Track circuit fault diagnosis and life prediction method, device, equipment and medium

Abstract

The application discloses a track circuit fault diagnosis and life prediction method, device, equipment and medium, and relates to the technical field of track circuit fault diagnosis and life prediction. The method acquires multi-modal signals of a track circuit under different working states; performs adaptive decomposition, feature extraction and time-frequency feature fusion on the multi-modal signals, detects whether the track circuit has a fault and a fault type when the fault exists by using a fault diagnosis model; the fault diagnosis model is obtained based on a width learning system training, a BLS incremental algorithm is used during training, and a method of actively selecting a training set is used to correct data used in each iteration training; according to the multi-modal signals of the track circuit without a fault under an adjustment state, a life prediction model is used to determine the residual life of each device of the track circuit; the life prediction model comprises a width learning system and a bidirectional long short-term memory neural network. The application can comprehensively perceive the overall health state of the track circuit.

Description

Technical Field

[0001] This application relates to the field of track circuit fault diagnosis and life prediction technology, and in particular to a method, apparatus, equipment and medium for track circuit fault diagnosis and life prediction. Background Technology

[0002] Track circuits are the core signaling equipment of railway train operation control systems. Their stable operation directly determines train dispatching efficiency and traffic safety. Providing intelligent operation and maintenance for track circuits throughout their entire lifecycle is a key link in ensuring the long-term reliable and continuous stable operation of the railway system, and an important development direction for improving the level of intelligent operation of railway transportation. Therefore, fault diagnosis and prediction of track circuits are of paramount importance.

[0003] In track circuit fault diagnosis, traditional methods rely heavily on convolutional neural networks, deep convolutional neural networks, or fault type analysis modeling. While these methods can accurately identify existing faults, they only passively respond to current faults and cannot proactively predict potential faults or assess the aging trends of equipment over long-term operation, thus failing to meet the "prevention is better than cure" maintenance requirements. In track circuit fault prediction, traditional methods often employ long short-term memory networks, Gath-Geva (GG) fuzzy clustering, convolutional neural networks with recurrent units, or spatiotemporal feature extraction. However, research often focuses on fault prediction for single devices (such as compensation capacitors, analog cables, and transmitters) or predicting the changing trends of single monitoring data (such as track-out voltage and track-in voltage). This fails to achieve comprehensive monitoring and analysis of the entire track circuit system and prevents the construction of a complete intelligent maintenance system spanning the entire equipment lifecycle. Summary of the Invention

[0004] The purpose of this application is to provide a method, apparatus, equipment, and medium for fault diagnosis and life prediction of track circuits, which can comprehensively perceive the overall health status of track circuits.

[0005] To achieve the above objectives, this application provides the following solution.

[0006] In a first aspect, this application provides a method for fault diagnosis and life prediction of track circuits, including: Acquire multi-mode signals of the track circuit under different operating states; different operating states include adjustment state and shunt state; The multimodal signal is adaptively decomposed, and the modal components obtained by adaptive decomposition are feature extracted to obtain the time-frequency features of the multimodal signal. By fusing the time-frequency features of the multimodal signals, fused features are obtained; Based on the fusion features, a fault diagnosis model is used to detect whether there is a fault in the track circuit and the type of fault when a fault exists. The fault diagnosis model is trained based on a wide learning system. When training the wide learning system, the BLS (Broad Learning System) incremental algorithm is used, and the data used in each iteration of training is corrected by actively selecting the training set. If no fault is detected in the track circuit, the remaining lifespan of each device in the track circuit is determined using a lifespan prediction model based on the multimodal signal of the track circuit in the adjustment state without fault; the lifespan prediction model includes a width learning system and a bidirectional long short-term memory neural network connected in sequence.

[0007] Optionally, the multimode signal includes: a voltage signal and a frequency shift keying signal; the voltage signal includes: main rail output voltage, main rail input voltage, small rail output voltage, small rail input voltage, transmitter output voltage, receiver receiving voltage, transmitter cable analog network disk voltage, receiver cable analog network disk voltage, transmitter matching transformer voltage, and receiver matching transformer voltage.

[0008] Optionally, the multimodal signal is adaptively decomposed, and the modal components obtained by the adaptive decomposition are feature extracted to obtain the time-frequency features of the multimodal signal. Specifically, this includes: adaptively decomposing the multimodal signal using symplectic geometric mode decomposition to obtain multiple modal components; and using minimum envelope entropy to extract features from the multiple modal components to obtain the time-domain and frequency-domain features of the multimodal signal.

[0009] Optionally, the process of training the width learning system in each iteration includes: The training sample set is divided into categories, and the category with the fewest samples is determined as the basic sample category; Retain all samples from the base sample category and select from each sample category other than the base sample category. These samples together constitute the training subset; among them... The number of samples in the basic sample categories; the basic sample categories are the th [number] in the training sample set. Each sample category; For the first The scaling factor for each sample category satisfies condition, For the first Number of samples in each sample category; Using the aforementioned training subset, the width learning system obtained from the previous iteration is trained to obtain the width learning system after the current iteration. The width learning system trained in this iteration is used to predict the posterior probability of a test sample belonging to each fault category. Based on the posterior probability, the BLS incremental algorithm is used to determine whether the width learning system after this iteration of training needs incremental nodes, so as to obtain the width learning system for the next iteration of training.

