A method and system for unsupervised diagnosis of rail transit electromechanical equipment

Through unsupervised diagnostic methods and clustering algorithms, the Pareto principle is used to select features and automatically update models, and the problem of lack of label data in fault diagnosis of rail transit electromechanical equipment is solved, the accuracy and efficiency of diagnosis are improved, and the equipment operation is ensured. effective diagnostic capabilities.

CN118551289BActive Publication Date: 2024-11-05NANJING KANGNI MECHANICAL & ELECTRICAL
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411017785.3
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-07-29
Publication Date
2024-11-05
Estimated Expiration
2044-07-29

AI Technical Summary

Technical Problem

The lack of labeled data in the fault diagnosis of rail transit electromechanical equipment leads to low diagnostic efficiency and poor accuracy, and it is difficult for existing methods to conduct effective fault simulation experiments in actual operation.

Method used

Unsupervised diagnostic method is adopted to collect operational data, extract time domain, frequency domain, and time frequency domain features, use Pareto principle to select key features, combine soft clustering algorithm and GMM model, automatically update the diagnostic model and eliminate outlier samples , realize unsupervised clustering and diagnosis.

Benefits of technology

It improves the accuracy and credibility of fault diagnosis of rail transit electromechanical equipment, reduces feature dimensions, improves model performance, and ensures the effectiveness and efficiency of the diagnostic model in long-term operation.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118551289B_ABST
    Figure CN118551289B_ABST
Patent Text Reader

Abstract

The present invention provides an unsupervised diagnosis method and system for rail transit electromechanical equipment, the diagnosis method comprising: collecting rail transit electromechanical equipment operation data to form an original operation data set; extracting the time domain, frequency domain, and time-frequency domain features of the data set to form an original feature set, scoring each feature of the original feature set, selecting features using the Pareto principle to form a filtered feature set; using the filtered feature set as a training set, cyclically training to obtain a training set model; obtaining the clustering results of the training set model, training to obtain all diagnostic models within the range of cluster numbers, comparing the clustering results of each diagnostic model with the training set model, obtaining the optimal diagnostic model, outputting the clustering results, classifying and marking each cluster cluster sample, and updating the training set for the next round of model training. The present invention can automatically complete the end-to-end unsupervised fault diagnosis process of signal preprocessing, feature selection, abnormal sample elimination, unsupervised model construction, and abnormal state recognition of key components.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of rail transit safety, and in particular to an unsupervised diagnosis method and system for rail transit electromechanical equipment. Background Art

[0002] Various electromechanical equipment in the field of rail transit are widely used as an important part of urban transportation. It is inevitable that rail transit electromechanical equipment will fail during operation. Once a failure occurs, it will not only affect transportation efficiency, but also endanger passenger safety. At present, the fault diagnosis of rail transit electromechanical equipment mainly relies on regular inspections and manual maintenance. However, this troubleshooting method has problems such as low efficiency, high cost and poor real-time performance. It is increasingly important to develop technologies that can monitor the status of rail transit electromechanical equipment in real time and accurately diagnose faults.

[0003] At present, data-driven rail transit electromechanical equipment fault diagnosis methods are a hot topic in research and application. By analyzing the data generated during equipment operation, machine learning algorithms are used to identify and warn potential faults at an early stage. For example, the vibration signals of electromechanical equipment during operation are analyzed to monitor the operating status of key components of the equipment, and the temperature of the equipment during operation is monitored to identify overheating of components. Combined with deep learning algorithms, the accuracy and automation level of fault diagnosis can be further improved.

[0004] The existing method for detecting the operating status data of rail transit electromechanical equipment has realized the detection of the basic operating conditions of the equipment, and effectively improved the emergency rescue capability of the rail transit system in response to sudden failures. However, as equipment that needs to operate for a long time, rail transit electromechanical equipment has high design requirements, high operation and maintenance safety requirements, and high frequency of manual maintenance. The failure rate of rail transit electromechanical equipment is low, and historical data of failures is difficult to obtain. For the above reasons, fault simulation experiments are also difficult to conduct on rail transit equipment that is actually in operation. Therefore, it is very difficult to obtain labeled fault data for rail transit electromechanical equipment, which requires that the fault diagnosis of rail transit electromechanical equipment in actual work needs to be carried out in the absence of labeled data guidance. Summary of the invention

[0005] Purpose of the invention: The first purpose of the present invention is to provide an automatically updated, high-performance, and high-accuracy unsupervised diagnosis method for rail transit electromechanical equipment; the second purpose is to provide an unsupervised diagnosis system for rail transit electromechanical equipment corresponding to the unsupervised diagnosis method for rail transit electromechanical equipment.

[0006] Technical solution: An unsupervised diagnosis method for rail transit electromechanical equipment, comprising the following steps:

[0007] (1) Collect the operating data of rail transit electromechanical equipment to form the original operating data set;

[0008] (2) Extract the time domain, frequency domain, and time-frequency domain features of the original running data set to form an original feature set, score each feature of the original feature set, and select features based on the scoring results using the Pareto principle to form a filtered feature set;

[0009] (3) Provide a soft clustering algorithm model, give a range of cluster numbers, use the filtered feature set as the training set, cyclically train the soft clustering algorithm model until convergence, obtain the initial parameters of the unsupervised clustering algorithm model, repeat this step, and obtain all training set unsupervised clustering algorithm models within the range of cluster numbers;

[0010] (4) Select any one of the unsupervised clustering algorithm models in the training set, obtain the training set clustering result of the model, then merge the test data with the training set data as the input of the soft clustering algorithm model, train all diagnostic models within the clustering number range, and compare the clustering results of each diagnostic model with the unsupervised clustering algorithm model in the training set in turn, give a consistency score, and obtain the diagnostic model with the best consistency score with the unsupervised clustering algorithm model in the training set. If the consistency score corresponding to the diagnostic model meets the set value, output the corresponding clustering result of the test data. If it does not meet the set value, select another unused training set unsupervised clustering algorithm model and repeat this step until a diagnostic model that meets the set value is obtained, and output the test data clustering result of the diagnostic model;

[0011] (5) Based on the clustering results of the test data output by the diagnostic model, classify and mark the samples in each cluster, issue an alarm for the cluster samples that do not meet the set values, and update the training set based on the classification and marking results for the next round of diagnostic model training.

[0012] Furthermore, step (2) includes the following steps:

[0013] (21) Performing integrated empirical mode decomposition on the original operation data set, extracting the mode with the largest information entropy in the original operation data signal, calculating the time domain, frequency domain, and time-frequency domain features of the original operation data set, and forming an original feature set containing multiple dimensions;

[0014] (22) Score each feature in the original feature set according to the redundancy index or graph index;

[0015] (23) According to the Pareto principle, the cumulative scores of the features are calculated in descending order until the cumulative score accounts for more than or equal to 80%. The corresponding features are then taken out to form a filtered feature set.

[0016] Further, the soft clustering algorithm model includes: a density-based soft clustering algorithm, a model-based soft clustering algorithm or a fuzzy-based soft clustering algorithm; the density-based soft clustering algorithm includes a DBSCAN algorithm or an OPTICS algorithm, the model-based soft clustering algorithm includes a GMM algorithm, and the fuzzy-based soft clustering algorithm includes an FCM algorithm or an FMM algorithm;

[0017] Unsupervised clustering algorithm models include: center-based unsupervised clustering algorithm, density-based unsupervised clustering algorithm or connectivity-based unsupervised clustering algorithm; center-based unsupervised clustering algorithms include K-Means, Bisecting K-Means, GMM algorithms, density-based unsupervised clustering algorithms include DBSCAN, HDBSCAN, OPTICS, Mean Shift, DensityPeak Clustering algorithms, connectivity-based unsupervised clustering algorithms include hierarchical clustering, K nearest neighbor, ROCK, SOM, spectral clustering algorithms.

[0018] Furthermore, when the FCM algorithm model is selected as the soft clustering algorithm model in step (3), the following steps are included:

[0019] (31) Select a cluster number c within the cluster number range, randomize c cluster centers, use the potential function method to determine the initial cluster center of the FCM model, and use the sample data of the training set to cyclically train the FCM model until the FCM model converges;

[0020] (32) Use the EM algorithm to iterate parameters and converge to obtain the optimal unsupervised clustering algorithm model under the cluster number c, which is used as the unsupervised clustering algorithm model of the training set;

[0021] (33) Repeat the above steps (31)-(32) to obtain the unsupervised clustering algorithm model for all training sets within the range of cluster numbers.

