel function with good smoothing and local adaptability. In this embodiment, Epanechnikov kernel function is used to weight the sum of historical fault samples, so as to estimate the probability density function of fault occurrence. Assuming that the historical fault samples are x1, x2, …, x n , the kernel function is K(x), and the formula of kernel density estimation is:

[0093]

[0094] where, is the estimated probability density function, n is the sample size, h is the bandwidth, which controls the smoothness of the kernel function. The larger the bandwidth, the smoother the estimated probability density function, but some local information may be lost; the smaller the bandwidth, the closer the estimated probability density function to the true distribution of the sample, but overfitting may occur.

[0095] A local adaptive bandwidth selection mechanism is introduced to dynamically adjust the smoothing coefficient of the kernel function according to the sample space density. The sample space density can be measured by calculating the number of samples around the sample point. When the number of samples around the sample point is large, it means that the sample space density is large, at this time the bandwidth can be appropriately reduced to improve the estimation accuracy; when the number of samples around the sample point is small, it means that the sample space density is small, at this time the bandwidth can be appropriately increased to avoid overfitting. Through this local adaptive bandwidth selection mechanism, the kernel density estimation can more accurately reflect the distribution of the fault samples.

[0096] A probability contour mapping model is constructed, and a fast Fourier transform is used to speed up the two-dimensional convolution operation to generate a probability heat map of the fault mode. Fast Fourier transform is an efficient algorithm for calculating discrete Fourier transform, which can convert time domain signals to frequency domain signals, thereby greatly improving the speed of convolution operation. In this embodiment, the probability density function obtained by kernel density estimation is subjected to two-dimensional convolution operation by using fast Fourier transform to obtain a probability heat map of the fault mode. In the probability heat map, the darker the color, the higher the probability of failure, so that the distribution of the fault mode can be intuitively displayed, facilitating the analysis and judgment of the staff.

[0097] Embodiment 5:

[0098] In the verification feedback layer, the application updates the mode library through an online incremental learning mechanism to adapt to the changing track circuit operation data, ensuring the accuracy and timeliness of fault prediction.

[0099] A drift detection module is designed, and its main purpose is to test the changes in the distribution of the monitoring data. In the actual operation of the track circuit, due to environmental factors (such as changes in temperature and humidity), gradual aging of equipment, and fluctuations in train operation conditions, the distribution of monitoring data may change. In order to timely capture these changes, the drift detection module uses the statistical hypothesis testing method, specifically using the Kolmogorov-Smirnov test. Assuming that the current collected monitoring data sequence is X = {x1, x2, …, x m}, the historical data sequence is Y = {y1, y2, …, y n} The Kolmogorov-Smirnov test calculates the maximum distance D between the empirical cumulative distribution functions F X (x) of the two data sequences, i.e. Y (x) and F α (x) is greater than the critical value D at a significance level a, it is determined that concept drift has occurred, indicating that the distribution of the current monitoring data is significantly different from the distribution of the historical data.

[0100] When the drift detection module detects concept drift, it activates the incremental support vector machine to adjust the boundary of the new samples. The incremental support vector machine (ISVM) is developed on the basis of the traditional support vector machine, which can efficiently process new samples on the basis of the existing model, avoiding the high-cost operation of retraining the model for all data. Assuming that the classification hyperplane of the existing support vector machine model is w T φ(x) + b = 0, where w is the weight vector, φ(x) is a function that maps the input data x to a high-dimensional feature space, and b is the bias term. When there is a new sample x new , ISVM adjusts w and b so that the classification hyperplane can better adapt to the new data distribution. Specifically, ISVM updates the Lagrange multipliers using the information of the new sample, and then adjusts the classification hyperplane. For example, using the incremental version of the SMO (Sequential Minimal Optimization) algorithm, a pair of samples that violate the KKT (Karush-Kuhn-Tucker) condition is selected for optimization each time, and the model parameters are updated gradually to effectively learn the new samples and make the classification boundary more consistent with the characteristics of the current data.

