A method, device and equipment for diagnosing faults in railway switches.

Abstract

This invention relates to the field of railway fault diagnosis, specifically to a method, apparatus, and equipment for diagnosing railway switch faults, comprising the following steps: S1 acquiring the original vibration signal of the switch, and decomposing the original vibration signal of the switch into modal signals through variational mode decomposition; S2 extracting multi-scale entropy features from the decomposed modal signals, wherein the multi-scale entropy features are multi-scale wavelet coherent weighted permutation entropy, multi-scale weighted diversity entropy, and multi-scale Mel spectrum fusion entropy; S3 using the multi-scale entropy features to train a support vector machine classifier, thereby constructing multiple independent diagnostic models and obtaining multiple diagnostic results; S4 integrating the multiple diagnostic results through a hard-voting decision fusion strategy to obtain the fault type of the switch.

Description

Technical Field

[0001] This invention relates to the field of railway fault diagnosis, and specifically to a method, apparatus and equipment for diagnosing railway switch faults. Background Technology

[0002] With the increasing importance of transportation in modern society, railways, as a vital mode of passenger and freight transport, play a crucial role in national economic and social development through their efficiency and safety. Railway signaling systems, as a key component ensuring train operation safety, are particularly important for stability. Among these systems, switches (i.e., turnout control devices), as core equipment, are responsible for controlling the switching of train directions and are widely deployed along railway lines. However, due to their consistently complex and harsh operating environment, they are highly susceptible to malfunction. According to railway statistics, switch failures account for approximately 40% of all railway signaling system failures, and these failures have the potential to cause serious safety accidents and economic losses. Therefore, developing an efficient and reliable method for diagnosing switch failures is of significant practical importance for improving the safety and intelligence level of railway transportation systems.

[0003] Despite the existence of various fault diagnosis methods, fault diagnosis of switches still faces many challenges. On the one hand, traditional methods often rely on expert experience and rule-based reasoning, lacking universality; on the other hand, due to the low overall failure rate of switches, the amount of fault data obtained is limited, making it difficult to support the training of complex models. Therefore, in recent years, researchers have begun to introduce machine learning techniques to achieve more intelligent fault detection. Among these, entropy methods have become a research hotspot due to their advantages such as less reliance on prior knowledge, simple parameter settings, and strong adaptability to small sample data. Compared with methods such as deep learning, entropy methods are easier to deploy in engineering practice and have promising application prospects.

[0004] The concept of entropy originates from information theory and is primarily used to measure the complexity of signals. To more accurately characterize the features of nonlinear signals, scholars have proposed various improved entropy algorithms based on classical information entropy, such as permutation entropy, diversity entropy, and spectral entropy, which have been widely applied in fields such as machinery, medicine, and power. However, these entropy methods are not specifically designed for the operating conditions of switchgear, and their applicability in this scenario still has certain limitations. Specifically, some entropy features lack sufficient discriminative power when identifying different types of faults, and a single entropy feature cannot fully reflect the complexity of the fault signal, exhibiting weak sensitivity to certain anomaly types. Therefore, how to combine multiple entropy feature expressions to improve the accuracy and robustness of switchgear fault diagnosis has become an important problem that urgently needs to be solved. Summary of the Invention

[0005] To address the shortcomings of existing railway switch fault diagnosis methods, such as insufficient discrimination of partial entropy features in identifying different types of faults, the inability of a single entropy feature to fully reflect the complexity of fault signals, and weak sensitivity to certain anomaly types, this invention provides a railway switch fault diagnosis method, comprising the following steps:

[0006] S1 collects the original vibration signal of the switch and decomposes the original vibration signal of the switch into modal signals through variational mode decomposition;

[0007] S2 extracts multi-scale entropy features from the decomposed modal signals. The multi-scale entropy features are multi-scale wavelet coherent weighted permutation entropy, multi-scale weighted diversity entropy, and multi-scale Mel spectrum fusion entropy.

[0008] S3 uses the multi-scale entropy features to train a support vector machine classifier, thereby constructing multiple independent diagnostic models and obtaining multiple diagnostic results;

[0009] S4 integrates the multiple diagnostic results through a hard-voting decision fusion strategy to obtain the fault type of the switch.

