Turnout working condition comprehensive monitoring method and system based on feature extraction

Abstract

The invention discloses a turnout working condition comprehensive monitoring method and system based on feature extraction, relates to the technical field of rail transit, and overcomes the defect that the traditional vibration spectrum analysis is difficult to reflect the change of electrical characteristics. The turnout target monitoring part is monitored in a mode of combining the high-frequency vibration signal and the contact resistance data, monitoring information is enriched, comprehensive understanding of turnout working conditions is improved, the automation level of turnout monitoring is improved, the requirement for manual intervention is reduced, and the working efficiency is improved. Maintenance and management of the turnout target monitoring part are more scientific and systematized, targeted maintenance decisions can be made according to real-time monitoring data, and therefore operation efficiency and safety of equipment are improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of rail transit, and in particular relates to a method and system for comprehensive monitoring of turnout working conditions based on feature extraction. Background Art

[0002] Track turnouts are mainly composed of switch machines, switch cores, point rails, slide plates, and connecting parts. The switch machine includes base rails, point rails, and connecting parts. Its main function is to guide the train to the correct path. The switch core is a key component of the turnout, which is used to fix the turnout so that it remains stable when the train passes. The close fit between the point rail and the base rail is crucial to the safety of the train. When the train approaches the turnout, the point rail of the switch will move to the selected line under the repulsive force and the guidance of the base rail. This process utilizes the wheel flange guidance principle and controls the position and movement of the wheel flange to keep the train running on the predetermined route. The currently commonly used online monitoring technology for track switches is mainly sensor technology. Among them, sensor technology often uses vibration sensors to obtain vibration data of track switches, and analyzes and processes the vibration data to obtain a vibration spectrum. The vibration spectrum is used to reflect the dynamic characteristics of the track switch, such as whether the switch machine and the point rail move smoothly, and whether there is any jamming or wear. However, single parameter monitoring has certain limitations. For example, the vibration spectrum can only reflect the dynamic characteristics of the switch machine and the point rail, and cannot provide information on changes in electrical characteristics, which makes it difficult to detect early faults. Summary of the invention

[0003] In view of the defects in the prior art, the present invention provides a method and system for comprehensive monitoring of turnout working conditions based on feature extraction to solve the above technical problems.

[0004] A method for comprehensive monitoring of turnout working conditions based on feature extraction, the method comprising the following steps: Deploy a lightweight LSTM network to obtain multi-source feature data of the target monitoring part of the turnout, wherein the multi-source feature data includes high-frequency vibration signals and contact resistance data, and align and fuse the multi-source feature data based on the acquisition time; De-noising, data conversion and data set partitioning are performed on the aligned and fused multi-source feature data, thereby calculating the rank correlation coefficient of the multi-source feature data, and constructing a correlation matrix based on the rank correlation coefficient using a graph theory method; A feature fusion algorithm is used to compare the similarity between the correlation matrix and a preset fault template library, and correlation data with a similarity exceeding a preset threshold is output as a warning signal.

[0005] Among them, the fusion of multi-source data can more comprehensively reflect the operating status of the turnout, overcome the limitations of single parameter monitoring, and by combining electrical characteristics with dynamic characteristics, make the identification of early faults more sensitive, making the maintenance and management of the turnout more scientific and systematic.

[0006] Preferably, when calculating the rank correlation coefficient of the multi-source feature data, the following formula is specifically used: , , , ; in, is a time series high frequency vibration signal, For the synchronously collected resistance data, For sliding window The rank of the internal high frequency vibration signal, is the collection time, is the half width of the sliding window, For the corresponding sliding window The rank of the internal resistance data, is the time decay weight, is the attenuation coefficient, is the weighted average rank of the high-frequency vibration signal, is the weighted average rank of the high-frequency vibration signal.

[0007] Here, the rank correlation coefficient is obtained through vibration signals and resistance data for data processing. Compared with the linear relationship processing method used in general data processing, it effectively captures the nonlinear and monotonic relationships between variables, is insensitive to outliers, and can provide reliable correlation evaluation even when there is noise or extreme values ​​in the data.