[0010] Alternatively, based on the output weight matrix of the multilayer perceptron, the formula can be used... and The output of the width learning system is corrected and mapped to the posterior probability; where, For the input feature data, To output the weight matrix, For the output weight matrix, the first... One output; For the width learning system One output, For the corresponding true value, This represents the number of training sample groups; Let be the posterior probability.

[0011] Optionally, based on the posterior probability, the BLS incremental algorithm is used to determine whether the width learning system after this iteration of training needs incremental nodes to obtain the width learning system for the next iteration of training. Specifically, this includes: if the posterior probability is less than a probability threshold, then adding a set of augmentation nodes to the width learning system after this iteration of training to obtain the width learning system for the next iteration of training; if the posterior probability is greater than or equal to the probability threshold, then determining the width learning system after this iteration of training as the width learning system for the next iteration of training.

[0012] Optionally, based on the multimodal signals of the track circuit in the adjustment state without any faults, the remaining lifespan of each device in the track circuit is determined using a lifespan prediction model. Specifically, this includes: normalizing the multimodal signals of the track circuit in the adjustment state without any faults to obtain normalized multimodal signals; inputting the normalized multimodal signals into the width learning system in the lifespan prediction model to output deep features; inputting the deep features into a bidirectional long short-term memory neural network to output the remaining lifespan of each device in the track circuit; the remaining lifespan is represented by the percentage of remaining lifespan at the current moment.

[0013] Secondly, this application provides a track circuit fault diagnosis and life prediction device, comprising: The signal acquisition module is used to acquire multi-mode signals of the track circuit under different operating states; the different operating states include adjustment state and shunt state; The decomposition module is used to adaptively decompose the multimodal signal and extract features from the modal components obtained by adaptive decomposition to obtain the time-frequency features of the multimodal signal. The fusion module is used to fuse the time-frequency features of the multimodal signals to obtain fused features; The fault diagnosis module is used to detect whether there is a fault in the track circuit and the type of fault when the fault exists, based on the fused features and the fault diagnosis model. The fault diagnosis model is trained based on a width learning system. When training the width learning system, the BLS incremental algorithm is used, and the data used in each iteration of training is corrected by actively selecting the training set. The lifespan prediction module is used to determine the remaining lifespan of each device in the track circuit based on the multimodal signal of the track circuit in the adjustment state if no fault is detected in the track circuit. The lifespan prediction model includes a width learning system and a bidirectional long short-term memory neural network connected in sequence.

[0014] Thirdly, this application provides a computer device, including: a memory, a processor, and a computer program stored in the memory and capable of running on the processor, wherein the processor executes the computer program to implement the track circuit fault diagnosis and life prediction method described in any one of the above.

[0015] Fourthly, this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the track circuit fault diagnosis and life prediction method described in any one of the above descriptions.

[0016] According to the specific embodiments provided in this application, this application has the following technical effects.

[0017] This application provides a method, apparatus, equipment, and medium for fault diagnosis and life prediction of track circuits. Based on a wide learning system, a fault diagnosis model and a life prediction model are established. The fault diagnosis model can actively detect whether there are faults in the track circuit under different working conditions. The life prediction model can assess the aging trend of each device in the track circuit after long-term operation, predict the remaining life of each device in the track circuit, comprehensively perceive the overall health status of the track circuit, and can run through the entire life cycle of the track circuit. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of this application or related technologies, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a schematic diagram of the ZPW-2000A non-insulated track circuit section structure provided in an embodiment of this application.

[0020] Figure 2 This is a flowchart illustrating a method for diagnosing track circuit faults and predicting lifespan, provided in an embodiment of this application.

[0021] Figure 3 This is a schematic diagram of the structure of the width learning system provided in an embodiment of this application.

[0022] Figure 4 This is a schematic diagram illustrating the execution process of the modified integrated width learning system model provided in an embodiment of this application.

[0023] Figure 5 This is a schematic diagram of the training process of BLS-BiLSTM (Bidirectional Long Short-Term Memory) provided in the embodiments of this application.

[0024] Figure 6 This is a schematic diagram of the BLS fault diagnosis confusion matrix provided in an embodiment of this application.

[0025] Figure 7 This is a schematic diagram of the confusion matrix for fault diagnosis of MIBLS (Modified Integrated Broad Learning System) provided in an embodiment of this application.

[0026] Figure 8 This is a schematic diagram comparing the predicted remaining lifespan of the rail compensation capacitor in the track circuit provided in the embodiments of this application.

[0027] Figure 9 A schematic diagram of remaining lifetime prediction based on BLS-BiLSTM provided for embodiments of this application.

[0028] Figure 10 A schematic diagram of BLS-based remaining lifetime prediction provided for an embodiment of this application.

[0029] Figure 11 This is a schematic diagram of LSTM-based remaining lifetime prediction provided for an embodiment of this application.

[0030] Figure 12 A schematic diagram showing the comparison of remaining lifetime predictions for the track circuit tuning unit provided in the embodiments of this application.

[0031] Figure 13 A schematic diagram showing the comparison of the remaining life prediction of the air-core coil of the track circuit provided in the embodiments of this application.

[0032] Figure 14 A schematic diagram showing the comparison of remaining lifetime predictions for track circuit transmitters provided in this application embodiment.

[0033] Figure 15 This is a schematic diagram comparing the prediction of the remaining lifespan of the simulated network disk in the track circuit provided in the embodiments of this application.

[0034] Figure 16 A schematic diagram showing the comparison of remaining lifetime predictions for track circuit receivers provided in this application embodiment.