[0010] Preferably, in step S1, variational mode decomposition is used to decompose a non-stationary signal into multiple eigenmode functions, specifically as follows:

[0011] The constraints are ,

[0012] in It is the preset number of modes. Represents the time derivative. Represents the Dirac function, For the first One modal function, For the first The center frequency of a mode represents the frequency at which that mode primarily oscillates. Imaginary unit, satisfying 2 = 1, The original vibration signal, It is a time variable;

[0013] By introducing Lagrange multipliers and secondary penalty factor This transforms the constrained optimization problem into an unconstrained optimization problem:

[0014]

[0015] here For inner product operations, the solution is obtained using the alternating direction method of multipliers, for each mode. and its corresponding center frequency All are solved iteratively in the frequency domain using the following formula:

[0016]

[0017]

[0018]

[0019] in Represents the original vibration signal Fourier transform; This represents the frequency domain representation of the k-th mode in the (n+1)-th iteration; This represents the Lagrange multiplier in the frequency domain during the (n+1)th iteration. This is the step size parameter.

[0020] Preferably, step S2, extracting the multi-scale wavelet coherent weighted permutation entropy from the decomposed modal signal, includes the following steps:

[0021] (1) Coarsening: Modal signals Coarsening into multiple scale signals The method is as follows:

[0022] s is the scaling factor, where This represents the value of the i-th sampling point. The length of the original signal;

[0023] (2) Weighted permutation entropy calculation: Construct an m-dimensional embedding vector:

[0024]

[0025] Where m is the embedding dimension, To delay the time, sort the vectors in ascending order:

[0026]

[0027] The signal has m! possible arrangements, and each embedding vector... Corresponding to a permutation pattern Where (r1, r2, ..., rm) is the sorted index sequence of the orbital vectors, defining the permutation pattern corresponding to this embedded vector. ;

[0028] For each permutation pattern We calculate the weighted frequency of occurrence of each subsequence and define the weight of the subsequence as follows:

[0029]

[0030] Among them, the mean for:

[0031]

[0032] Arrangement pattern The weighted probabilities are:

[0033]

[0034] in The weights are normalized to the extreme values, and the denominator ensures that:

[0035]

[0036] Finally, the weighted permutation entropy is calculated using Shannon entropy:

[0037] ;

[0038] (3) Calculate the wavelet coherence between modes: Wavelet coherence is used to quantify the time-frequency correlation between two modes. Let the mode pair be X(s) and Y(s). The calculation steps are as follows:

[0039] (a) Perform Morlet wavelet transform: Select the analytic Morlet wavelet mother function:

[0040]

[0041] in Calculate the continuous wavelet transform coefficients for the center frequency:

[0042]

[0043] Where a and b are the scale and translation factors, respectively;

[0044] (b) Calculate the cross spectrum and the autospectrum:

[0045]

[0046]

[0047]

[0048] (c) Two-dimensional Gaussian smoothing: using parameters , Perform two-dimensional Gaussian convolution:

[0049]

[0050] (d) Define wavelet coherence:

[0051]

[0052] in It is a regularization term to prevent division by zero. The global average wavelet coherence is:

[0053]

[0054] in The number of discrete points in frequency and time;

[0055] (4) Construct a multi-scale wavelet coherent weighted permutation entropy index

[0056] By combining wavelet coherence and weighted permutation entropy, a new entropy feature is constructed:

[0057] .

[0058] Preferably, step S2, extracting the multi-scale weighted diversity entropy from the decomposed modal signals, includes the following steps:

[0059] (1) Coarsening: The original signal Coarsening into multiple scale signals Specifically:

[0060]

[0061] Where s is the scale factor and N is the length of the original signal;

[0062] (2) Phase space reconstruction: After coarsening, the scale signal is reconstructed. Embedded into phase space to reconstruct its potential dynamic characteristics, let the embedding dimension be m, construct the orbital vector:

[0063] The complete orbital matrix is ​​constructed as follows:

[0064]

[0065] L represents the sequence length of the scaled signal after coarsening;

[0066] (3) Calculate the similarity after noise reduction, including the following steps:

[0067] (a) Calculate cosine similarity:

[0068]

[0069] (b) Calculate the variance of the local window: The i-th orbit is:

[0070]

[0071] The (i+1)th orbit is:

[0072] With two adjacent orbits and The merged window is formed as follows:

[0073]

[0074] window The variance is calculated as follows:

[0075]

[0076] The window mean is:

[0077]

[0078] (c) Calculate similarity:

[0079]

[0080] in, Variance after min-max normalization

[0081]

[0082] (4) Construct the state probability distribution:

[0083] Similarity The range of values ​​for is [ [1,1] is divided into ε equal-width intervals, and the number of occurrences in each interval is counted. The quantity is used to obtain the probability of each state. ;

[0084] (5) Calculate the multi-scale weighted diversity entropy value corresponding to the probability calculation of each state, specifically: , This represents the total number of states.