[0101] A memory replay buffer is constructed to store historical sample data. In order to balance the weights of new and old samples and prevent the model from forgetting important information in old samples when learning new samples, i.e., to avoid catastrophic forgetting, an importance sampling strategy is used. In the memory replay buffer, each sample is assigned an importance weight. One way to calculate the importance weight of a sample is based on the similarity of the sample to the current data distribution. For example, the kernel density estimation is used to calculate the probability density p(x i under the current data distribution, then the importance weight w i of the sample x i can be defined as i ​When sampling samples from the memory replay buffer for model training, sampling is performed according to the importance weights, and the greater the weight of a sample, the higher the probability of being sampled. In this way, in the training process, the model can fully consider the information of new and old samples, both using the real-time information of new samples to adapt to the change of data distribution and retaining the long-term regularity contained in old samples, thereby improving the stability and generalization ability of the model.

[0102] Embodiment 6:

[0103] In the process of extracting high-frequency phase mutation features by complex wavelet coherence analysis, the Hilbert transform is performed on the analytic signal of the impedance spectrum signal to extract the instantaneous phase derivative. The role of Hilbert transform is to convert the real signal into an analytic signal, so that the instantaneous phase information of the signal can be easily extracted. By taking the derivative of the phase of the analytic signal, the instantaneous phase derivative can be obtained.

[0104] A complex wavelet transform domain phase gradient detector is constructed to identify the position of the mutation point through the phase consistency criterion. The phase consistency criterion is a method for judging feature points based on the phase information of the signal. In the complex wavelet transform domain, when the phase information in multiple scales and directions has consistency, it is considered that there is a mutation point at this position. By calculating the phase gradient and combining the phase consistency criterion, the position of the high-frequency phase mutation point can be accurately identified.

[0105] An adaptive threshold segmentation algorithm is used to eliminate false mutation interference and retain significant phase jump events. The adaptive threshold segmentation algorithm can automatically adjust the threshold according to the local features of the signal. In this embodiment, by analyzing the signal strength, frequency and other features at the phase mutation point, a threshold is dynamically determined. When the amplitude of the phase mutation exceeds this threshold, it is considered to be a significant phase jump event and is retained; otherwise, it is considered to be false mutation interference and is removed.

[0106] In terms of dynamically adjusting the sparsity of the state transition matrix of the attention mechanism, a state correlation degree index is defined, and the difference between the state distributions of adjacent time points is measured based on KL divergence. Let the state distributions of adjacent time points be P1 and P2, respectively, and the calculation formula of the KL divergence is:

[0107]

[0108] where i represents the value of the state. The greater the value of the KL divergence, the greater the difference between the two state distributions, and the smaller the correlation degree between the states.

[0109] The sparsity control coefficient is generated by a gated recurrent unit. The gated recurrent unit generates a sparsity control coefficient according to input state correlation indicators and the like through a gating mechanism and a recurrent structure of the gated recurrent unit. The coefficient can be used to adjust which elements in the state transition matrix should be zero and which elements should be retained, thereby controlling the sparsity of the state transition matrix.

[0110] The projection gradient descent method is used to solve the constrained matrix optimization problem, to ensure the normalization of the transition probability. In each iteration of the projection gradient descent method, the gradient of the objective function is first calculated, and then the state transition matrix is updated in the opposite direction of the gradient while satisfying the constraint condition (the sum of the transition probabilities is 1). Through continuous iteration, the state transition matrix gradually converges to an optimal solution, while ensuring the normalization of the transition probability.

[0111] In the aspect of designing a dynamic mutation operator using a multi-objective optimization framework, a mutation strength evaluation function is constructed, which includes a weighted combination of a population convergence index and a distribution index. Let the population convergence index be C, the distribution index be D, and the weighting coefficients be w1 and w2, respectively. Then the mutation strength evaluation function is:

[0112] S = w1C + w2D

[0113] where S represents the mutation strength. The population convergence index C can be measured by the variance of the fitness of individuals in the population. The smaller the variance, the more convergent the population is. The distribution index D can be measured by the coverage range of the population in the parameter space. The larger the coverage range, the better the distribution is.

[0114] The NSGA-II algorithm is used to perform a Pareto front search on the mutation parameters. The NSGA-II algorithm finds the optimal balanced solution between multiple objectives (such as minimizing mutation strength and maximizing population diversity) through non-dominated sorting and congestion calculation, and obtains the Pareto front. The combination of mutation parameters on the Pareto front has achieved a good balance between different objectives.