[0022] Furthermore, in step (4), the training set data samples that are in the same data cluster in the clustering results of the training set unsupervised clustering algorithm model and the diagnostic model are counted, and the training set data samples are marked as stable samples. The consistency score is the proportion of stable samples in the training set data. By setting a consistency threshold, the diagnostic model that exceeds the threshold is obtained as the final diagnostic model; in some preferred schemes, the diagnostic model that meets the set value is: after calculating the consistency score of all training set unsupervised clustering algorithm models, the diagnostic model with the highest consistency score is obtained.

[0023] Furthermore, step (3) further includes the following steps:

[0024] Calculate the Xie-Beni index of each training set unsupervised clustering algorithm model, select the cluster number with the smallest Xie-Beni index as the optimal cluster number, output the corresponding optimal training set unsupervised clustering algorithm model, and sort the remaining training set unsupervised clustering algorithm models in ascending order according to the Xie-Beni index.

[0025] Furthermore, in step (4), the clustering result of the optimal training set unsupervised clustering algorithm model is obtained, and then the test data and the training set data are merged as the input of the soft clustering algorithm model, and all diagnostic models within the clustering number range are trained, and the clustering results of each diagnostic model and the optimal training set unsupervised clustering algorithm model are compared in turn, and a consistency score is given. The diagnostic model with the best consistency score with the optimal training set unsupervised clustering algorithm model is selected. If the consistency score corresponding to the diagnostic model meets the set value, the corresponding clustering result of the test data is output. If it does not meet the set value, this step is repeated for the next training set unsupervised clustering algorithm model until a diagnostic model that meets the set value is obtained, and the clustering result of the test data of the diagnostic model is output.

[0026] Furthermore, step (2) further includes the following steps:

[0027] Calculate the SPE index and T of each sample in the filtered feature set 2 Indicators are obtained by performing mean analysis to obtain comprehensive statistical indicators, and outlier sample data in the training set that exceed the control limit of the comprehensive statistical indicators are eliminated to obtain training set samples.

[0028] Furthermore, in step (5), the classification label includes:

[0029] The samples in the cluster samples that only contain the data to be tested and whose data volume is greater than or equal to the set value are marked as unknown state samples, and a warning is issued; the samples in the cluster samples that only contain the data to be tested and whose data volume is less than the set value are marked as outlier samples; the samples in the cluster samples that contain the data to be tested and the training set data are marked as healthy state samples and marked.

[0030] Furthermore, in step (5), the test data marked as healthy state samples are added to the training set, and the equipment components corresponding to the test data marked as unknown state samples are inspected. If it is determined that the component is in a faulty state, the sample is marked as a faulty sample, and a supervised diagnosis model is constructed. Thereafter, the test data is first diagnosed by the supervised diagnosis model, and then enters the diagnosis model for diagnosis.

[0031] The present invention also provides an unsupervised diagnosis system for rail transit electromechanical equipment, comprising:

[0032] Operation data collection module: used to collect the operation data of rail transit electromechanical equipment to form the original operation data set;

[0033] Feature selection module: used to extract the time domain, frequency domain, and time-frequency domain features of the original running data set to form an original feature set, score each feature of the original feature set, and select features based on the scoring results using the Pareto principle to form a filtered feature set;

[0034] Clustering model parameter module: used to provide a soft clustering algorithm model. Given a range of cluster numbers, the filtered feature set is used as a training set. The soft clustering algorithm model is trained cyclically until convergence, and the initial parameters of the unsupervised clustering algorithm model are obtained. This step is repeated to obtain all training set unsupervised clustering algorithm models within the range of cluster numbers.

[0035] Clustering model training module: used to select any one from the unsupervised clustering algorithm models in the training set, obtain the training set clustering result of the model, and then merge the test data with the training set data as the input of the soft clustering algorithm model, train all diagnostic models within the clustering number range, and compare the clustering results of each diagnostic model with the unsupervised clustering algorithm model of the training set in turn, give a consistency score, and obtain the diagnostic model with the best consistency score with the unsupervised clustering algorithm model of the training set. If the consistency score corresponding to the diagnostic model meets the set value, the corresponding clustering result of the test data is output. If it does not meet the set value, another unused training set unsupervised clustering algorithm model is selected to repeat this module until a diagnostic model that meets the set value is obtained, and the clustering result of the test data of the diagnostic model is output;

[0036] Clustering result processing module: used to classify and mark each cluster sample according to the clustering results of the test data output by the diagnostic model, issue an alarm for the cluster samples that do not meet the set values, and update the training set according to the classification and marking results for the next round of diagnostic model training.

[0037] Furthermore, the feature selection module includes the following submodules:

[0038] Feature extraction module: used to perform integrated empirical mode decomposition on the original operation data set, extract the mode with the largest information entropy in the original operation data signal, calculate the time domain, frequency domain, and time-frequency domain features of the original operation data set, and form an original feature set containing multiple dimensions;

[0039] Feature scoring module: used to score each feature in the original feature set according to the redundancy index or graph index;

[0040] Feature filtering module: It is used to calculate the cumulative scores of features in descending order according to the Pareto principle until the cumulative score accounts for more than or equal to 80%, and then extract the corresponding features to form a filtered feature set.

[0041] Further, the soft clustering algorithm model includes: a density-based soft clustering algorithm, a model-based soft clustering algorithm or a fuzzy-based soft clustering algorithm; the density-based soft clustering algorithm includes a DBSCAN algorithm or an OPTICS algorithm, the model-based soft clustering algorithm includes a GMM algorithm, and the fuzzy-based soft clustering algorithm includes an FCM algorithm or an FMM algorithm;

[0042] Unsupervised clustering algorithm models include: center-based unsupervised clustering algorithm, density-based unsupervised clustering algorithm or connectivity-based unsupervised clustering algorithm; center-based unsupervised clustering algorithms include K-Means, Bisecting K-Means, GMM algorithms, density-based unsupervised clustering algorithms include DBSCAN, HDBSCAN, OPTICS, Mean Shift, DensityPeak Clustering algorithms, connectivity-based unsupervised clustering algorithms include hierarchical clustering, K nearest neighbor, ROCK, SOM, spectral clustering algorithms.

[0043] Furthermore, in the clustering model parameter module, when the FCM model is selected as the soft clustering algorithm model, the following submodules are included:

[0044] FCM model training module: used to select a cluster number c within the cluster number range, randomize c cluster centers, determine the initial cluster center of the FCM model using the potential function method, and use the sample data of the training set to cyclically train the FCM model until the FCM model converges;

[0045] Unsupervised clustering algorithm model training module: used to use the EM algorithm to iterate parameters and converge to obtain the optimal unsupervised clustering algorithm model under the cluster number c as the training set unsupervised clustering algorithm model;

[0046] Cyclic training module: used to repeatedly run the FCM model training module and the unsupervised clustering algorithm model training module to obtain all training set unsupervised clustering algorithm models within the range of cluster numbers.

[0047] Furthermore, in the clustering model training module, the training set data samples that are in the same data cluster in the clustering results of the training set unsupervised clustering algorithm model and the diagnostic model are counted, and the training set data samples are marked as stable samples. The consistency score is the proportion of stable samples in the training set data.

[0048] Furthermore, the clustering model parameter module also includes the following submodules:

[0049] Unsupervised clustering algorithm model sorting module: used to calculate the Xie-Beni index of each training set unsupervised clustering algorithm model, select the cluster number with the smallest Xie-Beni index as the optimal cluster number, output the corresponding optimal training set unsupervised clustering algorithm model, and sort the remaining training set unsupervised clustering algorithm models in ascending order according to the Xie-Beni index.

[0050] Furthermore, in the clustering model parameter module, the clustering result of the optimal training set unsupervised clustering algorithm model is obtained, and then the test data and the training set data are merged as the input of the soft clustering algorithm model, and all diagnostic models within the clustering number range are trained, and the clustering results of each diagnostic model and the optimal training set unsupervised clustering algorithm model are compared in turn, and a consistency score is given. The diagnostic model with the best consistency score with the optimal training set unsupervised clustering algorithm model is selected. If the consistency score corresponding to the diagnostic model meets the set value, the corresponding clustering result of the test data is output. If it does not meet the set value, this module is repeatedly executed for the next-ranked training set unsupervised clustering algorithm model until a diagnostic model that meets the set value is obtained, and the clustering result of the test data of the diagnostic model is output.

[0051] Furthermore, the feature selection module also includes the following submodules:

[0052] Outlier sample removal module: used to calculate the SPE index and T of each sample in the filtered feature set 2 Indicators are obtained by performing mean analysis to obtain comprehensive statistical indicators, and outlier sample data in the training set that exceed the control limit of the comprehensive statistical indicators are eliminated to obtain training set samples.