[0085] Preferably, step S2, extracting the multi-scale Mel spectrum fusion entropy from the decomposed modal signal, includes the following steps:

[0086] (1) Time-frequency representation: The frequency domain representation of the signal is obtained by windowing the signal subframes and performing a short-time Fourier transform.

[0087]

[0088] in, Indicates the first Fourier coefficients of each subframe and frequency index k. This represents the window function, where M is the frequency resolution and n is the time index.

[0089] (2) Mel filter bank and power spectrum transformation: Design a triangular Mel filter bank that is dense in the low-frequency region and sparse in the high-frequency region to map the linear power spectrum onto the Mel frequency scale:

[0090]

[0091] in Represents the linear power spectrum. This is the response of the i-th Mel filter at frequency k, and then the Mel power spectrum is converted to decibels:

[0092] ,in It is a tiny positive number used to avoid zero in logarithmic operations;

[0093] (3) Multi-scale block average pooling: Entropy values ​​at different scales are extracted by performing non-overlapping block average pooling on the time-frequency matrix.

[0094]

[0095] Where s is the scale factor, The coordinates of the pooled matrix. For row index, For column indexes;

[0096] (4) Gray-scale mapping: mapping the matrix Converted to a grayscale image matrix via grayscale mapping. :

[0097]

[0098] in, It is a matrix after undergoing min-max normalization, normalized to the range [0,1].

[0099] (5) Entropy calculation: Based on the probability of each gray level (0–255) appearing in the gray matrix, calculate the corresponding multi-scale Mel spectrum fusion entropy value:

[0100]

[0101] in This represents the probability of a pixel with grayscale value i appearing in the matrix.

[0102] Preferably, in step S4, the final diagnostic results are fused using a hard voting strategy, wherein:

[0103] When the judgments of the three models are inconsistent, the majority vote result shall be used as the final decision.

[0104] If a tie occurs in the voting, the model with the highest accuracy in cross-validation will determine the final classification result.

[0105] A fault diagnosis device for railway switches, comprising:

[0106] The variational mode decomposition module is used to acquire the original vibration signal of the switch and decompose the original vibration signal of the switch into modal signals through variational mode decomposition.

[0107] The extraction module is used to extract multi-scale entropy features from the decomposed modal signals. The multi-scale entropy features are multi-scale wavelet coherent weighted permutation entropy, multi-scale weighted diversity entropy, and multi-scale Mel spectrum fusion entropy.

[0108] The training module is used to train a support vector machine classifier using the multi-scale entropy features respectively, thereby constructing multiple independent diagnostic models and obtaining multiple diagnostic results.

[0109] The hard voting module is used to integrate multiple diagnostic results through a hard voting decision fusion strategy to obtain the fault type of the switch.

[0110] A railway switch fault diagnosis device, characterized in that it includes the railway switch fault diagnosis device as described above.

[0111] By adopting the above scheme, the present invention has the following advantages and beneficial effects: The present invention proposes a novel data-driven switch fault diagnosis method. The core idea is to combine variational mode decomposition (VMD) with multiple enhanced multi-scale entropy features, and improve classification robustness and accuracy through a decision fusion mechanism. This method first uses VMD to decompose the original vibration signal and extract multiple intrinsic mode components (IMFs) to reduce signal complexity and enhance feature separability. Subsequently, three types of optimized multi-scale entropy features are extracted from the decomposed signal: Multi-scale wavelet coherence weighted permutation entropy (MWC-WPE): combining wavelet coherence and permutation entropy ideas to improve the ability to identify phase relationships and abrupt changes; Multi-scale weighted diversity entropy (MWDE): based on cosine similarity and weighting mechanism, to enhance the characterization of the overall signal consistency; Multi-scale Mel spectrum fusion entropy (MMFE): converting the vibration signal into a Mel spectrum and then performing entropy calculation to achieve the fusion extraction of time-frequency domain information.