[0008] Preferably, when aligning and fusing the multi-source feature data based on the acquisition time, the following steps are specifically included: A three-dimensional coordinate system is constructed, in which the three dimensions represent the instantaneous frequency, energy concentration and frequency change rate of the high-frequency vibration signal respectively; Perform cluster analysis on the data coordinate points in the three-dimensional coordinate system to form several signal clusters, and remove data points far away from these signal clusters based on the K-means clustering algorithm; The IEEE 1588 precision clock protocol is used to align the high-frequency vibration signal and contact resistance data.

[0009] By constructing a three-dimensional coordinate system, the instantaneous frequency, energy concentration and frequency change rate of the high-frequency vibration signal can be considered at the same time. When the K-means clustering algorithm is used to perform cluster analysis on the data coordinate points, the different modes and characteristics of the signal can be effectively identified, reducing the impact of noise on the analysis results.

[0010] Preferably, when clustering the coordinate points in the three-dimensional coordinate system to obtain a number of signal clusters, the following steps are specifically adopted: Get the Euclidean distance between any data coordinate point and other data coordinate points; Using the ISODATA clustering algorithm, determining the intra-class distance based on the obtained multiple Euclidean distances; The data coordinate points in the three-dimensional coordinate system are integrated into several signal clusters according to the intra-cluster distance.

[0011] Using Euclidean distance to quantify the similarity between data points can accurately reflect the actual distribution of the data, lay the foundation for the subsequent ISODATA clustering algorithm to adjust the number and shape of clusters, and avoid information loss when the number of clusters is fixed.

[0012] Preferably, the ISODATA clustering algorithm is used, and when determining the intra-class distance based on the obtained multiple Euclidean distances, a metric loss function is used to determine the intra-class distance; The metric loss function is specifically as follows: , ; in, For the The total loss value of the iteration is is the number of clusters in the current iteration, For the The number of high-frequency vibration signals in clusters, For the The first A vector of samples, For the The mean of the sample vectors in the clusters, For the the centroid of the current iteration of the class, is the intra-class dispersion threshold, is the regularization coefficient, For the The number of samples of the class, For the The sample size weight of the class, is the standard deviation of the intra-class distance.

[0013] By using a metric loss function to determine the intra-class distance, the data points in the same cluster can be made closer, while the distances between different clusters can be more obvious, reducing the overlap between clusters and adapting to the needs of different distributions.

[0014] Preferably, when removing data points far away from the signal clusters according to the K-means clustering algorithm, the following steps are specifically included: After normalizing the multi-source feature data, the number of clusters K is determined using the elbow rule, and K cluster centers and cluster labels of each multi-source feature data are obtained according to the K-means clustering algorithm; The intra-class distance determines a screening threshold of data points, and data points exceeding the threshold are eliminated based on the screening threshold.

[0015] First, by normalizing the multi-source feature data, the dimensional differences between different feature data can be eliminated, thereby assisting the K-means clustering algorithm to quickly identify the structure and pattern of the data, quickly eliminate abnormal data points that are too far away from the cluster center, and ensure the stability and consistency of the clustering results.

[0016] Preferably, when performing denoising, data conversion and data set division on the aligned and fused multi-source feature data, the following steps are specifically included: Select Daubechies wavelet to perform wavelet decomposition on the multi-source feature data to obtain detail coefficients at multiple frequencies, perform soft threshold processing on the detail coefficients to remove noise, and then perform wavelet reconstruction on the denoised detail coefficients; The multi-source data features after wavelet reconstruction are processed by a standardized processing method, and the multi-source data features are scaled to the interval [0, 1]; The multi-source data features after interval division are divided into data sets according to the set ratio.

[0017] Soft threshold processing of detail coefficients using Daubechies wavelet can effectively remove noise, retain the main features of the signal, and capture the changing characteristics of the signal at different frequencies. Combined with the subsequent step of scaling the data to the [0, 1] interval, it helps to eliminate the dimensional differences between different detail coefficients.