[0115] The fuzzy logic controller dynamically adjusts the multi-objective weighting coefficients according to the iteration number and convergence speed. The fuzzy logic controller takes the iteration number and convergence speed as input according to pre-set fuzzy rules, and outputs dynamically adjusted weighting coefficients w1 and w2. For example, when the iteration number is small and the convergence speed is slow, the weight of the distribution index w2 can be appropriately increased to improve the diversity of the population; when the iteration number is large and the convergence speed is fast, the weight of the population convergence index w1 can be appropriately increased to speed up the convergence speed of the algorithm.

[0116] It is to be understood that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting; it is not intended to exclude myriad other embodiments of the present application that other inventors can develop based on the same general inventive concepts embodied by the described embodiments. That is, although the present application is described in terms of particular embodiments and illustrative figures, it should be apparent that the scope of the present application is not limited to these specific embodiments.

[0117] While the embodiments of the application have been shown and described herein, it will be understood by those skilled in the art that many changes, modifications, substitutions and alterations to these embodiments can be made without departing from the principles and spirits of the application, and it is intended that the scope of the application be limited solely by the scope of the appended claims and the equivalents thereof.

Claims

1. A method for predicting track circuit compensation capacitor faults, characterized in that: include: Collect real-time impedance spectrum data of track circuits through a distributed capacitance sensor network; Performing multi-scale wavelet decomposition on the real-time impedance spectrum data to extract frequency domain energy distribution characteristics and transient abnormal fluctuation characteristics; A dynamic Bayesian network model is constructed, the frequency domain energy distribution characteristics and transient abnormal fluctuation characteristics are used as observation node inputs, and the posterior probability distribution of the implicit state node is updated through a variational inference algorithm; Generate a fault risk indicator sequence based on the posterior probability distribution, and design an adaptive threshold adjustment strategy using a reinforcement learning framework, wherein the reinforcement learning framework takes minimizing the volatility of the risk indicator and balancing the false alarm rate as optimization goals; constructing a mixed integer dynamic programming model according to the adaptive threshold adjustment strategy, and outputting fault warning level data through the mixed integer dynamic programming model; A multi-stage decision tree is constructed based on the fault warning level data. The multi-stage decision tree includes a feature screening layer, a pattern matching layer and a verification feedback layer. The feature screening layer iteratively eliminates redundant features based on the information gain rate, the pattern matching layer uses the kernel density estimation method to construct a fault mode probability distribution map, and the verification feedback layer updates the pattern library through an online incremental learning mechanism.

2. The method for predicting track circuit compensation capacitor failure according to claim 1, characterized in that: The multi-scale wavelet decomposition comprises: The impedance spectrum data were transformed into a time-frequency energy matrix using the Morlet wavelet basis function. Performing multifractal denoising on the time-frequency energy matrix, and dividing the effective frequency band and the noise frequency band by using a Hurst exponent threshold; A dual-channel feature extraction network is constructed. The first channel uses discrete wavelet packet transform to extract low-frequency energy entropy features, and the second channel uses complex wavelet coherence analysis to extract high-frequency phase mutation features. The low-frequency energy entropy feature and the high-frequency phase mutation feature are tensor-concatenated through a feature fusion layer to generate a joint time-frequency feature vector.

3. The method for predicting track circuit compensation capacitor failure according to claim 1, characterized in that: The dynamic Bayesian network model adopts a hierarchical hidden Markov structure, including: Define three layers of implicit state nodes: the first layer represents the degree of capacitor aging, the second layer represents the loss of insulation medium, and the third layer reflects the abnormal contact resistance; A two-way message passing mechanism is designed. The forward propagation uses a particle filter algorithm to estimate the instantaneous state probability, and the backward propagation uses a time series backpropagation algorithm to correct the historical state confidence. An interpretability constraint module is constructed, and the attention mechanism is used to dynamically adjust the sparsity of the state transfer matrix to ensure the traceability of the physical process.