[0035] Figure 17 A schematic diagram of the intelligent operation and maintenance model framework for track circuits based on a width learning system provided in this application embodiment.

[0036] Figure 18 This is a schematic diagram of the structure of a computer device provided in an embodiment of this application. Detailed Implementation

[0037] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0038] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0039] Reference Figure 1 The structure of the ZPW-2000A non-insulated track circuit section shown illustrates the working state of the track circuit. Figure 1In the SPT cable section, "SPT cable" refers to a railway digital signal cable. GJ stands for track relay, XGJ for small track relay, and XGJG for small track condition relay. XGJG's activation condition is that the XGJ in the adjacent section is activated, meaning the small track behind is idle. XG and XGH represent small track information and small track information feedback conditions, respectively. These two signals work together to help the receiver in this section determine the small track status. The track section generates a frequency-shifted signal via a transmitter, which is injected into the main track circuit and the tuning area of ​​the transmitting end within this track section. The two frequency-shifted signals are transmitted independently and ultimately reach the receiver in this track section and the receiver in the adjacent track section, respectively. The frequency-shifted signal reaching the receiver in the adjacent track section performs a condition check on the tuning area status between this track section and the adjacent track section. The tuning area status check result is transmitted from the adjacent track section receiver to the receiver in this track section. The receiver in this track section receives the operating status, which is divided into adjustment status and branching status. In the "Adjustment" state, which is the idle state where the track circuit is complete and not occupied by any train, the track relay armature is engaged, and the track circuit indicator light turns green, indicating that a locomotive is allowed to enter and occupy the track. In the "Splitter" state, which is the state where a train enters the track circuit and the vehicle's wheelset short-circuits the rails, forming a splitter circuit, the track relay's engagement weakens, its armature drops, and the track circuit indicator light turns red, indicating that the track circuit is occupied.

[0040] In one exemplary embodiment, such as Figure 2 As shown, a method for fault diagnosis and life prediction of track circuits is provided. This method is executed by computer equipment, specifically by a computer device such as a terminal or server alone, or by a terminal and a server together. In the embodiments of this application, it includes the following steps 101 to 105.

[0041] Step 101: Obtain the multi-mode signals of the track circuit under different operating states; different operating states include adjustment state and branching state.

[0042] Step 102: Perform adaptive decomposition on the multimodal signal and extract features from the modal components obtained by adaptive decomposition to obtain the time-frequency features of the multimodal signal.

[0043] Step 103: Fuse the time-frequency features of the multimodal signals to obtain fused features.

[0044] Step 104: Based on the fusion features, use the fault diagnosis model to detect whether there is a fault in the track circuit and the type of fault when there is a fault; the fault diagnosis model is trained based on a width learning system. When training the width learning system, the BLS incremental algorithm is used, and the data used in each iteration of training is corrected by actively selecting the training set.

[0045] Step 105: If no fault is detected in the track circuit, the remaining lifespan of each device in the track circuit is determined using the lifespan prediction model based on the multimodal signal of the track circuit in the adjustment state without fault; the lifespan prediction model includes a width learning system and a bidirectional long short-term memory neural network connected in sequence.

[0046] By implementing steps 101 to 105 above, through feature extraction and feature fusion, and by utilizing and improving a wide learning system, a fast modeling speed and accurate intelligent operation and maintenance diagnosis and prediction can still be achieved even when facing a relatively large dataset.

[0047] In another exemplary embodiment of this application, current sensors and voltage sensors are used at various devices in the monitoring machine and track circuit, respectively. The track circuit schematic acquires voltage signals and frequency-shift keying (FSK) signals from the track circuit. The acquired vibration and current signals are then time-aligned and downsampled (i.e., downsampling) to balance signal quality and computational cost, constructing the original dataset. The multimodal signals include voltage signals and frequency-shift keying signals. The voltage signals include: main rail output voltage, main rail input voltage, secondary rail output voltage, secondary rail input voltage, transmitter output voltage, receiver received voltage, transmitter cable analog network disk voltage, receiver cable analog network disk voltage, transmitter matching transformer voltage, and receiver matching transformer voltage.

[0048] In another exemplary embodiment of this application, the time-domain and frequency-domain features of the voltage signal and FSK signal of the track circuit are extracted to construct a time-frequency domain dataset. While performing noise reduction on the signal data using Symptotic Geometry Mode Decomposition (SGMD), adaptive decomposition is also performed, and feature extraction is performed on the decomposed modal components using minimum envelope entropy. Therefore, step 102 can be replaced by steps 201-202.

[0049] Step 201: Adaptively decompose the multimodal signal using symplectic geometric mode decomposition to obtain multiple modal components.

[0050] The general process of symplectic geometric mode decomposition is as follows: Based on Takens embedding theorem, the original dynamic signal is reconstructed into a phase space matrix. Autocorrelation analysis is performed on the phase space matrix to obtain a covariance symmetric matrix. Then, a Hamiltonian matrix is ​​constructed, followed by a symplectic orthogonal matrix for signal reconstruction. The signal is then decomposed into components with different trends and frequency bands using diagonal averaging.

[0051] The detailed process of symplectic geometric mode decomposition is as follows: 1) Symplectic geometric matrix transformation According to Takens' embedding theorem, a topological equivalent representation of the behavior of an original multidimensional system can be reconstructed from a one-dimensional signal using a delay method. In other words, the original time series can be mapped to a multidimensional state space. Consider a real non-zero time series. , , , Representing time 1, time 2, and time 3 respectively. The signal The numerical value represents the number of samples.