[0112] Each entropy feature is used to train a support vector machine (SVM) classifier. Finally, multiple classification results are integrated through a hard-voting decision fusion strategy to improve the overall diagnostic performance, especially in identifying complex or ambiguous fault types that were previously difficult to distinguish. Attached Figure Description

[0113] Figure 1 This is a basic flowchart of the present invention;

[0114] Figure 2 This is a model framework diagram of the present invention;

[0115] Figure 3 This is a list of the accuracy of different diagnostic models corresponding to different entropy features in this invention. Detailed Implementation

[0116] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to represent selected embodiments of the invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0117] The preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings, so that the advantages and features of the present invention can be more easily understood by those skilled in the art, thereby providing a clearer and more explicit definition of the scope of protection of the present invention.

[0118] Reference manual attached Figure 1-2 This invention proposes a novel data-driven method for fault diagnosis of switchgear. The core idea is to combine variational mode decomposition (VMD) with various enhanced multiscale entropy features, and improve the classification robustness and accuracy through a decision fusion mechanism.

[0119] Specifically, this method first uses VMD to decompose the original vibration signal and extract multiple intrinsic mode components (IMFs) to reduce signal complexity and enhance feature separability. Then, it extracts three types of optimized multi-scale entropy features from the decomposed signal:

[0120] Multi-scale wavelet coherence weighted permutation entropy (MWC-WPE): Combining the concepts of wavelet coherence and permutation entropy, this method enhances the ability to identify phase relationships and abrupt changes.

[0121] Multi-scale weighted diversity entropy (MWDE): Based on cosine similarity and weighting mechanisms, it enhances the characterization of the overall consistency of signals.

[0122] Multi-scale Mel spectrum fusion entropy (MMFE): After converting the vibration signal into a Mel spectrum, entropy is calculated to achieve the fusion and extraction of time-frequency domain information.

[0123] Each entropy feature is used to train a support vector machine (SVM) classifier. Finally, multiple classification results are integrated through a hard-voting decision fusion strategy to improve the overall diagnostic performance, especially in identifying complex or ambiguous fault types that were previously difficult to distinguish.

[0124] The method was validated on real vibration data provided by Xi'an Railway Signal Co., Ltd., with a diagnostic accuracy of up to 99.52%. It showed good stability and generalization ability on multiple test datasets, significantly outperforming traditional methods and showing broad prospects for engineering applications.

[0125] Reference manual attached Figure 1 The basic process of fault diagnosis for railway switches includes the following steps: First, vibration signals are collected, that is, raw vibration signal data are obtained from the operation of the railway switch; second, the collected signals are decomposed into multiple modal components; then, feature extraction is performed on the decomposed modal signals to extract time-frequency domain features such as multi-scale entropy; next, the extracted features are input into a classification model for fault classification and diagnosis; finally, the diagnosis results are output, recording the health status or fault type of the switch, thus completing the entire fault diagnosis process.

[0126] Step 1: Perform mode decomposition on the signal

[0127] We employ Variational Mode Decomposition (VMD) to perform mode decomposition on the original railway switch vibration signal. VMD is an adaptive, non-recursive signal decomposition technique used to decompose non-stationary signals into multiple intrinsic mode functions (IMFs). The core difference between VMD and Empirical Mode Decomposition (EMD) is that VMD reformulates the signal decomposition problem as a constrained variational optimization problem:

[0128]

[0129] The constraints are

[0130]

[0131] in It is the preset number of modes. Represents the time derivative. Represents the Dirac function, For the first One modal function, For the first The center frequency of a mode represents the frequency at which that mode primarily oscillates. Imaginary unit, satisfying 2 = 1, It is a time variable.

[0132] By introducing Lagrange multipliers and secondary penalty factor The constrained optimization problem is transformed into an unconstrained optimization problem:

[0133]

[0134] This problem can be solved using the Alternating Direction Multiplier Method (ADMM). The core idea of ​​ADMM is to alternately update one variable while keeping two variables fixed. Here... For inner product operations, the solution is obtained using the alternating direction method of multipliers, for each mode. and its corresponding center frequency All are solved iteratively in the frequency domain using the following formula:

[0135]

[0136]

[0137]

[0138] in Represents the original signal Fourier transform; This represents the frequency domain representation of the k-th mode in the (n+1)-th iteration; This represents the Lagrange multiplier in the frequency domain during the (n+1)th iteration. This is the step size parameter.

[0139] VMD has good noise resistance and decomposition accuracy, and is an important step in signal preprocessing in this study.