[0018] Preferably, the lightweight LSTM network adopts a stacked multi-layer LSTM network and a fully connected layer. The first layer of LSTM network is used to obtain the time series features when collecting the multi-source feature data; The second layer of LSTM network obtains further abstract time-dependent features based on the time series features; The fully connected layer is used to convert the abstract time output sequence output by the second layer LSTM network into a fixed-length vector.

[0019] The first layer of LSTM network can effectively capture the dynamic patterns of data changing over time. Based on the extraction results of the first layer of LSTM network, the feature abstraction process of the second layer of LSTM network enables the model to have a deeper understanding of the complex patterns in the data. Through the combination of two layers of LSTM, the model can process more complex time series data. Through layers of abstraction and transformation, the model can effectively extract and retain key information and reduce information loss. Finally, a fully connected layer is used to convert the abstract time series output by the second layer of LSTM network into a fixed-length vector. This fixed-length output format simplifies subsequent processing.

[0020] A comprehensive monitoring system for turnout working conditions based on feature extraction, comprising: Lightweight LSTM network, multi-source feature data acquisition module, data cleaning module, data processing module; The multi-source characteristic data acquisition module is used to collect high-frequency vibration signals and contact resistance data of the target monitoring part of the turnout; The lightweight LSTM network is used to process time series data and capture the temporal dependencies in multi-source feature data; The data cleaning module constructs a correlation matrix based on the high-frequency vibration signal and contact resistance data collected by the multi-source feature data collection module; The data processing module is used to obtain the comparison result between the correlation matrix and the preset fault template library, and output the correlation data whose similarity exceeds the preset value as a warning signal.

[0021] Preferably, the data cleaning module further includes a data alignment and fusion module, a data set partitioning module, a rank correlation coefficient acquisition module, and a matrix construction module; The data alignment and fusion module is used to align and fuse the multi-source feature data based on the acquisition time; The data set division module is used to perform denoising, data conversion and data set division on the aligned and fused multi-source feature data; The rank correlation coefficient acquisition module is used to calculate the rank correlation coefficient based on the divided data set; The matrix construction module is used to construct a correlation matrix based on the rank correlation coefficient using a graph theory method.

[0022] The beneficial effects of the present invention are as follows: this scheme overcomes the fact that traditional vibration spectrum analysis is difficult to reflect changes in electrical characteristics. By combining high-frequency vibration signals with contact resistance data, the target monitoring parts of the turnout are monitored. While enriching the monitoring information, the comprehensive understanding of the turnout working conditions is improved, the automation level of turnout monitoring is improved, and the need for manual intervention is reduced, making the maintenance and management of the turnout target monitoring part more scientific and systematic, and being able to make targeted maintenance decisions based on real-time monitoring data, thereby improving the operating efficiency and safety of the equipment. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0024] Figure 1 A flow chart of a method for comprehensive monitoring of turnout working conditions based on feature extraction provided by the present invention; Figure 2 A schematic structural diagram of a comprehensive monitoring system for turnout operating conditions based on feature extraction provided by the present invention. DETAILED DESCRIPTION

[0025] In the description of the present invention, it should be understood that the terms "center", "longitudinal", "lateral", "length", "width", "thickness", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside", "clockwise", "counterclockwise", "axial", "radial", "circumferential" and the like indicate orientations or positional relationships based on the orientations or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the referred device or element must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be understood as limiting the present invention.

[0026] The disclosure below provides many different embodiments or examples to realize different structures of the present invention. In order to simplify the disclosure of the present invention, the parts and settings of specific examples are described below. Of course, they are only examples, and the purpose is not to limit the present invention.

[0027] The embodiments of the invention are described in detail below with reference to the accompanying drawings.