[0053] Furthermore, in the clustering result processing module, the classification tags include:

[0054] The samples in the cluster samples that only contain the data to be tested and whose data volume is greater than or equal to the set value are marked as unknown state samples, and a warning is issued; the samples in the cluster samples that only contain the data to be tested and whose data volume is less than the set value are marked as outlier samples; the samples in the cluster samples that contain the data to be tested and the training set data are marked as healthy state samples and marked.

[0055] Furthermore, in the clustering result processing module, the test data marked as healthy state samples are added to the training set, and the equipment components corresponding to the test data marked as unknown state samples are inspected. If it is determined that the component is in a faulty state, the sample is marked as a faulty sample, and a supervised diagnosis model is constructed. After that, the test data is first diagnosed through the supervised diagnosis model, and then enters the diagnosis model for diagnosis.

[0056] Beneficial effects: Compared with the prior art, the present invention has the following significant effects:

[0057] 1. Aiming at the problem that the original feature dimension of rail transit electromechanical equipment operation data is too large, resulting in unsatisfactory clustering effect, the present invention uses a feature evaluation algorithm to score the features, selects key features according to the scoring results and the Pareto principle, and constructs a filtered feature set, thereby reducing the feature dimension while retaining the interpretability of the feature space, greatly improving the performance of the unsupervised diagnosis model.

[0058] 2. In view of the difficulty in setting the initial parameters of the unsupervised diagnostic model and the problem of reliance on manual selection, the present invention designs a set of model construction and update processes. First, an auxiliary clustering model is used to obtain the initial parameters of the Gaussian mixture model, the model is selected using historical clustering results, and the training set is continuously updated using the clustering results of the data to be tested, so that the diagnostic model has a lasting and effective diagnostic capability, thereby increasing the credibility and accuracy of the diagnosis.

[0059] 3. In order to solve the problem that there are a large number of abnormal samples in the operating data collected by rail transit electromechanical equipment, which affects the accuracy of model diagnosis, the present invention first improves the signal-to-noise ratio of the original data through signal processing, and then uses the statistical control limit model to propose outlier samples in the training set to achieve the optimization of the data set. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] Figure 1 It is a flow chart of the unsupervised diagnosis method of rail transit electromechanical equipment in Example 1 of the present invention.

[0061] Figure 2 It is a flow chart of feature selection in Example 1 of the present invention.

[0062] Figure 3 This is a flow chart of clustering model training and diagnosis in Example 1 of the present invention.

[0063] Figure 4 It is the original vibration signal of the example sample of embodiment 2 of the present invention.

[0064] Figure 5 It is an example sample filtering signal of embodiment 2 of the present invention.

[0065] Figure 6 This is a heat map of the similarity matrix of the example data of Example 2 of the present invention.

[0066] Figure 7 It is a bar chart of the cumulative score of the information compression index in Example 2 of the present invention.

[0067] Figure 8 It is a bar graph of the cumulative score of the atlas index in Example 2 of the present invention.

[0068] Fig. 9 This is a visualization diagram of the data features of the motor bearing before feature selection in Example 2 of the present invention.

[0069] Fig.10 This is a visualization diagram of the data features of the motor bearing after feature selection in Example 2 of the present invention.

[0070] Fig.11 It is a similarity matrix heat map of the maximum information compression index of the escalator main line data in Example 2 of the present invention.

[0071] Fig.12 This is a statistical distribution diagram of example data in Example 2 of the present invention.

[0072] Fig.13 This is a visualization diagram of the initial cluster centers and clustering results of the example training set data in Example 2 of the present invention.

[0073] Fig.14 This is a visualization diagram of the GMM clustering results after adding the test data in Example 2 of the present invention.

[0074] Fig.15 This is a visualization diagram of the GMM clustering results after the motor bearing data of Example 2 of the present invention is added to the data to be tested. DETAILED DESCRIPTION

[0075] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments.

[0076] Embodiment 1:

[0077] See also Figure 1 As shown, this embodiment provides an unsupervised diagnosis method for rail transit electromechanical equipment, including the following steps.

[0078] See also Figure 2 As shown, the following steps 1-3 are the overall process of feature selection in this embodiment.

[0079] Step 1: Vibration signal collection of key components: Collect the operating data of rail transit electromechanical equipment to form the original operating data set.

[0080] By setting sensors at different locations of rail transit electromechanical equipment, vibration signals of key components are collected to form the original operation data set.

[0081] Rail transit electromechanical equipment includes: escalators, vertical elevators, car doors, platform screen doors, ventilation and air-conditioning equipment, automatic ticket vending and checking systems, etc.

[0082] Step 2: Original feature set construction: Extract the time domain, frequency domain, and time-frequency domain features of the original running data set to form the original feature set.

[0083] The operating environment of rail transit electromechanical equipment is complex, and the original vibration signal contains a lot of noise that is irrelevant to the state of the equipment itself. Therefore, it is necessary to perform integrated empirical mode decomposition (EEDM) on the vibration signal, calculate the information entropy of each mode, and take the modal component with the largest information entropy as the feature signal for subsequent calculations, extract its statistical characteristics in the time domain and frequency domain, and perform wavelet packet decomposition on the feature signal at the same time, calculate the energy proportion at each level as the time-frequency domain feature, and constitute the original feature set.

[0084] Step 3: Feature scoring and feature selection: Score each feature of the original feature set, and select features based on the scoring results using the Pareto principle to form a filtered feature set.

[0085] The original feature set has a high dimensionality and must contain some redundant and irrelevant features, so these features need to be removed to improve the diagnostic performance of the model. In some specific embodiments, the following unsupervised feature selection indicators are selected to score each feature: Redundancy index , Spectrum Indicators .

[0086] 3.1. Redundancy index The calculation is as follows. First, the maximum information compression index is calculated. , which is an index used to describe the similarity between features:

[0087]

[0088] Where: and is a random variable and The variance of yes and The covariance of yes and Pearson correlation coefficient between .

[0089] Calculate the maximum information compression index between all features , forming a similarity matrix A, each element in this matrix Representative features With features The maximum information compression index between .

[0090] Using the k-nearest neighbor algorithm, the features are divided into several similar clusters based on the similarity matrix A.

[0091] First, determine the initial parameter k, k is the size of each cluster, which is used to control the number of features in the cluster during the clustering process. For each feature , find its maximum information compression index The highest k nearest neighbor features, these features and themselves form a cluster.

[0092] According to the importance of the feature to the feature cluster, the feature is scored to obtain the redundancy index , calculated as follows:

[0093]

[0094] Calculate The maximum information compression index between other features in the cluster The maximum information compression index between all features in the cluster The proportion of the sum is the redundancy index. , redundancy index The larger the value, the more The more representative it is in the cluster, the lower its redundancy.

[0095] 3.2. Graph Indicators The calculation is as follows: First, a similarity matrix S is established based on the original feature set. Each element in the matrix

[0096]

[0097] Where: is a parameter used to control the similarity decay speed, and are the values ​​of the i-th and j-th instances on the l-th feature, respectively.

[0098] Construct an undirected graph from the similarity matrix S , where V is the vertex set and E is the edge set; each vertex in the graph Corresponding to a feature in the dataset , the weight of the edge between each pair of vertices That is , based on the graph G, define the adjacency matrix W and degree matrix D, and then calculate the Laplacian matrix of the graph G , the graph index of each feature is calculated based on the Laplace matrix L and the feature ranking function.

[0099] In some specific embodiments, a normalized intercept feature ranking function is used:

[0100]

[0101]

[0102] Where: is a feature vector, consisting of the values ​​of all instances under this feature, is the eigenvector The normalized weighted representation of .

[0103] The smaller the value of this function, the smoother the eigenvector is on the graph, the smaller the variation of eigenvalues ​​between adjacent instances is, and the better the consistency with the graph results is.

[0104] In addition to the above feature ranking functions, the following two feature ranking functions can also be used:

[0105]

[0106] Where: is the eigenvector The cosine similarity with the jth non-trivial eigenvector, is a nontrivial eigenvalue.

[0107] Another feature ranking function:

[0108]

[0109] Where: k is the known number of clusters, is the eigenvector The cosine similarity with the jth non-trivial eigenvector, is a nontrivial eigenvalue.

[0110] The above three feature ranking functions all use the spectral properties of the graph to evaluate the relevance of features, but each has its own focus: The consistency of all eigenvectors with the graph structure is considered; Eliminating the influence of trivial eigenvectors, the features can be evaluated more accurately; It focuses on the top k-1 most discriminative eigenvectors and can be expected to achieve better results when dealing with noisy data.