[0052] (1) In the formula: For embedded dimensions, To delay time, Indicates the row number. Choose the appropriate embedding dimension and delay time The corresponding reconstruction matrix can be obtained. Since the embedding dimension affects the final result, power spectral density (PSD) is used to determine the embedding dimension. First, the embedding dimension is calculated... The PSD is obtained; then the maximum peak frequency in the PSD is estimated. If the normalized frequency is less than the threshold 10 -3 ,but ;otherwise .in, Indicates the sampling frequency.

[0053] right Autocorrelation analysis was performed to obtain the covariance symmetric matrix. : (2) Then construct the Hamiltonian matrix. : (3) After constructing the Hamiltonian matrix, Squaring yields ， but This is also a Hamiltonian matrix. Furthermore, a symplectic orthogonal matrix is ​​constructed. : (4) In the formula: It is an orthogonal symplectic matrix, possessing the properties of symplectic matrices. Therefore, during this matrix transformation, the structural form of the Hamiltonian matrix remains intact. It is an upper triangular matrix, and all elements below the main diagonal are 0. It is an intermediate coupling term. Calculate the upper triangular matrix. eigenvalues , These represent the first eigenvalue, the second eigenvalue, and the third eigenvalue, respectively. The eigenvalues, as known from the properties of the Hamiltonian matrix, are: (5) In the formula: Indicates the first Square root of each eigenvalue The distribution is a covariance symmetric matrix The symplectic geometric spectrum; ; =1,2,···, .

[0054] Through Hamiltonian matrix Construction, symplectic geometry theory is an efficient method for solving embedded dimension time series. Therefore, the Hamiltonian matrix is ​​employed. Replace symplectic orthogonal matrix , It is a Xin-You matrix. It can be constructed from real matrices, thus providing convenience for the study of time series.

[0055] Using symplectic geometric transformation Decompose the matrix and construct Then the Hamiltonian matrix is ​​used. Replace symplectic orthogonal matrix ,make , The transformation coefficient matrix, This is the reconstructed trajectory matrix. The construction steps for each matrix are as follows: First, the transformation coefficient matrix is ​​obtained through the eigenvectors of the unitary matrix and the trajectory matrix. : (6) In the formula: Indicates the first The transformation coefficient matrix of eigenvalues, Indicates the first A symplectic orthogonal matrix with eigenvalues, for The transformation yields the single-component components: (7) In the formula, Indicates the first The initial single-component matrix of each eigenvalue. Then the initial single-component trajectory matrix. for: (8) In the formula: , express Z It is 3D matrix.

[0056] because yes A dimensional matrix, therefore, it is necessary to... By reordering and averaging the diagonal, the reconstructed matrix can be obtained. Transform into a new set of lengths equal to The time series of the numerical values ​​are thus obtained. The new length of the group is equal to The sum of the numerical time series is the original time series. For Define the first matrix Line number Column elements are ;in and . express and The smaller value in express and The larger of the values. If ,but , Represents the matrix used to reconstruct the time series; otherwise, , Indicates the first Line number The elements of the column. Therefore, the reconstructed time series is as follows: (9) in, Represents the time series of reconstruction. This represents the summation index variable during reconstruction. After averaging along the diagonal, Convert to time series . It is by A matrix composed of initial eigencomponents.

[0057] 2) Single component reconstruction By constructing the trajectory matrix and performing matrix decomposition, we can obtain... The initial signal consists of several single components, but these components are not independent. They may share the same periodicity, frequency, and other characteristics. Therefore, it is necessary to reconstruct each initial component. This application uses periodic similarity as an evaluation index to reconstruct the components. Due to environmental interference, the signal usually contains a large amount of noise, and its period is not regular. Therefore, an iteration termination condition is set.

[0058] first For one For a 3D matrix, since the main components are distributed in the first column, a single component is... Perform a periodic similarity comparison with the remaining components. Components with high similarity are reconstructed to obtain the first component, SGC1. Components participating in the reconstruction of SGC1 are then... After removing the middle, the remaining matrix can be represented as The residual signal is obtained by summing the residual matrix. The normalized mean square error between the obtained residual term and the original signal is calculated as follows: (10) in, This represents the summation index variable used when calculating the normalized mean squared error. The iteration count is [number of iterations]. The entire decomposition process terminates when the Normalized Mean Squared Error (NMSE) is less than a given threshold; otherwise, the residual matrix is ​​used as the original matrix for iteration, and the above iteration process is repeated until the iteration stopping condition is met, thus obtaining the final decomposition result. ,Right now: (11) In the formula, The number of component sequences obtained.

[0059] Step 202: Use the minimum envelope entropy to extract features from multiple modal components to obtain the time-domain and frequency-domain features of the multimodal signal.

[0060] The following features are extracted from each SGC component: Time domain features: mean, variance, standard deviation. Frequency domain features: Perform FFT (Fast Fourier Transform) on each SGC to obtain its spectrum, calculate the centroid frequency, mean square frequency, frequency variance, and spectral entropy, and extract the amplitude and frequency values ​​of the top 3 dominant frequencies.

[0061] In another exemplary embodiment of this application, after the time-frequency features of the 11 signals of the track circuit are extracted, all features are spliced ​​together to form a global high-dimensional feature vector.