[0140] Step 2: Extracting features from modal signals

[0141] We first propose a novel entropy feature—Multi-scale Wavelet Coherence Weighted Permutation Entropy (MWC-WPE)—to improve the distinguishability between different fault types. In previous studies, permutation entropy (PE) is typically calculated only in the frequency domain. Unlike existing entropy calculation methods, the proposed MWC-WPE introduces wavelet coherence. By calculating the wavelet coherence coefficients between mode pairs after VMD decomposition, the entropy value can simultaneously reflect the time-frequency characteristics of the signal. Even if two faults have the same value under the traditional MWPE, the difference in their wavelet coherence allows the final MWC-WPE to effectively distinguish between the two types of faults. The specific calculation steps are as follows:

[0142] Step 1: Coarsening treatment

[0143] Original signal Coarsening into multiple scale signals The method is as follows:

[0144] s is the scaling factor, where These represent the values ​​of the 1st sampling point, the 2nd sampling point, ..., the cth sampling point. The index of the original signal. The length of the original signal;

[0145] Step 2: Calculate the Weighted Permutation Entropy (WPE)

[0146] Constructing m-dimensional embedding vectors:

[0147]

[0148] Where m is the embedding dimension, To delay the time, sort the vectors in ascending order:

[0149]

[0150] The signal has m! possible arrangements, and each embedding vector... Corresponding to a permutation pattern Where (r1, r2, ..., rm) is the sorted index sequence of the orbital vectors, and...

[0151] Defines the permutation pattern corresponding to this embedding vector. ;

[0152] For each permutation pattern We calculate the weighted frequency of occurrence of each subsequence and define its weight as follows:

[0153]

[0154] Among them, the mean for:

[0155]

[0156] Arrangement pattern The weighted probabilities are:

[0157]

[0158] in The weights are normalized to the extreme values, and the denominator ensures that:

[0159]

[0160] Finally, the weighted permutation entropy is calculated using Shannon entropy:

[0161] ;

[0162] Step 3: Calculate wavelet coherence between modes

[0163] Wavelet coherence is used to quantify the time-frequency correlation between two modes. Let the mode pair be X(s) and Y(s). The calculation steps are as follows:

[0164] 1. Perform Morlet wavelet transform

[0165] Select the Morlet wavelet mother function for analysis:

[0166]

[0167] in The center frequency.

[0168] Calculate the continuous wavelet transform (CWT) coefficients:

[0169]

[0170] Where a and b are the scale and translation factors, respectively.

[0171] 2. Calculate the cross spectrum and the autospectrum:

[0172]

[0173]

[0174]

[0175] 3. Two-dimensional Gaussian smoothing:

[0176] Use parameters , Perform two-dimensional Gaussian convolution:

[0177]

[0178] 4. Define wavelet coherence:

[0179]

[0180] in It is a regular expression to prevent division by zero.

[0181] The global average wavelet coherence is:

[0182]

[0183] in This represents the number of discrete points in frequency and time.

[0184] Step 4: Construct the MWC-WPE index

[0185] By combining wavelet coherence and WPE, a new entropy feature is constructed:

[0186] .

[0187] Secondly, we propose a novel entropy calculation method based on Diversity Entropy (DE)—Multiscale Weighted Diversity Entropy (MWDE). This method introduces variance information in the time domain, rather than relying solely on traditional cosine similarity, thus making fuller use of the amplitude variation characteristics between adjacent orbits. This definition, while retaining the consistency advantages of the original DE, significantly enhances the sensitivity to local amplitude information, thereby improving the ability to distinguish different fault types. The specific calculation steps of MWDE are as follows:

[0188] Step 1: Coarsening treatment

[0189] Original signal Coarsening into multiple scale signals The method is as follows:

[0190]

[0191] Where s is the scaling factor and N is the length of the original signal.

[0192] Step 2: Phase Space Reconstruction

[0193] After coarse-graining, the signal is embedded into the phase space to reconstruct its underlying dynamic characteristics. Let the embedding dimension be m, then the orbital vector is constructed as follows:

[0194] The complete orbital matrix is ​​constructed as follows:

[0195]

[0196] L represents the length of the coarsened signal sequence;

[0197] Step 3: Calculate the improved similarity (enhanced noise resistance)

[0198] To improve robustness to noise, this invention introduces an improved similarity measurement method that combines cosine similarity with a local variance penalty term.