[0111] 3.3. The Pareto method is based on the widely used Pareto principle, which is used to optimize resource allocation and improve efficiency. In the fault diagnosis process, irrelevant features will significantly reduce the performance of the classifier and the accuracy of pattern recognition. Therefore, it is necessary to find a set of representative important features to reduce the data dimension. Based on the Pareto principle, in the feature selection process, it can be considered that the most influential features (accounting for about 20% in number) are expected to produce most of the classification effects (about 80%).

[0112] According to the scoring results of the above feature selection indicators, the features are ranked in descending order according to the scores, and the cumulative percentages of the feature scores are calculated in sequence until the cumulative features are greater than or equal to 80%. The corresponding features are sufficient to complete most of the classification functions, and these features are combined into a filtering feature set.

[0113] See also Figure 3As shown, the following steps 4-7 are the algorithm flow for constructing the diagnostic model of this embodiment.

[0114] Step 4: Outlier sample removal: Calculate the value of each sample in the filtered feature set. Indicators and The standard is obtained by performing mean analysis to obtain the comprehensive statistical indicator fai, and the outlier sample data in the training set that exceeds the control limit of the comprehensive statistical indicator fai is eliminated to obtain the training set samples.

[0115] The operating environment of rail transit electromechanical equipment is complex, and the vibration signals collected by each component are easily disturbed by the external environment. From the perspective of feature space, there are outlier samples in the data set. In actual work, their number cannot be ignored. Retaining outlier samples will cause the model to judge the abnormal state as normal, reducing the accuracy of algorithm diagnosis. However, obtaining labeled normal state data or manually removing outlier samples is time-consuming and laborious, and it is necessary to design an algorithm to automatically remove outlier samples in the training set.

[0116] 4.1. Perform covariance decomposition on the data matrix and select the number of principal components.

[0117] Computational Dataset The covariance matrix of , perform eigenvalue decomposition on it, and sort it in descending order according to the size of the eigenvalue. The formula is as follows:

[0118]

[0119] Where: for The eigenvalue matrix of , yes The eigenvector matrix of yes The first A column contains all the pivot information. yes The remaining columns contain non-pivotal information, n is a single feature The number of elements contained in , unless otherwise specified, the above definition is used in each sub-step of step 4.

[0120] 4.2. Decompose the data space to obtain the principal component subspace and the residual subspace.

[0121] right The eigenvalue decomposition is performed as follows:

[0122]

[0123] Where: is the principal subspace, is the residual subspace, is the score matrix.

[0124] 4.3. Construct detection indicators and criteria.

[0125] Assumptions It is a dataset A sample in the main element subspace is constructed as a comprehensive index , to measure the change of sample vector in principal component space, the calculation formula is:

[0126]

[0127] Where: is the control limit with confidence level α, and its calculation formula is:

[0128]

[0129] Where: is a distribution with A and nA degrees of freedom and confidence level α.

[0130] The indicator is used to measure the projection transformation of the sample vector in the residual space, and its calculation formula is:

[0131]

[0132] Where: It represents the control limit with confidence level α, and its calculation formula is as follows:

[0133]

[0134] Where: , for The eigenvalues ​​of the covariance matrix of is the threshold of the standard normal distribution at the confidence level α.

[0135] 4.4. Construct comprehensive statistical indicator fai.

[0136] right and Perform mean analysis and construct a comprehensive statistical indicator fai, whose calculation formula is as follows:

[0137]

[0138] Where: represents the control limit with confidence level α, .

[0139] The calculation formula is as follows:

[0140]

[0141] Where: , , .

[0142] 4.5. Control Limits As a basis, when the statistical indicator fai exceeds the control limit, the sample is an outlier sample and is eliminated to obtain the screened training set sample.

[0143] Step 5: Construction of unsupervised clustering algorithm model for training set: Given a range of cluster numbers, use the filtered feature set as the training set, cyclically train the soft clustering algorithm model until convergence, determine the cluster center and weight parameters of the unsupervised clustering algorithm model, repeat this step, and obtain all unsupervised clustering algorithm models for training sets within the range of cluster numbers.

[0144] A soft clustering algorithm is a type of clustering algorithm that can assign a probability distribution to each data point. It allows each sample to belong to multiple clusters at the same time with different probabilities, and obtains the probability distribution by calculating the relationship between each sample and each cluster. In this embodiment, the soft clustering algorithm includes density-based clustering algorithms (DBSCAN, OPTICS, etc.), model-based clustering algorithms (GMM, etc.), and fuzzy-based clustering algorithms (FCM, FMM, etc.). The soft clustering algorithm can provide an initial partitioning result based on the data distribution, thereby providing initial parameters for the unsupervised clustering algorithm; among them, the FCM (Fuzzy Clustering Mean) algorithm, that is, the fuzzy C-means clustering algorithm, is a widely used soft clustering algorithm, and the FCM algorithm is used as an example for explanation below.

[0145] Unsupervised clustering algorithms are used to cluster data without pre-labeled data tags. These algorithms can automatically discover structures and patterns in data without manual labeling or guidance. In this embodiment, the unsupervised clustering algorithms include center-based clustering algorithms, such as K-Means, Bisecting K-Means, GMM, etc. The clustering centers provided by the FCM algorithm can be directly used as initial parameters to assist in model fitting; density-based clustering algorithms, such as DBSCAN, HDBSCAN, OPTICS, Mean Shift, DensityPeak Clustering algorithms; connectivity-based clustering algorithms, such as hierarchical clustering, K nearest neighbor, ROCK, SOM, spectral clustering algorithms, the results of the FCM algorithm can assist in determining algorithm parameters or optimizing the clustering process, for example, providing an initial graph structure for the spectral clustering algorithm.

[0146] Taking into account the characteristics of the daily operation data of the target object (large data volume, unstable health status), the data features have multimodal distribution characteristics, and there are often clusters of different shapes and sizes in the feature space. The model-based clustering method is suitable for processing such data sets with complex distribution characteristics. Therefore, in some specific implementations, the GMM model is preferred as the unsupervised clustering algorithm model.

[0147] Rail transit electromechanical equipment has a long scrap age, so its monitoring system needs to always have effective diagnostic capabilities throughout its long operating cycle. As the remaining life of the equipment decreases, the status of each component in the equipment will inevitably deteriorate within an acceptable range of maintenance, which means that the diagnostic model needs to be updated in a timely manner when the component status changes.

[0148] The GMM model is selected for illustration below. The classic method for estimating GMM model parameters is the EM algorithm, but the algorithm is sensitive to the initial values ​​of the parameters and cannot guarantee convergence to the global optimum. Therefore, the present invention uses the potential function method to determine the initial clustering center of the FCM algorithm, uses the training set sample data to train the FCM algorithm model, and determines the GMM model parameters according to the training results.

[0149] 5.1. According to empirical rules, the range of cluster number selection is , n is the number of samples, take a cluster number c within this range, and randomly initialize c cluster centers: .

[0150] 5.2. For each data point , calculate it to each cluster center Distance , using distance to calculate membership , to represent the data points The probability of clustering to j, The calculation formula is:

[0151]

[0152] Where m is a real number greater than 1, which is used to control the fuzziness. The larger the value of m, the more fuzzy the membership.

[0153] 5.3. Based on membership The calculation results of , construct the membership matrix U, and update each cluster center according to the membership matrix U , The calculation formula is:

[0154]

[0155] Where: is the degree of membership, is the data point, m is the membership degree The real number given in the calculation.

[0156] 5.4. Repeat steps 5.2 and 5.3 until the algorithm converges.

[0157] 5.5. Use the EM algorithm to iterate parameters, initialize the cluster centers and memberships obtained from steps 4.2 to 4.4, and converge to obtain the optimal GMM model when the number of clusters is c, which is the training set GMM model when the number of clusters is c.

[0158] 5.6. Repeat steps 5.1 to 5.5 to obtain the training set GMM model under all cluster numbers.

[0159] Step 6: Output the diagnostic model: Select any one from the GMM models in the training set, obtain the clustering result of the GMM model M of the training set, merge the test data with the training set data as the input of the FCM algorithm model, train all GMM diagnostic models within the clustering number range, and compare the clustering results of each GMM diagnostic model model with the GMM model M of the training set in turn, and give a consistency score; select the GMM diagnostic model m with the best consistency score with the GMM model M of the training set. If the consistency score corresponding to the GMM diagnostic model m meets the set value, output the clustering result of the test data corresponding to the GMM diagnostic model m. If it does not meet the set value, randomly select the next one from the remaining unused GMM models of the training set and repeat this step until a GMM diagnostic model that meets the set value is obtained, and the clustering result of the test data of the model is output.