[0062] In another exemplary embodiment of this application, traditional deep neural networks already possess excellent feature extraction capabilities. However, when faced with complex operating conditions such as track circuits, the sheer volume of data and the complexity of the network structure inevitably lead to slow modeling speeds, and even problems such as gradient explosion and getting trapped in local optima. The Breadth Learning System (BLS) possesses the ability to learn and approximate in time-varying big data environments. Compared to the structure of deep neural networks that expands towards "depth," BLS focuses on constructing the structure towards "width."

[0063] like Figure 3 As shown, BLS consists of an input layer, a hidden layer (feature nodes and augmentation nodes concatenated), and an output layer. Its learning process is as follows: For a given training dataset , yes dimensional training samples, yes Corresponding label samples, Represents the number of samples. It is the number of categories in the sample. Indicates input feature data, Indicate the fault type. First, the training samples... Convert to feature nodes With the Feature nodes (mapping feature groups) The mapping relationship is as follows: (12) in: It can be a linear or nonlinear activation function. and These are the weights and biases, which are set randomly.

[0064] Secondly, after obtaining all the feature nodes, Group feature nodes are concatenated as Then its first Enhanced relationships between group nodes It can be represented as: (13) Where: activation function A nonlinear activation function This represents a rectified linear unit function with leakage. Represents variables; and The weights and biases are also randomly set.

[0065] After obtaining all the enhanced nodes, Group feature nodes are concatenated as Then, all feature nodes and enhancement nodes are concatenated and combined to obtain the hidden layer. Its form is: (14) in, Indicates the dimension of the data.

[0066] Finally, the relationship between the hidden layers and the output layer of the BLS model can be expressed as: (15) in: These are the labels output by the model. These are the weights from the hidden layer to the output layer of the BLS model.

[0067] , , , The model is randomly generated and does not require training; network training only requires training. The training problem is usually solved using ridge regression theory. First, the objective function of BLS is constructed as follows: (16) in: It is a regularization parameter. This is used to prevent overfitting; the latter is the least squares error term.

[0068] The fault diagnosis model in this application is trained based on a wide learning system, and improves upon the wide learning system through a modified ensemble concept: Track circuits exhibit a wide variety of fault types, and their classification tasks often face severe class imbalance problems. To avoid the excessive impact of imbalanced datasets on the model's learning performance, this application introduces a modified ensemble concept to improve the traditional BLS model. At the data level, a proactive selection of the training set is used before training to ensure a more balanced data distribution. At the algorithm level, an incremental BLS algorithm is employed to better adapt to complex and imbalanced datasets.

[0069] Actively selecting the training set is mainly based on the value of the samples, which requires the use of posterior probabilities. However, the output of the traditional BLS classification algorithm may not be in the range of [0,1]. Therefore, the output weight matrix of a multilayer perceptron (MLP) is utilized. The output of the improved BLS, namely MBLS (Modified Broad Learning System), can be corrected and mapped to posterior probabilities. express The Each output can be represented as: (17) in It is the first BLS classification model One output, That corresponds to the actual value. Therefore: (18) At this point, BLS only corrected the output format and did not change the characteristics of traditional BLS. In imbalanced classification tasks, performance bias still exists (favoring the output type as the normal type with a large amount of data). Further, the concepts of active selection of training subsets and ensemble are introduced to build the MIBLS model.

[0070] Given training samples ,in It is all that belong to the first A collection of class samples. Definition The number of samples in the sample is Training samples The imbalance ratio (IR) is: (19) A smaller IR value indicates a higher degree of class imbalance in the dataset. Assuming... and ,Right now The largest number of samples were found in the middle. The fewest samples are in the subset. When constructing the training subset in each iteration, retain... All samples from each other Select from ( It is a scale factor that satisfies (Condition) samples are used to construct a training subset. It is obvious that That is, the imbalance of the training subset is much smaller than that of the original training set.

[0071] MBLS uses this training subset to train a weak classification model, and then inputs the test set data into this classification model to output the prediction results. If the accuracy is lower than a set value, the model will fail to perform its prediction. Adding a set of enhancement nodes improves the prediction performance of MBLS, which can be represented as: (20) in and The weights and biases are random. The new hidden layer is... Then the new weight for: (twenty one) in: This represents a definition matrix used to adjust weights to introduce boosting nodes to improve model performance. This represents a definition matrix that ensures the model updates weights appropriately under different conditions, thereby improving prediction performance. .

[0072] The process involves T iterations, and the final class label of the sample is determined by voting on this set of outputs. T is related to the degree of data imbalance and is set as follows: (twenty two) in This indicates rounding up. In other words, the maximum number of iterations a model can perform depends on the imbalance of the dataset; the greater the imbalance, the more iterations are required, and vice versa.

[0073] Figure 4 Part (a) illustrates the iterative process of the proposed MIBLS model, where MBLS i MB i y i Let represent the outputs of the MBLS algorithm, the classification model, and the predictions of the classification model on the test set in the i-th iteration, respectively. T represents the maximum number of iterations. In each iteration, MIBLS selects a certain number of samples from each class in the training set and uses them to train MBLS. i , obtain MB i Finally, use MB i Predict the posterior probability that a test sample belongs to each class, and then determine the MBLS based on the posterior probability. i+1 The algorithm determines whether incremental nodes are needed and then selects a subset of the training set for the next iteration from the training set. The iteration stops when the maximum number of iterations T is reached. Finally, the class label of the test sample is determined by voting on the y values ​​obtained in each iteration. The specific voting process is as follows: Figure 4 As shown in part (b) of the document.

[0074] As can be seen from the above, the process of each iteration of training in the width learning system can be replaced by the following steps 301 to 305.