4. The method for predicting track circuit compensation capacitor failure according to claim 1, characterized in that: The mixed integer dynamic programming model uses a hierarchical genetic algorithm to globally optimize fault prediction parameters, wherein the hierarchical genetic algorithm integrates an elite retention strategy and an adaptive crossover mutation probability mechanism; include: Encode the fault prediction parameters into a hybrid chromosome structure, which includes discrete fault type identification genes and continuous threshold adjustment genes; Latin hypercube sampling is used when initializing the population to ensure uniform coverage of the parameter space; In the crossover operation, a multi-point crossover strategy is used for discrete genes, and a simulated binary crossover is implemented for continuous genes; A dynamic mutation operator is designed to adaptively adjust the mutation intensity according to the population diversity index, and Gaussian perturbation mutation is triggered when the Hamming distance is lower than the set threshold.

5. The method for predicting track circuit compensation capacitor failure according to claim 1, characterized in that: The kernel density estimation method for constructing a failure mode probability distribution diagram includes: The Epanechnikov kernel function is used to perform non-parametric density estimation on historical fault samples; A local adaptive bandwidth selection mechanism is introduced to dynamically adjust the kernel function smoothing coefficient according to the sample space density; A probability contour mapping model is constructed, and the two-dimensional convolution operation is accelerated by fast Fourier transform to generate a probability heat map of the fault mode.

6. The method for predicting track circuit compensation capacitor failure according to claim 1, characterized in that: The online incremental learning mechanism update pattern library includes: Design a drift detection module to test changes in monitoring data distribution; When concept drift is detected, the incremental support vector machine is activated to adjust the boundaries of the newly added samples; Construct a memory replay buffer and use an importance sampling strategy to balance the weights of new and old samples to prevent catastrophic forgetting.

7. The method for predicting track circuit compensation capacitor failure according to claim 2, characterized in that: The complex wavelet coherence analysis to extract high-frequency phase mutation features includes: Perform Hilbert transform on the analytical signal of the impedance spectrum signal and extract the instantaneous phase derivative; A phase gradient detector is constructed in the complex wavelet transform domain, and the position of the mutation point is identified by the phase consistency criterion. An adaptive threshold segmentation algorithm is used to eliminate pseudo-mutation interference and retain significant phase jump events.

8. The method for predicting track circuit compensation capacitor failure according to claim 3, characterized in that: The attention mechanism dynamically adjusts the sparsity of the state transfer matrix including: Define the state correlation index, which measures the difference in state distribution at adjacent moments based on KL divergence; Generate sparsity control coefficients through gated recurrent units to constrain the proportion of non-zero elements in the state transfer matrix; The projected gradient descent method is used to solve the constrained matrix optimization problem to ensure the normalization of the transition probability.

9. The method for predicting track circuit compensation capacitor failure according to claim 4, characterized in that: The dynamic mutation operator is designed using a multi-objective optimization framework, including: Construct a mutation intensity evaluation function, which includes a weighted combination of population convergence index and distribution index; The NSGA-II algorithm is used to perform Pareto front search on the mutation parameters; The multi-objective weight coefficients are dynamically adjusted according to the number of iterations and convergence speed through the fuzzy logic controller.

10. A track circuit compensation capacitor fault prediction system, characterized in that: include: Data acquisition module: used to collect real-time impedance spectrum data of track circuits through a distributed capacitance sensor network; Feature extraction module: performing multi-scale wavelet decomposition on the real-time impedance spectrum data to extract frequency domain energy distribution characteristics and transient abnormal fluctuation characteristics; Probability update module: constructs a dynamic Bayesian network model, takes the frequency domain energy distribution characteristics and transient abnormal fluctuation characteristics as observation node input, and updates the posterior probability distribution of the implicit state node through the variational inference algorithm; Threshold Optimization Module: Generates a fault risk indicator sequence based on the posterior probability distribution and designs an adaptive threshold adjustment strategy using a reinforcement learning framework. The reinforcement learning framework takes minimizing the volatility of the risk indicator and balancing the false alarm rate as optimization goals; Early warning decision module: constructing a mixed integer dynamic programming model according to the adaptive threshold adjustment strategy, and outputting fault early warning level data through the mixed integer dynamic programming model; Decision tree construction module: Construct a multi-stage decision tree based on the fault warning level data. The multi-stage decision tree includes a feature screening layer, a pattern matching layer and a verification feedback layer. The feature screening layer iteratively eliminates redundant features based on the information gain rate, the pattern matching layer uses the kernel density estimation method to construct a fault mode probability distribution map, and the verification feedback layer updates the pattern library through an online incremental learning mechanism.

Images