[0199] 1. Calculate cosine similarity:

[0200]

[0201] Calculate the variance of the local window:

[0202] A window is formed by merging two adjacent tracks. For example, the i-th track is:

[0203]

[0204] The (i+1)th orbit is:

[0205]

[0206] With two adjacent orbits and The merged window is formed as follows:

[0207]

[0208] window The variance is calculated as follows:

[0209]

[0210] The window mean is:

[0211]

[0212] 3. Calculate similarity:

[0213]

[0214] in, It is the variance value after min-max normalization, specifically:

[0215] .

[0216] Step 4: Construct the state probability distribution

[0217] Similarity The range of values ​​for is [ [1,1] is divided into ε equal-width intervals, and the number of occurrences in each interval is counted. The quantity is used to obtain the probability of each state. .

[0218] Step 5: Calculate the MWDE value

[0219] Calculate the corresponding multi-scale weighted diversity entropy (MWDE) value based on the probabilities of each state:

[0220] , The total number of states;

[0221] MMFE (Multiscale Mel-frequency Fusion Entropy) is an entropy calculation method based on feature entropy (FE), incorporating the concept of the Mel spectrum to better align with human perception when processing complex signals. The Mel spectrum is a spectral form that maps actual frequencies to a scale consistent with human auditory perception, emphasizing low-frequency details and compressing high-frequency information, and is widely used in speech recognition and audio processing. By introducing the Mel spectrum, MMFE can more accurately characterize the perceptual properties of signals, thereby improving the separability of fault features. In practical fault diagnosis applications, MMFE exhibits superior performance. Its specific calculation steps are as follows:

[0222] Step 1: Time-Frequency Representation

[0223] The frequency domain representation of the signal is obtained by windowing the signal subframes and performing a short-time Fourier transform (STFT).

[0224]

[0225] in, Indicates the first Fourier coefficients of each subframe and frequency index k. Represents the window function. M is the time offset (subframe index), M is the frequency resolution, and n is the time index.

[0226] Step 2: Mel filter bank and power spectrum transformation

[0227] Design a triangular Mel filter bank that is dense in the low-frequency region and sparse in the high-frequency region to map the linear power spectrum to the Mel frequency scale:

[0228]

[0229] in Represents the linear power spectrum. This is the response of the i-th Mel filter at frequency k, and then the Mel power spectrum is converted to decibels:

[0230]

[0231] in It is a tiny positive number used to avoid zero in logarithmic operations.

[0232] Step 3: Multi-scale block average pooling

[0233] Unlike previous entropy calculation methods, this part extracts entropy values ​​at different scales by performing non-overlapping block average pooling on the time-frequency matrix H:

[0234]

[0235] Where s is the scale factor, The coordinates of the pooled matrix. For row index, For column indexes.

[0236] Step 4: Grayscale mapping

[0237] matrix Converted to a grayscale image matrix via grayscale mapping. :

[0238]

[0239] in, It is a matrix after undergoing min-max normalization, normalized to the range [0,1].

[0240] Step 5: Entropy Calculation

[0241] Calculate the corresponding MMFE value based on the probability of each gray level (0–255) appearing in the gray-scale matrix:

[0242]

[0243] in This represents the probability of a pixel with grayscale value i appearing in the matrix.

[0244] Step 3: Model Classification and Fault Diagnosis

[0245] The entropy values ​​of MWC-WPE, MWDE, and MMFE were calculated for the modal signals and used as feature inputs for fault diagnosis. For each feature, a support vector machine (SVM) classifier was trained, thus constructing three independent diagnostic models. The trained classifiers were tested on a validation set to evaluate the overall framework performance. The final diagnostic results were fused using a hard voting strategy: when the judgments of the three models were inconsistent, the majority vote result was used as the final decision; if a tie occurred, the model with the highest accuracy in cross-validation determined the final classification result.

[0246] Reference manual attached Figure 3 To verify the superiority of the proposed diagnostic framework, this invention also selected several methods for comparison, including MWPE-SVM, MDE-SVM, and MFE-SVM, and compared them on real railway switch vibration data. In addition, the Discrete Entropy (DISE) method was introduced for comparison, with its parameter settings kept identical. Another comparison method is the traditional Time-Frequency Domain Feature (TFDF) method, which includes time-domain features (such as mean amplitude, root mean square, standard deviation, skewness, crest factor, impact factor, and margin factor) and frequency-domain features (such as spectral centroid, spectral bandwidth, spectral flatness, spectral roll-off rate, and zero-crossing rate).