[0075] Step 301: Divide the samples in the training sample set into categories, and determine the category with the fewest samples as the basic sample category.

[0076] Step 302: Retain all samples from the base sample category and select from each sample category other than the base sample category. These samples together constitute the training subset; among them... The number of samples in the basic sample categories; the basic sample categories are the th [number] in the training sample set. Each sample category; For the first The scaling factor for each sample category satisfies condition, For the first The number of samples in each sample category.

[0077] Step 303: Using the training subset, train the width learning system obtained in the previous iteration to obtain the width learning system after the current iteration.

[0078] Step 304: Use the width learning system trained in this iteration to predict the posterior probability of the test sample belonging to each fault category.

[0079] Based on the output weight matrix of the multilayer perceptron, the output of the width learning system is corrected and mapped to the posterior probability using formulas (19) and (20); where, For the input feature data, To output the weight matrix, For the output weight matrix, the first... One output; For the width learning system One output, For the corresponding true value, This represents the number of training sample groups; Let be the posterior probability.

[0080] Step 305: Based on the posterior probability, use the BLS incremental algorithm to determine whether the width learning system after this iteration of training needs incremental nodes, and obtain the width learning system for the next iteration of training.

[0081] If the posterior probability is less than the probability threshold, a set of augmentation nodes is added to the width learning system after this iteration of training to obtain the width learning system for the next iteration of training; if the posterior probability is greater than or equal to the probability threshold, the width learning system after this iteration of training is determined as the width learning system for the next iteration of training.

[0082] In another exemplary embodiment of this application, BiLSTM is a variant of RNN (Recurrent Neural Network) that combines forward and backward information. It effectively improves upon the gradient vanishing and gradient exploding problems caused by long-term dependency mechanisms in traditional recurrent neural networks when processing long-term sequences, especially the gradient vanishing problem. Since the data from the orbital circuit is sampled at fixed periods over a long period, BiLSTM has excellent processing performance for this data.

[0083] BiLSTM performs forward and backward computations using two LSTM networks to utilize information from the entire sequence. To address the mutual influence between different feature values, an LSTM layer can be added to BiLSTM to capture trend changes caused by the correlation between different feature values.

[0084] Therefore, this example combines it with width learning to construct the lifetime prediction module (i.e., BLS-BiLSTM) in a large model.

[0085] Track circuit signal data from different sections were selected as the training set. Max-min normalization was applied to each data point monitored by the microcomputer, mapping the results to the range of 0 to 1. This normalization was then used as the feature input, and the lifespan percentage of each piece of equipment in the track circuit was used as the model output. The model underwent supervised training, with the input and output labels in the following format: .in: Represents the training set, This represents the N-dimensional input feature vector of a certain section of the track circuit at the current moment; This represents the percentage of remaining lifespan for Class M track circuit equipment at the current moment, with each value ranging from [0,1].

[0086] Will As input to BLS, the input data is initially mapped to feature nodes. Then, the feature nodes are enhanced to generate enhanced nodes, and the nodes are then concatenated and combined to form deep features.

[0087] Deep features are input into a BiLSTM network. Leveraging the unique characteristics of the BiLSTM memory cell structure, the network outputs M percentage lifetime values ​​for each device at the current moment to build a model. The model training process is as follows: Figure 5 .

[0088] That is, in step 105 above, the remaining lifespan of each device in the track circuit is determined using a lifespan prediction model based on the multimodal signal of the track circuit in the adjustment state without any faults. This may include: normalizing the multimodal signal of the track circuit in the adjustment state without any faults to obtain a normalized multimodal signal; inputting the normalized multimodal signal into the width learning system in the lifespan prediction model to output deep features; inputting the deep features into a bidirectional long short-term memory neural network to output the remaining lifespan of each device in the track circuit; the remaining lifespan is represented by the percentage of remaining lifespan at the current moment.

[0089] In another exemplary embodiment of this application, the application may also provide intelligent operation and maintenance suggestions for each device based on fault diagnosis and life prediction results, combined with expert experience.

[0090] Taking multiple sections of a certain railway track circuit as examples, the effectiveness of the method in this application is explained.

[0091] The fault diagnosis and life prediction modules were co-trained using data from section 1 and simulation data. The fault diagnosis module was tested using data from section 2, and the life prediction module was tested using data from section 3. Considering the class imbalance problem in the fault diagnosis task, Weight-F1 was used as the evaluation metric, and its calculation formula is as follows: (twenty three) In the formula, Number of fault types Fault Category The sample proportion Fault Category The F1 score.

[0092] The proposed method is compared with a fault diagnosis model based on BLS architecture and a remaining lifetime prediction model based on CNN (Convolutional Neural Network) + LSTM. Their performance on the test set is shown in Table 1. In Table 1, the BLS-based comparison method can only perform fault diagnosis or prediction tasks independently, the CNN+LSTM-based comparison method can only perform prediction tasks (remaining lifetime prediction), while the proposed method can perform both diagnosis and prediction tasks in tandem, and achieves better diagnosis and lifetime prediction results than the comparison models on the test set.

[0093] Table 1 Diagnostic and Predictive Results

[0094] To further demonstrate the model's performance and versatility, Figure 6 and Figure 7 The confusion matrix between the 11 fault types of track circuits diagnosed by BLS and MIBLS and the actual fault types is presented. Figure 6 In the confusion matrix of the BLS diagnostic results shown, fault categories F2, F4, F5, F7, F8, F9, F10, and F11 all showed misdiagnosis, indicating a significant bias in the BLS diagnostic results. Figure 7 In the confusion matrix of the MIBLS diagnostic results shown, only a very small number of fault samples in F2 and F4 were misdiagnosed as F4 and F6, and one case of F11 was misdiagnosed as normal. This shows that the MIBLS model has a higher accuracy in diagnosing each type of fault.