[0160] In some specific embodiments, the training set data samples that are in the same data cluster in the clustering results of the training set GMM model M and the GMM diagnostic model m are counted, and the training set data samples are marked as stable samples, and the consistency score is the proportion K of stable samples in the training set data.

[0161] The closer the k value is to 1, the higher the robustness of the model. For GMM diagnostic model m, the addition of the test data may supplement the equipment status information not included in the training set data. However, under the premise that the clustering result of the training set GMM model M for the training set data is reasonable, the clustering result of the training set data by GMM diagnostic model m should be as consistent as possible with that of the training set GMM model M. Therefore, the GMM diagnostic model with the largest K value is selected as the final GMM diagnostic model. .

[0162] Step 7: Output diagnostic results: According to the clustering results output by the GMM diagnostic model, classify and mark each cluster sample, issue an alarm for cluster samples that do not meet the set value, and update the training set according to the classification and marking results for the next round of GMM diagnostic model training.

[0163] Examining the final GMM diagnostic model For the clustering results of the training set data and the data to be tested, the samples in the clustering cluster samples that only contain the data to be tested and whose data volume is greater than or equal to the set value are marked as unknown state samples, and a warning is issued; the samples in the clustering cluster samples that only contain the data to be tested and whose data volume is less than the set value are marked as outlier samples; the samples in the clustering cluster samples that contain the data to be tested and the training set data are marked as healthy state samples and marked.

[0164] The test data marked as healthy state samples are added to the training set, and the equipment components corresponding to the test data marked as unknown state samples are inspected. If the component is determined to be in a faulty state, the sample is marked as a faulty sample, and a supervised diagnosis model is constructed. When performing subsequent diagnosis, the test data is first diagnosed through the supervised diagnosis model, and then enters the GMM diagnosis model for diagnosis, thereby realizing the cyclic training and updating of the model.

[0165] In some specific embodiments, the following preferred steps may be added to step 5.

[0166] 5.7. Use Xie-Beni index to evaluate the clustering effect of the model.

[0167] Calculate the Xie-Beni index of each training set GMM model, select the cluster number with the smallest Xie-Beni index as the optimal cluster number, output the corresponding optimal training set GMM model, and sort the remaining training set GMM models in ascending order according to the Xie-Beni index. The calculation formula of the Xie-Beni index is as follows:

[0168]

[0169] Where: is the sample set in the jth cluster, is the center of the jth cluster, Is a cluster of sample points.

[0170] Corresponding to the above step 5.7, the following improvement is made in step 6: obtain the clustering result of the optimal training set GMM model, use the test data and the training set data as the input of the FCM algorithm model, train all GMM diagnostic models within the clustering number range, and compare the clustering results of each GMM diagnostic model with the optimal training set GMM model to give a consistency score; select the GMM diagnostic model with the best consistency score with the optimal training set GMM model. If the consistency score corresponding to the GMM diagnostic model meets the set value, output the clustering result of the test data corresponding to the GMM diagnostic model. If it does not meet the set value, repeat this step for the next training set GMM model until a GMM diagnostic model that meets the set value is obtained, and output the clustering result of the test data of the model.

[0171] Since the Xie-Beni index is first used to evaluate the clustering effect of the model, a GMM diagnostic model that meets the consistency score can be obtained more quickly, thereby optimizing the efficiency of this step.

[0172] The present solution is described below in conjunction with a specific embodiment.

[0173] Embodiment 2:

[0174] In this embodiment, an escalator is selected as an example of rail transit electromechanical equipment to illustrate the implementation effect of the unsupervised diagnosis method for rail transit electromechanical equipment.

[0175] Step 1: Install the accelerometer sensor at key parts of the escalator, including the main frame, gear box, motor, main drive shaft, truss, downstairs shaft, etc., and collect the vibration signals of the key parts of the escalator, such as Figure 4 As shown, the original vibration signal is obtained.

[0176] Step 2: Perform integrated empirical mode decomposition on the original vibration signal, calculate the information entropy of each mode, and take the modal component with the largest information entropy as the characteristic signal for subsequent calculations. Figure 5 As shown in the figure, the obtained filtered signal extracts its statistical characteristics in time domain and frequency domain, and at the same time, the characteristic signal is decomposed by wavelet packet, and the energy proportion at each level is calculated as the time-frequency domain feature. The time domain, frequency domain, and time-frequency domain features together constitute the original feature set.

[0177] The time domain and frequency domain characteristics and their calculation methods are shown in Table 1 and Table 2 below:

[0178] Table 1

[0179]

[0180] Table 2

[0181]

[0182] In the table: v is the timing signal, Its frequency domain representation is: is the signal power spectrum.

[0183] Wavelet packet energy Energy ratio characteristics of wavelet packets The calculation is as follows:

[0184]

[0185]

[0186]

[0187] Where: is the i-th wavelet transform coefficient of the signal in the k-th wavelet packet.

[0188] In this embodiment, the signal is decomposed by three-layer wavelet packets, and the features of each domain are combined to finally obtain a 46-dimensional original feature set.

[0189] Step 3: Score each feature of the original feature set, that is, use the method described in step 3 of Example 1 to calculate the redundancy index , Spectrum Indicators , see Figure 6 As shown, some example data graph indicators in this embodiment are Similarity matrix heat map of .

[0190] According to the above two scoring results, the Pareto principle is used to select features respectively, and the experimental data of the host test bench is used as an example to illustrate.

[0191] See also Figure 7 and Figure 8 As shown, they are redundancy index , Spectrum Indicators The cumulative score of the original 46-dimensional features is calculated using the redundancy index of the maximum information compression index. In the evaluation, the 11 features with better performance accounted for 80% of the score, while in the graph index In the evaluation, the 9 features with better performance accounted for 80% of the score, which is in line with the Pareto principle.

[0192] See also Fig. 9 and Fig.10 As shown in the figure, the fault data and normal data feature visualization results of the motor bearing before and after feature selection. It can be seen that through the optimization extraction of features, the spacing between samples in different data clusters becomes smaller, and the distance between clusters becomes larger. At the same time, outlier samples still maintain their characteristics in the sensitive feature space.

[0193] Test the above feature selection algorithm on the escalator mainline data. Fig.11 As shown in the figure, it is a similarity matrix heat map obtained in the calculation of the maximum information compression index. The darker the color in the figure, the higher the similarity. It can be seen that in the original feature set, there is a redundant relationship between some features. The scores of each feature under the two evaluation indicators are shown in Table 3 below. The cumulative score results also conform to the Pareto principle.

[0194] Table 3

[0195]

[0196] Step 4: According to the method described in step 4 of embodiment 1, calculate the Indicators and Indicators are obtained by performing mean analysis to obtain the comprehensive statistical indicator fai, and outlier sample data in the training set that exceeds the control limit of the comprehensive statistical indicator fai are eliminated to obtain the training set samples.

[0197] See also Fig.12 As shown in the figure, the control limits of the example data and the statistics of each sample in the training set. In this example, 220 samples that exceed the control limits are eliminated to avoid affecting the effect of subsequent clustering.

[0198] The example data used in this embodiment is the vibration signal data of the key components (motor) of the escalator in a subway station. The monitoring system collects 2.56 seconds of vibration signals every hour, and the sampling frequency of the acceleration sensor used is 51.2kHz. The training set data was collected from March to August 2023, with a total of 2690 samples; the test set data was collected from August to October of the same year, with a total of 933 samples.

[0199] Step 5: According to the method described in step 5 of Example 1, a range of cluster numbers is given, the filtered feature set is used as the training set, the FCM model is cyclically trained until convergence, the cluster center and weight parameters of the GMM model are determined, and this step is repeated to obtain all training set GMM models within the range of cluster numbers, and the clustering effect of the training set GMM model is scored using the Xie-Beni index, the cluster number with the smallest Xie-Beni index is selected as the optimal cluster number, the corresponding optimal training set GMM model is output, and the remaining training set GMM models are sorted in ascending order according to the Xie-Beni index.

[0200] See also Fig.13As shown in the figure, the optimal clustering results and cluster centers of the example data in this embodiment. For the example data, when the number of clusters is 3, the Xie-Beni index of the corresponding training set GMM model is the smallest. It can be seen that the initial parameters provided by the FCM model effectively guide the fitting of the GMM model.

[0201] Step 6: First, obtain the clustering result of the optimal training set GMM model, use the test data and the training set data as the input of the FCM algorithm model, train all GMM diagnostic models within the clustering number range, and compare the clustering results of each GMM diagnostic model with the optimal training set GMM model in turn, and give a consistency score; select the GMM diagnostic model with the best consistency score with the optimal training set GMM model. If the consistency score corresponding to the GMM diagnostic model meets the set value, output the clustering result of the test data. If it does not meet the set value, repeat this step for the next training set GMM model until a GMM diagnostic model that meets the set value is obtained, and output the clustering result of the test data of the model.