[0028] like Figure 1 As shown, a comprehensive monitoring method for turnout working conditions based on feature extraction, the method comprises the following steps: Deploy a lightweight LSTM network to obtain multi-source feature data of the target monitoring part of the turnout, wherein the multi-source feature data includes high-frequency vibration signals and contact resistance data, and align and fuse the multi-source feature data based on the acquisition time; De-noising, data conversion and data set partitioning are performed on the aligned and fused multi-source feature data, thereby calculating the rank correlation coefficient of the multi-source feature data, and constructing a correlation matrix based on the rank correlation coefficient using a graph theory method; A feature fusion algorithm is used to compare the similarity between the correlation matrix and a preset fault template library, and correlation data with a similarity exceeding a preset threshold is output as a warning signal.

[0029] This solution overcomes the fact that traditional vibration spectrum analysis is difficult to reflect changes in electrical characteristics. It monitors the target monitoring parts of the turnout by combining high-frequency vibration signals with contact resistance data. While enriching the monitoring information, it also improves the comprehensive understanding of the turnout working conditions, improves the automation level of turnout monitoring, reduces the need for manual intervention, and makes the maintenance and management of the target monitoring parts of the turnout more scientific and systematic. Targeted maintenance decisions can be made based on real-time monitoring data, thereby improving the operating efficiency and safety of the equipment.

[0030] More specifically, when calculating the rank correlation coefficient of the multi-source feature data, the following formula is specifically used: , , , ; in, is a time series high frequency vibration signal, For the synchronously collected resistance data, For sliding window The rank of the internal high frequency vibration signal, is the collection time, is the half width of the sliding window, For the corresponding sliding window The rank of the internal resistance data, is the time decay weight, is the attenuation coefficient, is the weighted average rank of the high-frequency vibration signal, is the weighted average rank of the high-frequency vibration signal.

[0031] By calculating the rank correlation coefficient, the high-frequency vibration signal and the contact resistance data are integrated into a correlation matrix, revealing the subtle changes between the vibration signal and the resistance data, thereby improving the sensitivity of early fault detection. This multi-dimensional data fusion method helps to comprehensively evaluate the health status of the target monitoring parts of the turnout and overcome the limitations of single parameter monitoring. This data-driven decision support system can improve maintenance efficiency, optimize resource allocation, simplify data analysis processes, make the monitoring system more efficient and automated, and reduce the need for manual intervention. Take 0.01.

[0032] More specifically, when aligning and fusing the multi-source feature data based on the acquisition time, the following steps are specifically included: A three-dimensional coordinate system is constructed, in which the three dimensions represent the instantaneous frequency, energy concentration and frequency change rate of the high-frequency vibration signal respectively; Perform cluster analysis on the data coordinate points in the three-dimensional coordinate system to form several signal clusters, and remove data points far away from these signal clusters based on the K-means clustering algorithm; The IEEE 1588 precision clock protocol is used to align the high-frequency vibration signal and contact resistance data.

[0033] The multidimensional feature representation method of constructing a three-dimensional coordinate system provides a more comprehensive signal feature analysis, which can better capture the dynamic changes and complexity of the signal, and surpasses the limitations of traditional single parameter analysis. The use of high-precision time scale alignment such as the IEEE 1588 precision clock protocol can improve the accuracy of data fusion and ensure the reliability of the analysis results.

[0034] More specifically, when clustering the coordinate points in the three-dimensional coordinate system to obtain a number of signal clusters, the following steps are specifically adopted: Get the Euclidean distance between any data coordinate point and other data coordinate points; Using the ISODATA clustering algorithm, determining the intra-class distance based on the obtained multiple Euclidean distances; The data coordinate points in the three-dimensional coordinate system are integrated into several signal clusters according to the intra-cluster distance.

[0035] The ISODATA clustering algorithm is used to make the clustering results adapt to the actual distribution of the data, more in line with the data characteristics, and avoid the information loss caused by the fixed number of clusters. At the same time, the data coordinate points in the three-dimensional coordinate system can be integrated into several signal clusters according to the intra-class distance, which can effectively identify and extract potential patterns in the data. This enhanced data analysis capability helps to deeply understand the data characteristics and provide support for subsequent fault detection and early warning.