[0247] from Figure 3 It can be seen that the fault diagnosis framework proposed in this invention is superior to other methods. The decision fusion strategy proposed in this paper has better performance than the single entropy decision model in previous studies, with MFE performing the worst. Compared with the four methods MWPE-SVM, MDE-SVM, MFE-SVM, and MDISE-SVM, the fault diagnosis accuracy of the method proposed in this invention is improved by 1.21%, 0.86%, 11.02%, and 1.21%, respectively; compared with the TFDF method, the accuracy is improved by 2.3%. Each single entropy decision model performs differently when facing different types of faults, which is consistent with the previous analysis; however, the decision fusion strategy proposed in this invention can effectively integrate the advantages of different entropy features, making its diagnostic accuracy optimal for all fault types compared to single models, especially achieving 100% diagnostic accuracy for Type-B faults, which is significantly better than other methods, fully demonstrating the effectiveness of the three improved entropy feature and decision fusion strategies.

[0248] The above embodiments are only for illustrating the technical concept and features of the present invention, and are intended to enable those skilled in the art to understand the content of the present invention and implement it accordingly. They should not be construed as limiting the scope of protection of the present invention. All equivalent changes or modifications made in accordance with the spirit and essence of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A method for diagnosing faults in railway switches, characterized in that, Includes the following steps: S1 collects the original vibration signal of the switch and decomposes the original vibration signal of the switch into modal signals through variational mode decomposition; S2 extracts multi-scale entropy features from the decomposed modal signals. These multi-scale entropy features include multi-scale wavelet coherent weighted permutation entropy, multi-scale weighted diversity entropy, and multi-scale Mel spectrum fusion entropy. Extracting the multi-scale Mel spectrum fusion entropy from the decomposed modal signals involves the following steps: (1) Time-frequency representation: The frequency domain representation of the signal is obtained by windowing the signal subframes and performing a short-time Fourier transform. in, Indicates the first Subframes, frequency index Fourier coefficients, Represents the window function. It is frequency resolution. For time indexing; (2) Mel filter bank and power spectrum transformation: Design a triangular Mel filter bank that is dense in the low-frequency region and sparse in the high-frequency region to map the linear power spectrum onto the Mel frequency scale: in Represents the linear power spectrum. This is the response of the i-th Mel filter at frequency k, and then the Mel power spectrum is converted to decibels: ,in It is a tiny positive number used to avoid zero in logarithmic operations; (3) Multi-scale block average pooling: Entropy values ​​at different scales are extracted by performing non-overlapping block average pooling on the time-frequency matrix. in, As a scale factor, The coordinates of the pooled matrix. For row index, For column indexes; (4) Gray-scale mapping: mapping the matrix Converted to a grayscale image matrix via grayscale mapping. : in, It is a matrix after min-max normalization, normalized to the range [0,1]. (5) Entropy calculation: Based on the probability of each gray level (0–255) appearing in the gray matrix, calculate the corresponding multi-scale Mel spectrum fusion entropy value: in Indicates grayscale value The probability of a pixel appearing in the matrix; S3 uses the multi-scale entropy features to train a support vector machine classifier, thereby constructing multiple independent diagnostic models and obtaining multiple diagnostic results; S4 integrates the multiple diagnostic results through a hard-voting decision fusion strategy to obtain the fault type of the switch.

2. The railway switch fault diagnosis method according to claim 1, characterized in that, In step S1, variational mode decomposition is used to decompose a non-stationary signal into multiple eigenmode functions, specifically as follows: The constraints are , in It is the preset number of modes. Represents the time derivative. Represents the Dirac function, For the first One modal function, For the first The center frequency of a mode represents the frequency at which that mode primarily oscillates. Imaginary unit, satisfying , The original vibration signal, It is a time variable; By introducing Lagrange multipliers and secondary penalty factor This transforms the constrained optimization problem into an unconstrained optimization problem: here For inner product operations, the solution is obtained using the alternating direction method of multipliers, for each mode. and its corresponding center frequency All are solved iteratively in the frequency domain using the following formula: in Represents the original vibration signal Fourier transform; Indicates the first In the nth iteration Frequency domain representation of each mode; Indicates the first Lagrange multipliers in the frequency domain during the next iteration This is the step size parameter.