[0202] The training set data samples that are in the same data cluster in the clustering results of the training set GMM model M and the GMM diagnostic model m are counted and marked as stable samples. The consistency score is the proportion K of stable samples in the training set data.

[0203] Given a hyperparameter , if the K value of a model is less than the hyperparameter , then use the next model M for training until a value greater than or equal to the hyperparameter Model M is used as the final GMM diagnostic model Hyperparameters The selection of the value is mainly determined by the distribution characteristics of the data set. In this embodiment, the hyperparameter value selected based on experimental experience is 0.92.

[0204] Step 7: Examine the final GMM diagnostic model For the clustering results of the training set data and the data to be tested, the samples in the clustering cluster samples that only contain the data to be tested and whose data volume is greater than or equal to the set value are marked as unknown state samples, and a warning is issued; the samples in the clustering cluster samples that only contain the data to be tested and whose data volume is less than the set value are marked as outlier samples; the samples in the clustering cluster samples that contain the data to be tested and the training set data are marked as healthy state samples and marked.

[0205] At the same time, the test data marked as healthy state samples are added to the training set, and the equipment components corresponding to the test data marked as unknown state samples are inspected. If the component is determined to be in a faulty state, the sample is marked as a faulty sample, and a supervised diagnosis model is constructed. When performing subsequent diagnosis, the test data is first diagnosed through the supervised diagnosis model, and then enters the GMM diagnosis model for diagnosis, thereby realizing the cyclic training and updating of the model.

[0206] In this embodiment, taking the escalator main line data as an example, after adding the test data, the clustering results consistency of the optimal training set GMM model M and the GMM diagnosis model m under different cluster numbers are compared. Fig.14 As shown, the final output model For the clustering results of the test data, the right picture shows the data source label, and the left picture shows the clustering result label (different numbers represent different data clusters in the clustering results). In the data, the training set data are distributed in four data clusters, and the test data are divided into clusters 0, 1, and 3. Since there are a large number of normal state samples of the training set distributed in the three clusters, the test data are all diagnosed as normal state samples. Through the inspection of the components, it can be known that the escalator motor did not fail during the period from August to October, so the diagnostic accuracy is 100%.

[0207] In order to verify the effectiveness of the diagnostic method, the algorithm was tested using escalator host test bench fault simulation test data. The escalator host test bench consists of a three-phase asynchronous motor and a gearbox. The vibration signal has four channels: base, motor, gearbox x-axis, and gearbox y-axis. The sampling frequency of the acceleration sensor is 51.2kHz, and the acquisition time for each sample is 2.56 seconds. There are 49 normal state samples and 58 motor bearing inner ring wear fault samples. To simulate the actual engineering situation, 70% of the normal state data is used as the training set data, and the remaining 30% of the normal state data and all fault state data are used as test data for testing. The final diagnostic results of the model are as follows: Fig.15 As shown in Figure 1, cluster 1 of the test data does not contain any normal state samples in the training set, so the algorithm determines that it is unknown state data and issues an alarm. Cluster 0 contains a large number of normal state samples in the training set, so the samples in cluster 0 are diagnosed as normal state data and added to the training set data for the next diagnosis. The algorithm test results are shown in Table 4 below.

[0208] Table 4

[0209] Data to be tested Normal state Motor bearing inner ring wear status Sample size 15 58 Diagnostic accuracy 100% 100%

[0210] It can be seen from Table 4 that the unsupervised diagnosis method for rail transit electromechanical equipment provided by this embodiment has a high diagnostic accuracy in the application of escalator fault diagnosis, which repeatedly proves the implementation effect of this solution. At the same time, the diagnosis result can be used as labeled fault data to construct a supervised diagnosis model, thereby further improving the efficiency of fault diagnosis.

[0211] Embodiment 3:

[0212] This embodiment provides a diagnostic system corresponding to the unsupervised diagnostic method for rail transit electromechanical equipment in Embodiment 1, including the following modules.

[0213] Operation data collection module: used to collect the operation data of rail transit electromechanical equipment to form the original operation data set;

[0214] Feature selection module: used to extract the time domain, frequency domain, and time-frequency domain features of the original running data set to form an original feature set, score each feature of the original feature set, and select features based on the scoring results using the Pareto principle to form a filtered feature set;

[0215] Clustering model parameter module: used to provide a soft clustering algorithm model. Given a range of cluster numbers, the filtered feature set is used as a training set. The soft clustering algorithm model is trained cyclically until convergence, and the initial parameters of the unsupervised clustering algorithm model are obtained. This step is repeated to obtain all training set unsupervised clustering algorithm models within the range of cluster numbers.

[0216] Clustering model training module: used to select any one from the unsupervised clustering algorithm models in the training set, obtain the training set clustering result of the model, and then merge the test data with the training set data as the input of the soft clustering algorithm model, train all diagnostic models within the clustering number range, and compare the clustering results of each diagnostic model with the unsupervised clustering algorithm model of the training set in turn, give a consistency score, and obtain the diagnostic model with the best consistency score with the unsupervised clustering algorithm model of the training set. If the consistency score corresponding to the diagnostic model meets the set value, the corresponding clustering result of the test data is output. If it does not meet the set value, another unused training set unsupervised clustering algorithm model is selected to repeat this module until a diagnostic model that meets the set value is obtained, and the clustering result of the test data of the diagnostic model is output;

[0217] Clustering result processing module: used to classify and mark each cluster sample according to the clustering results of the test data output by the diagnostic model, issue an alarm for the cluster samples that do not meet the set values, and update the training set according to the classification and marking results for the next round of diagnostic model training.

[0218] In some specific embodiments, the feature selection module includes the following submodules:

[0219] Feature extraction module: used to perform integrated empirical mode decomposition on the original operation data set, extract the mode with the largest information entropy in the original operation data signal, calculate the time domain, frequency domain, and time-frequency domain features of the original operation data set, and form an original feature set containing multiple dimensions;

[0220] Feature scoring module: used to score each feature in the original feature set according to the maximum information compression index or graph index;

[0221] Feature filtering module: It is used to calculate the cumulative scores of features in descending order according to the Pareto principle until the cumulative score accounts for more than or equal to 80%, and then extract the corresponding features to form a filtered feature set.

[0222] In some specific embodiments, the soft clustering algorithm model includes: a density-based soft clustering algorithm, a model-based soft clustering algorithm, or a fuzzy-based soft clustering algorithm; the density-based soft clustering algorithm includes a DBSCAN algorithm or an OPTICS algorithm, the model-based soft clustering algorithm includes a GMM algorithm, and the fuzzy-based soft clustering algorithm includes an FCM algorithm or an FMM algorithm;

[0223] Unsupervised clustering algorithm models include: center-based unsupervised clustering algorithm, density-based unsupervised clustering algorithm or connectivity-based unsupervised clustering algorithm; center-based unsupervised clustering algorithms include K-Means, Bisecting K-Means, GMM algorithms, density-based unsupervised clustering algorithms include DBSCAN, HDBSCAN, OPTICS, Mean Shift, DensityPeak Clustering algorithms, connectivity-based unsupervised clustering algorithms include hierarchical clustering, K nearest neighbor, ROCK, SOM, spectral clustering algorithms.

[0224] In some specific embodiments, in the clustering model parameter module, when the FCM model is selected as the soft clustering algorithm model, the following submodules are included:

[0225] FCM model training module: used to select a cluster number c within the cluster number range, randomize c cluster centers, determine the initial cluster center of the FCM model using the potential function method, and use the sample data of the training set to cyclically train the FCM model until the FCM model converges;

[0226] Unsupervised clustering algorithm model training module: used to use the EM algorithm to iterate parameters and converge to obtain the optimal unsupervised clustering algorithm model under the cluster number c as the training set unsupervised clustering algorithm model;

[0227] Cyclic training module: used to repeatedly run the FCM model training module and the unsupervised clustering algorithm model training module to obtain all training set unsupervised clustering algorithm models within the range of cluster numbers.

[0228] In some specific embodiments, in the clustering model training module, the training set data samples that are in the same data cluster in the clustering results of the training set unsupervised clustering algorithm model and the diagnostic model are counted, and the training set data samples are marked as stable samples. The consistency score is the proportion of stable samples in the training set data.