[0036] More specifically, when determining the intra-class distance based on the obtained multiple Euclidean distances using the ISODATA clustering algorithm, a metric loss function is used to determine the intra-class distance; The metric loss function is specifically as follows: , ; in, For the The total loss value of the iteration is is the number of clusters in the current iteration, For the The number of high-frequency vibration signals in clusters, For the The first A vector of samples, For the The mean of the sample vectors in the clusters, For the the centroid of the current iteration of the class, is the intra-class dispersion threshold, is the regularization coefficient, For the The number of samples of the class, For the The sample size weight of the class, is the standard deviation of the intra-class distance.

[0037] Using a metric loss function to determine the intra-class distance can dynamically adjust the distance metric according to the characteristics of the data and improve the accuracy of clustering. This adaptability enables the algorithm to perform better when processing complex data such as high-frequency vibration signals and contact resistance data. It can effectively reduce the impact of outliers on the calculation of intra-class distances, improve the overall performance of the clustering model, and make the model perform better in classification and prediction tasks.

[0038] More specifically, when removing data points far away from the signal clusters according to the K-means clustering algorithm, the following steps are also included: After normalizing the multi-source feature data, the number of clusters K is determined using the elbow rule, and K cluster centers and cluster labels of each multi-source feature data are obtained according to the K-means clustering algorithm; The intra-class distance determines a screening threshold of data points, and data points exceeding the threshold are eliminated based on the screening threshold.

[0039] By analyzing the relationship between clustering error and the number of clusters, we can find the appropriate K value to optimize the clustering effect. By obtaining K cluster centers and cluster labels for each multi-source feature data, we can quickly identify the structure and pattern of the data, improve the efficiency of data analysis, and further reduce clustering fluctuations caused by outliers, thereby improving the interpretability of the results.

[0040] More specifically, when performing denoising, data conversion and data set division on the aligned and fused multi-source feature data, the following steps are also specifically included: Select Daubechies wavelet to perform wavelet decomposition on the multi-source feature data to obtain detail coefficients at multiple frequencies, perform soft threshold processing on the detail coefficients to remove noise, and then perform wavelet reconstruction on the denoised detail coefficients; The multi-source data features after wavelet reconstruction are processed by a standardized processing method, and the multi-source data features are scaled to the interval [0, 1]; The multi-source data features after interval division are divided into data sets according to the set ratio.

[0041] The Daubechies wavelet processing method performs well in processing non-stationary signals such as high-frequency vibration signals and contact resistance data, and can greatly improve the quality of data. At the same time, the multi-scale analysis capability of the Daubechies wavelet processing method itself enables a deeper understanding of complex signals and can better identify potential fault modes. The process of wavelet decomposition and reconstruction ensures the integrity and accuracy of the signal, so that subsequent analysis is based on high-quality data, ensuring the stability of model training. This flexibility makes the data set division more in line with actual needs and improves the generalization ability of the model.

[0042] More specifically, the lightweight LSTM network uses a stacked multi-layer LSTM network and a fully connected layer. The first layer of LSTM network is used to obtain the time series features when collecting the multi-source feature data; The second layer of LSTM network obtains further abstract time-dependent features based on the time series features; The fully connected layer is used to convert the abstract time output sequence output by the second layer LSTM network into a fixed-length vector.

[0043] The first-layer LSTM network focuses on obtaining the time series features of multi-source feature data. This time series feature extraction capability enables the model to understand the timing information of the data, and the second-layer LSTM network further obtains abstract time-dependent features based on the time series features extracted by the first layer. This hierarchical feature abstraction process enables the lightweight LSTM network to more deeply understand the complex patterns in non-stationary signal data such as high-frequency vibration signals and contact resistance data, and improves the modeling capabilities of nonlinear relationships and long- and short-term dependencies. Through layers of abstraction and transformation, the final fully connected layer converts the abstract time series output by the second-layer LSTM network into a fixed-length vector, which is convenient for subsequent classification or regression tasks. The model can effectively extract and retain key information and reduce information loss.