3. The railway switch fault diagnosis method according to claim 1, characterized in that, Step S2, which involves extracting the multi-scale wavelet coherent weighted permutation entropy from the decomposed modal signal, includes the following steps: (1) Coarsening: Modal signals Coarsening into multiple scale signals The method is as follows: , is the scaling factor, where Indicates the first Values ​​of each sampling point Indicates the first Each scale signal The length of the original signal; (2) Weighted permutation entropy calculation: Construct an m-dimensional embedding vector: Where m is the embedding dimension, To delay the time, sort the vectors in ascending order: The signal has m! possible arrangements, and each embedding vector... Corresponding to a permutation pattern ,in( , ,…, () is a sorted index sequence of orbital vectors, defining the permutation pattern corresponding to this embedded vector. For each permutation pattern We calculate the weighted frequency of occurrence and define the weight of the sequence as follows: Among them, the mean for: Arrangement pattern The weighted probabilities are: in The weights are normalized to the extreme values, and the denominator ensures that: Finally, the weighted permutation entropy is calculated using Shannon entropy: ; (3) Calculate the wavelet coherence between modes: Wavelet coherence is used to quantify the time-frequency correlation between two modes. Let the mode pair be... and The calculation steps are as follows: (a) Perform Morlet wavelet transform: Select the analytic Morlet wavelet mother function: in Calculate the continuous wavelet transform coefficients for the center frequency: in and These are the scale and translation factors, respectively. (b) Calculate the cross spectrum and the autospectrum: (c) Two-dimensional Gaussian smoothing: using parameters , Perform two-dimensional Gaussian convolution: (d) Define wavelet coherence: in It is a regularization term to prevent division by zero. The global average wavelet coherence is: in The number of discrete points in frequency and time; (4) Construct a multi-scale wavelet coherent weighted permutation entropy index By combining wavelet coherence and weighted permutation entropy, a new entropy feature is constructed: 。 4. The railway switch fault diagnosis method according to claim 1, characterized in that, Step S2, which involves extracting the multi-scale weighted diversity entropy from the decomposed modal signals, includes the following steps: (1) Coarsening: The original signal Coarsening into multiple scale signals Specifically: in, As a scale factor, The length of the original signal; (2) Phase space reconstruction: After coarsening, the scale signal is reconstructed. Embedded into phase space to reconstruct its potential dynamic characteristics, let the embedding dimension be m, construct the orbital vector: ， The complete orbital matrix is ​​constructed as follows: L represents the sequence length of the scaled signal after coarsening; (3) Calculate the similarity after noise reduction, including the following steps: (a) Calculate cosine similarity: (b) Calculate the variance of the local window: The orbits are: No. +1 track is: With two adjacent orbits and The merged window is formed as follows: window The variance is calculated as follows: The window mean is: (c) Calculate similarity: in, Variance after min-max normalization (4) Construct the state probability distribution: Similarity The range of values ​​for is [ 1,1], divide it evenly into Divide the intervals into equal-width intervals and count the number of occurrences in each interval. The quantity is used to obtain the probability of each state. ; (5) Calculate the multi-scale weighted diversity entropy value corresponding to the probability calculation of each state, specifically: , This represents the total number of states.

5. The railway switch fault diagnosis method according to claim 1, characterized in that, In step S4, the final diagnostic results are fused using a hard voting strategy, wherein: When the judgments of the three models are inconsistent, the majority vote result shall be used as the final decision. If a tie occurs in the voting, the model with the highest accuracy in cross-validation will determine the final classification result.

6. A fault diagnosis device for railway switches, characterized in that, Fault diagnosis is performed using the railway switch fault diagnosis method as described in any one of claims 1-5, including: The variational mode decomposition module is used to acquire the original vibration signal of the switch and decompose the original vibration signal of the switch into modal signals through variational mode decomposition. The extraction module is used to extract multi-scale entropy features from the decomposed modal signals. The multi-scale entropy features are multi-scale wavelet coherent weighted permutation entropy, multi-scale weighted diversity entropy, and multi-scale Mel spectrum fusion entropy. The training module is used to train a support vector machine classifier using the multi-scale entropy features respectively, thereby constructing multiple independent diagnostic models and obtaining multiple diagnostic results. The hard voting module is used to integrate multiple diagnostic results through a hard voting decision fusion strategy to obtain the fault type of the switch.

7. A fault diagnosis device for railway switches, characterized in that, Includes the railway switch fault diagnosis device as described in claim 6.

Images