[0229] In some specific embodiments, the clustering model parameter module further includes the following submodules:

[0230] Unsupervised clustering algorithm model sorting module: used to calculate the Xie-Beni index of each training set unsupervised clustering algorithm model, select the cluster number with the smallest Xie-Beni index as the optimal cluster number, output the corresponding optimal training set unsupervised clustering algorithm model, and sort the remaining training set unsupervised clustering algorithm models in ascending order according to the Xie-Beni index.

[0231] In some specific embodiments, in the clustering model parameter module, the clustering result of the optimal training set unsupervised clustering algorithm model is obtained, and then the test data and the training set data are merged as the input of the soft clustering algorithm model, and all diagnostic models within the clustering number range are trained, and the clustering results of each diagnostic model and the optimal training set unsupervised clustering algorithm model are compared in turn, and a consistency score is given, and the diagnostic model with the best consistency score with the optimal training set unsupervised clustering algorithm model is selected. If the consistency score corresponding to the diagnostic model meets the set value, the corresponding clustering result of the test data is output; if it does not meet the set value, this module is repeatedly executed for the next-ranked training set unsupervised clustering algorithm model until a diagnostic model that meets the set value is obtained, and the clustering result of the test data of the diagnostic model is output.

[0232] In some specific embodiments, the feature selection module further includes the following submodules:

[0233] Outlier sample removal module: used to calculate the outliers of each sample in the filter feature set. Indicators and Indicators are obtained by performing mean analysis to obtain the comprehensive statistical indicator fai, and outlier sample data in the training set that exceeds the control limit of the comprehensive statistical indicator fai are eliminated to obtain the training set samples.

[0234] In some specific embodiments, in the clustering result processing module, the classification label includes:

[0235] The samples in the cluster samples that only contain the data to be tested and whose data volume is greater than or equal to the set value are marked as unknown state samples, and a warning is issued; the samples in the cluster samples that only contain the data to be tested and whose data volume is less than the set value are marked as outlier samples; the samples in the cluster samples that contain the data to be tested and the training set data are marked as healthy state samples and marked.

[0236] In some specific embodiments, in the clustering result processing module, the test data marked as healthy state samples are added to the training set, and the equipment components corresponding to the test data marked as unknown state samples are inspected and repaired. If it is determined that the component is in a faulty state, the sample is marked as a faulty sample, and a supervised diagnosis model is constructed. After that, the test data is first diagnosed through the supervised diagnosis model, and then enters the diagnosis model for diagnosis.

[0237] In summary, in the above-mentioned specific implementation of the present invention, the key component operation data collected by rail transit electromechanical equipment is fully utilized. Without the need for manual labeling of data, the algorithm process can automatically complete the end-to-end unsupervised fault diagnosis process of signal preprocessing, feature selection, abnormal sample elimination, unsupervised model construction, and abnormal state identification of key components, and can continuously and automatically update the training set according to the diagnosis results, thereby ensuring that the diagnosis model always has effective diagnosis capabilities in the long-term mainline operation. At the same time, labeled fault data can be obtained based on the diagnosis results, thereby providing feasibility for building a supervised diagnosis model and further improving overall work efficiency.

Claims

1. An unsupervised diagnosis method for rail transit electromechanical equipment, characterized in that: The following steps are involved: (1) Collect the operating data of rail transit electromechanical equipment to form the original operating data set; (2) Extract the time domain, frequency domain, and time-frequency domain features of the original running data set to form an original feature set, score each feature of the original feature set, and select features based on the scoring results using the Pareto principle to form a filtered feature set; (3) Provide a soft clustering algorithm model, give a range of cluster numbers, use the filtered feature set as the training set, cyclically train the soft clustering algorithm model until convergence, obtain the initial parameters of the unsupervised clustering algorithm model, repeat this step, and obtain all training set unsupervised clustering algorithm models within the range of cluster numbers; (4) Select any one of the unsupervised clustering algorithm models in the training set, obtain the training set clustering result of the model, then merge the test data with the training set data as the input of the soft clustering algorithm model, train all diagnostic models within the clustering number range, and compare the clustering results of each diagnostic model with the unsupervised clustering algorithm model in the training set in turn, give a consistency score, and obtain the diagnostic model with the best consistency score with the unsupervised clustering algorithm model in the training set. If the consistency score corresponding to the diagnostic model meets the set value, output the corresponding clustering result of the test data. If it does not meet the set value, select another unused training set unsupervised clustering algorithm model and repeat this step until a diagnostic model that meets the set value is obtained, and output the test data clustering result of the diagnostic model; (5) Based on the clustering results of the test data output by the diagnostic model, classify and mark the samples in each cluster, issue an alarm for the cluster samples that do not meet the set values, and update the training set based on the classification and marking results for the next round of diagnostic model training.

2. The unsupervised diagnosis method for rail transit electromechanical equipment according to claim 1 is characterized in that: The step (2) comprises the following steps: (21) Performing integrated empirical mode decomposition on the original operation data set, extracting the mode with the largest information entropy in the original operation data signal, calculating the time domain, frequency domain, and time-frequency domain features of the original operation data set, and forming an original feature set containing multiple dimensions; (22) Score each feature in the original feature set according to the redundancy index or graph index; (23) According to the Pareto principle, the cumulative scores of the features are calculated in descending order until the cumulative score accounts for more than or equal to 80%. The corresponding features are then taken out to form a filtered feature set.

3. The unsupervised diagnosis method for rail transit electromechanical equipment according to claim 1 is characterized in that: The soft clustering algorithm model includes: a density-based soft clustering algorithm, a model-based soft clustering algorithm or a fuzzy-based soft clustering algorithm; the density-based soft clustering algorithm includes a DBSCAN algorithm or an OPTICS algorithm, the model-based soft clustering algorithm includes a GMM algorithm, and the fuzzy-based soft clustering algorithm includes an FCM algorithm or an FMM algorithm; The unsupervised clustering algorithm models include: a center-based unsupervised clustering algorithm, a density-based unsupervised clustering algorithm or a connectivity-based unsupervised clustering algorithm; the center-based unsupervised clustering algorithms include K-Means, Bisecting K-Means, and GMM algorithms; the density-based unsupervised clustering algorithms include DBSCAN, HDBSCAN, OPTICS, Mean Shift, and DensityPeak Clustering algorithms; the connectivity-based unsupervised clustering algorithms include hierarchical clustering, K nearest neighbor, ROCK, SOM, and spectral clustering algorithms.

4. The unsupervised diagnosis method for rail transit electromechanical equipment according to claim 3 is characterized in that: When the FCM algorithm model is selected as the soft clustering algorithm model in step (3), the following steps are included: (31) Select a cluster number c within the cluster number range, randomize c cluster centers, use the potential function method to determine the initial cluster center of the FCM model, and use the sample data of the training set to cyclically train the FCM model until the FCM model converges; (32) Use the EM algorithm to iterate parameters and converge to obtain the optimal unsupervised clustering algorithm model under the cluster number c, which is used as the unsupervised clustering algorithm model of the training set; (33) Repeat the above steps (31)-(32) to obtain the unsupervised clustering algorithm model for all training sets within the range of cluster numbers.

5. The unsupervised diagnosis method for rail transit electromechanical equipment according to claim 1 is characterized in that: In the step (4), the training set data samples that are in the same data cluster in the clustering results of the training set unsupervised clustering algorithm model and the diagnostic model are counted, and the training set data samples are marked as stable samples. The consistency score is the proportion of stable samples in the training set data.

6. The unsupervised diagnosis method for rail transit electromechanical equipment according to claim 1 is characterized in that: The step (3) further comprises the following steps: Calculate the Xie-Beni index of each training set unsupervised clustering algorithm model, select the cluster number with the smallest Xie-Beni index as the optimal cluster number, output the corresponding optimal training set unsupervised clustering algorithm model, and sort the remaining training set unsupervised clustering algorithm models in ascending order according to the Xie-Beni index.

7. The unsupervised diagnosis method for rail transit electromechanical equipment according to claim 6 is characterized in that: In the step (4), the clustering result of the optimal training set unsupervised clustering algorithm model is obtained, and then the test data and the training set data are merged as the input of the soft clustering algorithm model, and all diagnostic models within the clustering number range are trained, and the clustering results of each diagnostic model and the optimal training set unsupervised clustering algorithm model are compared in turn, and a consistency score is given, and the diagnostic model with the best consistency score with the optimal training set unsupervised clustering algorithm model is selected. If the consistency score corresponding to the diagnostic model meets the set value, the corresponding clustering result of the test data is output; if it does not meet the set value, this step is repeated for the next training set unsupervised clustering algorithm model until a diagnostic model that meets the set value is obtained, and the clustering result of the test data of the diagnostic model is output.