[0095] Figures 8-16The study demonstrates the ability of BLS, CNN-LSTM, and BLS-BiLSTM models to predict the remaining life of various equipment in the track circuit. The prediction results of BLS-BiLSTM for large batches of data (remaining life of rail compensation capacitors) and other small batches of data (remaining life of other equipment) are relatively consistent with the actual values, indicating that the model can provide relatively accurate prediction results.

[0096] This embodiment enables coordinated monitoring of multiple objects in the track circuit to achieve status monitoring and signal acquisition. It effectively fuses multi-source heterogeneous data using the symplectic geometric mode decomposition method, overcoming the problem of incomplete representation of single-mode feature information. It can process multi-mode data such as vibration and current from different sampling frequencies and operating conditions, establishing an intelligent operation and maintenance model for the track circuit based on wide learning. This model can simultaneously perform status monitoring, fault diagnosis, and lifespan prediction tasks, comprehensively perceiving the overall health status of the equipment and achieving system-level refined health management. Through multi-task collaborative linkage of diagnosis and prediction, the model demonstrates better performance than dedicated task models. In scenarios such as few-sample learning and continuous learning, the model can be quickly adapted and deployed using relatively few computing resources.

[0097] This application improves the accuracy of fault diagnosis and life prediction, and enhances the operation and maintenance efficiency of track circuits.

[0098] Based on the same inventive concept, this application also provides a track circuit fault diagnosis and life prediction device for implementing the track circuit fault diagnosis and life prediction method described above. The solution provided by this device is similar to the solution described in the above method; therefore, the specific limitations in one or more embodiments of the track circuit fault diagnosis and life prediction device provided below can be found in the limitations of the track circuit fault diagnosis and life prediction method described above, and will not be repeated here.

[0099] In one exemplary embodiment, a track circuit fault diagnosis and life prediction device is provided, comprising: The signal acquisition module is used to acquire multi-mode signals of the track circuit under different operating states; the different operating states include adjustment state and shunt state; The decomposition module is used to adaptively decompose the multimodal signal and extract features from the modal components obtained by adaptive decomposition to obtain the time-frequency features of the multimodal signal. The fusion module is used to fuse the time-frequency features of the multimodal signals to obtain fused features; The fault diagnosis module is used to detect whether there is a fault in the track circuit and the type of fault when the fault exists, based on the fused features and the fault diagnosis model. The fault diagnosis model is trained based on a width learning system. When training the width learning system, the BLS incremental algorithm is used, and the data used in each iteration of training is corrected by actively selecting the training set. The lifespan prediction module is used to determine the remaining lifespan of each device in the track circuit based on the multimodal signal of the track circuit in the adjustment state if no fault is detected in the track circuit. The lifespan prediction model includes a width learning system and a bidirectional long short-term memory neural network connected in sequence.

[0100] Reference Figure 17 The track circuit fault diagnosis and life prediction device of this application includes a data layer, a model layer, and an application layer. The data layer acquires multimodal signals from the device layer. Devices in the device layer include wires, sensors, relays, transmitters, matching transformers, main track circuits, digital cables, analog network disks, small rails, and receiver circuits. The multimodal signals in the data layer include AC, DC, and FSK signals. The model layer mainly consists of three parts: data analysis, state analysis, and downstream tasks (fault diagnosis and life prediction). First, the multimodal data is uniformly encoded into a word sequence through SGMD angle domain resampling, amplitude-distribution information separation, and data segmentation. Then, this sequence is input into a general base (an information integration base model based on a width learning system, i.e., a fault diagnosis model and a life prediction model), extracting health and degradation information into specific words. Finally, these specific words are used to perform various downstream health management tasks such as monitoring, diagnosis, and prediction. The application layer displays the integrated results. Through the above process, the proposed intelligent operation and maintenance model aims to possess characteristics such as high versatility, easy expansion, and evolvability.

[0101] In one exemplary embodiment, a computer device is provided, which may be a server or a terminal, and its internal structure diagram may be as follows. Figure 18As shown, the computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operating system and computer programs in the non-volatile storage media to run. The database stores fault diagnosis results and the remaining lifespan of each component in the track circuit. The I / O interfaces are used for exchanging information between the processor and external devices. The communication interface is used for communication with external terminals via a network connection. When the computer program is executed by the processor, it implements a method for track circuit fault diagnosis and lifespan prediction.

[0102] Those skilled in the art will understand that Figure 18 The structures shown are merely block diagrams of some structures related to the present application and do not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than shown in the figures, or combine certain components, or have different component arrangements. In an exemplary embodiment, a computer device is provided, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method embodiments.

[0103] In one exemplary embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0104] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties.

[0105] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM).

[0106] The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0107] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0108] This application uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of this application. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this application. In summary, the content of this specification should not be construed as a limitation of this application.