[0044] like Figure 2 As shown, a comprehensive monitoring system for turnout working conditions based on feature extraction includes: Lightweight LSTM network, multi-source feature data acquisition module, data cleaning module, data processing module; The multi-source characteristic data acquisition module is used to collect high-frequency vibration signals and contact resistance data of the target monitoring part of the turnout; The lightweight LSTM network is used to process time series data and capture the temporal dependencies in multi-source feature data; The data cleaning module constructs a correlation matrix based on the high-frequency vibration signal and contact resistance data collected by the multi-source feature data collection module; The data processing module is used to obtain the comparison result between the correlation matrix and the preset fault template library, and output the correlation data whose similarity exceeds the preset value as a warning signal.

[0045] Among them, the multi-source data acquisition module uses an acceleration sensor to collect high-frequency vibration signals from the target monitoring part of the turnout, and uses a contact resistance meter to measure the contact resistance of the turnout. At the same time, the multi-source data acquisition module is also equipped with a data acquisition card to convert the analog signals collected by the acceleration sensor into digital signals.

[0046] More specifically, the data cleaning module also includes a data alignment and fusion module, a data set partitioning module, a rank correlation coefficient acquisition module, and a matrix construction module; The data alignment and fusion module is used to align and fuse the multi-source feature data based on the acquisition time; The data set division module is used to perform denoising, data conversion and data set division on the aligned and fused multi-source feature data; The rank correlation coefficient acquisition module is used to calculate the rank correlation coefficient based on the divided data set; The matrix construction module is used to construct a correlation matrix based on the rank correlation coefficient using a graph theory method.

[0047] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein by equivalents. These modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present invention, and they should all be included in the scope of the claims and specification of the present invention.

Claims

1. A comprehensive monitoring method for turnout working conditions based on feature extraction, characterized in that: The steps of the method include: Deploy a lightweight LSTM network to obtain multi-source feature data of the target monitoring part of the turnout, wherein the multi-source feature data includes high-frequency vibration signals and contact resistance data, and align and fuse the multi-source feature data based on the acquisition time; De-noising, data conversion and data set partitioning are performed on the aligned and fused multi-source feature data, thereby calculating the rank correlation coefficient of the multi-source feature data, and constructing a correlation matrix based on the rank correlation coefficient using a graph theory method; A feature fusion algorithm is used to compare the similarity between the correlation matrix and a preset fault template library, and correlation data with a similarity exceeding a preset threshold is output as a warning signal.

2. The method for comprehensive monitoring of turnout working conditions based on feature extraction according to claim 1 is characterized in that: When calculating the rank correlation coefficient of the multi-source feature data, the following formula is specifically used: ， ， ， ； in, is a time series high frequency vibration signal, For the synchronously collected resistance data, For sliding window The rank of the internal high frequency vibration signal, is the collection time, is the half width of the sliding window, For the corresponding sliding window The rank of the internal resistance data, is the time decay weight, is the attenuation coefficient, is the weighted average rank of the high-frequency vibration signal, is the weighted average rank of the high-frequency vibration signal.

3. The method for comprehensive monitoring of turnout working conditions based on feature extraction according to claim 2 is characterized in that: When aligning and fusing the multi-source feature data based on the acquisition time, the following steps are specifically included: A three-dimensional coordinate system is constructed, in which the three dimensions represent the instantaneous frequency, energy concentration and frequency change rate of the high-frequency vibration signal respectively; Perform cluster analysis on the data coordinate points in the three-dimensional coordinate system to form several signal clusters, and remove data points far away from these signal clusters based on the K-means clustering algorithm; The IEEE 1588 precision clock protocol is used to align the high-frequency vibration signal and contact resistance data.