8. The unsupervised diagnosis method for rail transit electromechanical equipment according to claim 1 is characterized in that: The step (2) further comprises the following steps: Calculate the SPE index and T of each sample in the filtered feature set 2 Indicators are obtained by performing mean analysis to obtain comprehensive statistical indicators, and outlier sample data in the training set that exceed the control limit of the comprehensive statistical indicators are eliminated to obtain training set samples.

9. The unsupervised diagnosis method for rail transit electromechanical equipment according to claim 1, characterized in that: In step (5), the classification marking includes: The samples in the cluster samples that only contain the data to be tested and whose data volume is greater than or equal to the set value are marked as unknown state samples, and a warning is issued; the samples in the cluster samples that only contain the data to be tested and whose data volume is less than the set value are marked as outlier samples; the samples in the cluster samples that contain the data to be tested and the training set data are marked as healthy state samples and marked.

10. The unsupervised diagnosis method for rail transit electromechanical equipment according to claim 9, characterized in that: In the step (5), the test data marked as healthy state samples are added to the training set, and the equipment components corresponding to the test data marked as unknown state samples are inspected and repaired. If it is determined that the component is in a faulty state, the sample is marked as a faulty sample, and a supervised diagnosis model is constructed. Thereafter, the test data is first diagnosed by the supervised diagnosis model, and then enters the diagnosis model for diagnosis.

11. An unsupervised diagnosis system for rail transit electromechanical equipment, characterized in that: include: Operation data collection module: used to collect the operation data of rail transit electromechanical equipment to form the original operation data set; Feature selection module: used to extract the time domain, frequency domain, and time-frequency domain features of the original running data set to form an original feature set, score each feature of the original feature set, and select features based on the scoring results using the Pareto principle to form a filtered feature set; Clustering model parameter module: used to provide a soft clustering algorithm model. Given a range of cluster numbers, the filtered feature set is used as a training set. The soft clustering algorithm model is trained cyclically until convergence, and the initial parameters of the unsupervised clustering algorithm model are obtained. This step is repeated to obtain all training set unsupervised clustering algorithm models within the range of cluster numbers. Clustering model training module: used to select any one from the unsupervised clustering algorithm models in the training set, obtain the training set clustering result of the model, and then merge the test data with the training set data as the input of the soft clustering algorithm model, train all diagnostic models within the clustering number range, and compare the clustering results of each diagnostic model with the unsupervised clustering algorithm model of the training set in turn, give a consistency score, and obtain the diagnostic model with the best consistency score with the unsupervised clustering algorithm model of the training set. If the consistency score corresponding to the diagnostic model meets the set value, the corresponding clustering result of the test data is output. If it does not meet the set value, another unused training set unsupervised clustering algorithm model is selected to repeat this module until a diagnostic model that meets the set value is obtained, and the clustering result of the test data of the diagnostic model is output; Clustering result processing module: used to classify and mark each cluster sample according to the clustering results of the test data output by the diagnostic model, issue an alarm for the cluster samples that do not meet the set values, and update the training set according to the classification and marking results for the next round of diagnostic model training.

12. The unsupervised diagnosis system for rail transit electromechanical equipment according to claim 11, characterized in that: The feature selection module includes the following submodules: Feature extraction module: used to perform integrated empirical mode decomposition on the original operation data set, extract the mode with the largest information entropy in the original operation data signal, calculate the time domain, frequency domain, and time-frequency domain features of the original operation data set, and form an original feature set containing multiple dimensions; Feature scoring module: used to score each feature in the original feature set according to the redundancy index or graph index; Feature filtering module: It is used to calculate the cumulative scores of features in descending order according to the Pareto principle until the cumulative score accounts for more than or equal to 80%, and then extract the corresponding features to form a filtered feature set.

13. The unsupervised diagnosis system for rail transit electromechanical equipment according to claim 11, characterized in that: The soft clustering algorithm model includes: a density-based soft clustering algorithm, a model-based soft clustering algorithm or a fuzzy-based soft clustering algorithm; the density-based soft clustering algorithm includes a DBSCAN algorithm or an OPTICS algorithm, the model-based soft clustering algorithm includes a GMM algorithm, and the fuzzy-based soft clustering algorithm includes an FCM algorithm or an FMM algorithm; The unsupervised clustering algorithm models include: a center-based unsupervised clustering algorithm, a density-based unsupervised clustering algorithm or a connectivity-based unsupervised clustering algorithm; the center-based unsupervised clustering algorithms include K-Means, Bisecting K-Means, and GMM algorithms; the density-based unsupervised clustering algorithms include DBSCAN, HDBSCAN, OPTICS, Mean Shift, and DensityPeak Clustering algorithms; the connectivity-based unsupervised clustering algorithms include hierarchical clustering, K nearest neighbor, ROCK, SOM, and spectral clustering algorithms.

14. The unsupervised diagnosis system for rail transit electromechanical equipment according to claim 13, characterized in that: In the clustering model parameter module, when the FCM model is selected as the soft clustering algorithm model, the following submodules are included: FCM model training module: used to select a cluster number c within the cluster number range, randomize c cluster centers, determine the initial cluster center of the FCM model using the potential function method, and use the sample data of the training set to cyclically train the FCM model until the FCM model converges; Unsupervised clustering algorithm model training module: used to use the EM algorithm to iterate parameters and converge to obtain the optimal unsupervised clustering algorithm model under the cluster number c as the training set unsupervised clustering algorithm model; Cyclic training module: used to repeatedly run the FCM model training module and the unsupervised clustering algorithm model training module to obtain all training set unsupervised clustering algorithm models within the range of cluster numbers.

15. The unsupervised diagnosis system for rail transit electromechanical equipment according to claim 11, characterized in that: In the clustering model training module, the training set data samples that are in the same data cluster in the clustering results of the training set unsupervised clustering algorithm model and the diagnostic model are counted, and the training set data samples are marked as stable samples. The consistency score is the proportion of stable samples in the training set data.

16. The unsupervised diagnosis system for rail transit electromechanical equipment according to claim 11, characterized in that: The clustering model parameter module also includes the following submodules: Unsupervised clustering algorithm model sorting module: used to calculate the Xie-Beni index of each training set unsupervised clustering algorithm model, select the cluster number with the smallest Xie-Beni index as the optimal cluster number, output the corresponding optimal training set unsupervised clustering algorithm model, and sort the remaining training set unsupervised clustering algorithm models in ascending order according to the Xie-Beni index.

17. The unsupervised diagnosis system for rail transit electromechanical equipment according to claim 16, characterized in that: In the clustering model parameter module, the clustering result of the optimal training set unsupervised clustering algorithm model is obtained, and then the data to be tested and the training set data are merged as the input of the soft clustering algorithm model, and all diagnostic models within the clustering number range are trained, and the clustering results of each diagnostic model and the optimal training set unsupervised clustering algorithm model are compared in turn, and a consistency score is given. The diagnostic model with the best consistency score with the optimal training set unsupervised clustering algorithm model is selected. If the consistency score corresponding to the diagnostic model meets the set value, the corresponding clustering result of the data to be tested is output. If it does not meet the set value, this module is repeatedly executed for the next-ranked training set unsupervised clustering algorithm model until a diagnostic model that meets the set value is obtained, and the clustering result of the data to be tested of the diagnostic model is output.

18. The unsupervised diagnosis system for rail transit electromechanical equipment according to claim 11, characterized in that: The feature selection module also includes the following submodules: Outlier sample removal module: Calculate the SPE index and T of each sample in the filtered feature set 2 Indicators are obtained by performing mean analysis to obtain comprehensive statistical indicators, and outlier sample data in the training set that exceed the control limit of the comprehensive statistical indicators are eliminated to obtain training set samples.

19. The unsupervised diagnosis system for rail transit electromechanical equipment according to claim 11, characterized in that: In the clustering result processing module, the classification mark includes: The samples in the cluster samples that only contain the data to be tested and whose data volume is greater than or equal to the set value are marked as unknown state samples, and a warning is issued; the samples in the cluster samples that only contain the data to be tested and whose data volume is less than the set value are marked as outlier samples; the samples in the cluster samples that contain the data to be tested and the training set data are marked as healthy state samples and marked.

20. The unsupervised diagnosis system for rail transit electromechanical equipment according to claim 19, characterized in that: In the clustering result processing module, the test data marked as healthy state samples are added to the training set, and the equipment components corresponding to the test data marked as unknown state samples are inspected. If it is determined that the component is in a faulty state, the sample is marked as a faulty sample, and a supervised diagnosis model is constructed. After that, the test data is first diagnosed through the supervised diagnosis model, and then enters the diagnosis model for diagnosis.