Claims

1. A method for fault diagnosis and lifespan prediction of track circuits, characterized in that, include: Acquire multi-mode signals of the track circuit under different operating states; different operating states include adjustment state and shunt state; The multimodal signal is adaptively decomposed, and the modal components obtained by adaptive decomposition are feature extracted to obtain the time-frequency features of the multimodal signal. By fusing the time-frequency features of the multimodal signals, fused features are obtained; Based on the fusion features, a fault diagnosis model is used to detect whether there is a fault in the track circuit and the type of fault when a fault exists. The fault diagnosis model is trained based on a wide learning system. When training the wide learning system, the BLS incremental algorithm is used, and the data used in each iteration of training is corrected by actively selecting the training set. If no fault is detected in the track circuit, the remaining lifespan of each device in the track circuit is determined using a lifespan prediction model based on the multimodal signal of the track circuit in the adjustment state without fault; the lifespan prediction model includes a width learning system and a bidirectional long short-term memory neural network connected in sequence.

2. The method for fault diagnosis and lifespan prediction of track circuits according to claim 1, characterized in that, The multimodal signal includes: a voltage signal and a frequency shift keying signal; The voltage signals include: main rail output voltage, main rail input voltage, small rail output voltage, small rail input voltage, transmitter output voltage, receiver received voltage, transmitter cable analog network disk voltage, receiver cable analog network disk voltage, transmitter matching transformer voltage, and receiver matching transformer voltage.

3. The method for fault diagnosis and lifespan prediction of track circuits according to claim 1, characterized in that, The multimodal signal is adaptively decomposed, and the modal components obtained from the adaptive decomposition are feature extracted to obtain the time-frequency features of the multimodal signal, specifically including: The multimodal signal is adaptively decomposed using symplectic geometric mode decomposition to obtain multiple modal components; The time-domain and frequency-domain features of the multimodal signal are obtained by using the minimum envelope entropy to extract features from multiple modal components.

4. The method for fault diagnosis and life prediction of track circuits according to claim 1, characterized in that, The training process of the width learning system in each iteration includes: The training sample set is divided into categories, and the category with the fewest samples is determined as the basic sample category; Retain all samples from the base sample category and select from each sample category other than the base sample category. These samples together constitute the training subset; among them... The number of samples in the basic sample categories; the basic sample categories are the th [number] in the training sample set. Each sample category; For the first The scaling factor for each sample category satisfies condition, For the first Number of samples in each sample category; Using the aforementioned training subset, the width learning system obtained from the previous iteration is trained to obtain the width learning system after the current iteration. The width learning system trained in this iteration is used to predict the posterior probability of a test sample belonging to each fault category. Based on the posterior probability, the BLS incremental algorithm is used to determine whether the width learning system after this iteration of training needs incremental nodes, so as to obtain the width learning system for the next iteration of training.

5. The method for fault diagnosis and lifespan prediction of track circuits according to claim 4, characterized in that, Based on the output weight matrix of the multilayer perceptron, using the formula and The output of the width learning system is corrected and mapped to the posterior probability; where, For the input feature data, To output the weight matrix, For the output weight matrix, the first... One output; For the first width learning system One output, For the corresponding true value, This represents the number of training sample groups; Let be the posterior probability.

6. The method for fault diagnosis and lifespan prediction of track circuits according to claim 4, characterized in that, Based on the posterior probability, the BLS incremental algorithm is used to determine whether the width learning system after this iteration of training needs incremental nodes, resulting in the width learning system for the next iteration of training, specifically including: If the posterior probability is less than the probability threshold, then a set of enhancement nodes is added to the width learning system after this iteration of training to obtain the width learning system for the next iteration of training. If the posterior probability is greater than or equal to the probability threshold, then the width learning system trained in this iteration is determined as the width learning system for the next iteration.

7. The method for fault diagnosis and life prediction of track circuits according to claim 1, characterized in that, Based on the multi-mode signals of the track circuit in an adjusted state without faults, the remaining lifespan of each piece of equipment in the track circuit is determined using a lifespan prediction model, specifically including: The multimode signals of the track circuit without faults in the adjustment state are normalized to obtain normalized multimode signals; Normalized multimodal signals are input into the width learning system in the lifetime prediction model, and deep features are output. The deep features are input into a bidirectional long short-term memory neural network, which outputs the remaining lifetime of each device in the track circuit; the remaining lifetime is represented by the percentage of remaining lifetime at the current moment.

8. A track circuit fault diagnosis and life prediction device, characterized in that, include: The signal acquisition module is used to acquire multi-mode signals of the track circuit under different operating states; the different operating states include adjustment state and shunt state; The decomposition module is used to adaptively decompose the multimodal signal and extract features from the modal components obtained by adaptive decomposition to obtain the time-frequency features of the multimodal signal. The fusion module is used to fuse the time-frequency features of the multimodal signals to obtain fused features; The fault diagnosis module is used to detect whether there is a fault in the track circuit and the type of fault when the fault exists, based on the fused features and the fault diagnosis model. The fault diagnosis model is trained based on a wide learning system. When training the wide learning system, the BLS incremental algorithm is used, and the data used in each iteration of training is corrected by actively selecting the training set. The lifespan prediction module is used to determine the remaining lifespan of each device in the track circuit based on the multimodal signal of the track circuit in the adjustment state if no fault is detected in the track circuit. The lifespan prediction model includes a width learning system and a bidirectional long short-term memory neural network connected in sequence.

9. A computer device, comprising: A memory, a processor, and a computer program stored in the memory and capable of running on the processor, characterized in that the processor executes the computer program to implement the track circuit fault diagnosis and life prediction method according to any one of claims 1-7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the track circuit fault diagnosis and life prediction method as described in any one of claims 1-7.

Images