4. The method for comprehensive monitoring of turnout working conditions based on feature extraction according to claim 3 is characterized in that: When clustering the coordinate points in the three-dimensional coordinate system to obtain a number of signal clusters, the following steps are specifically adopted: Get the Euclidean distance between any data coordinate point and other data coordinate points; Using the ISODATA clustering algorithm, determining the intra-class distance based on the obtained multiple Euclidean distances; The data coordinate points in the three-dimensional coordinate system are integrated into several signal clusters according to the intra-cluster distance.

5. The method for comprehensive monitoring of turnout working conditions based on feature extraction according to claim 4 is characterized in that: Using the ISODATA clustering algorithm, when determining the intra-class distance based on the obtained multiple Euclidean distances, a metric loss function is used to determine the intra-class distance; The metric loss function is specifically as follows: , ; in, For the The total loss value of the iteration is, is the number of clusters in the current iteration, For the The number of high-frequency vibration signals in clusters, For the The first A vector of samples, For the The mean of the sample vectors in the clusters, For the the centroid of the current iteration of the class, is the intra-class dispersion threshold, is the regularization coefficient, For the The number of samples of the class, For the The sample size weight of the class, is the standard deviation of the intra-class distance.

6. The method for comprehensive monitoring of turnout working conditions based on feature extraction according to claim 3 is characterized in that: When the data points far away from the signal clusters are removed according to the K-means clustering algorithm, the following steps are specifically included: After normalizing the multi-source feature data, the number of clusters K is determined using the elbow rule, and K cluster centers and cluster labels of each multi-source feature data are obtained according to the K-means clustering algorithm; The intra-class distance determines a screening threshold of data points, and data points exceeding the threshold are eliminated based on the screening threshold.

7. The method for comprehensive monitoring of turnout working conditions based on feature extraction according to claim 1 is characterized in that: When performing denoising, data conversion and data set division on the aligned and fused multi-source feature data, the following steps are specifically included: Select Daubechies wavelet to perform wavelet decomposition on the multi-source feature data to obtain detail coefficients at multiple frequencies, perform soft threshold processing on the detail coefficients to remove noise, and then perform wavelet reconstruction on the denoised detail coefficients; The multi-source data features after wavelet reconstruction are processed by a standardized processing method, and the multi-source data features are scaled to the interval [0, 1]; The multi-source data features after interval division are divided into data sets according to the set ratio.

8. The method for comprehensive monitoring of turnout working conditions based on feature extraction according to claim 1 is characterized in that: The lightweight LSTM network uses a stacked multi-layer LSTM network and a fully connected layer. The first layer of LSTM network is used to obtain the time series features when collecting the multi-source feature data; The second layer of LSTM network obtains further abstract time-dependent features based on the time series features; The fully connected layer is used to convert the abstract time output sequence output by the second layer LSTM network into a fixed-length vector.

9. A comprehensive monitoring system for turnout working conditions based on feature extraction, characterized in that: include: Lightweight LSTM network, multi-source feature data acquisition module, data cleaning module, data processing module; The multi-source characteristic data acquisition module is used to collect high-frequency vibration signals and contact resistance data of the target monitoring part of the turnout; The lightweight LSTM network is used to process time series data and capture the temporal dependencies in multi-source feature data; The data cleaning module constructs a correlation matrix based on the high-frequency vibration signal and contact resistance data collected by the multi-source feature data collection module; The data processing module is used to obtain the comparison result between the correlation matrix and the preset fault template library, and output the correlation data whose similarity exceeds the preset value as a warning signal.

10. The comprehensive monitoring system for turnout working condition based on feature extraction according to claim 9 is characterized in that: The data cleaning module also includes a data alignment and fusion module, a data set partitioning module, a rank correlation coefficient acquisition module, and a matrix construction module; The data alignment and fusion module is used to align and fuse the multi-source feature data based on the acquisition time; The data set division module is used to perform denoising, data conversion and data set division on the aligned and fused multi-source feature data; The rank correlation coefficient acquisition module is used to calculate the rank correlation coefficient based on the divided data set; The matrix construction module is used to construct a correlation matrix based on the rank correlation coefficient using a graph theory method.

Images