A method and system for industrial equipment fault diagnosis based on multi-view integrated features

Abstract

This invention relates to a method and system for industrial equipment fault diagnosis based on multi-view integrated features, belonging to the field of industrial equipment fault diagnosis technology. The invention addresses the problem that existing methods suffer from significant performance degradation when processing high-noise, small-sample, or non-stationary operating condition data, severely restricting the accuracy and reliability of industrial equipment fault diagnosis. Key technical points include: acquiring a time-series dataset of industrial equipment operating fault states; performing feature extraction and enhancement on the time-series data; extracting feature vectors from the original time series, first-order difference time series, and dilated time series sample sets, and concatenating them into a high-dimensional initial feature vector set; performing multi-view integrated feature selection on the initial feature vector set; performing multi-metric scoring and initial view generation on the high-dimensional feature set; heterogeneous integration classification and diagnosis; and outputting the results. Through three core technologies—multi-view feature enhancement, dynamic selection, and heterogeneous integration—this invention solves the key bottlenecks of traditional fault diagnosis methods in industrial fault diagnosis, providing an innovative solution for high-precision, low-cost equipment health management. This invention balances feature diversity, expressive power, and model robustness, while adapting to the stringent constraints of industrial scenarios.

Description

Technical Field

[0001] This invention relates to the field of industrial equipment fault diagnosis technology, specifically to an industrial equipment fault diagnosis method and system based on multi-view integration features. Background Technology

[0002] With the increasing intelligence and complexity of industrial equipment, fault diagnosis technology based on condition monitoring signals has become a key means to ensure the reliable operation of industrial systems. However, in actual industrial scenarios, only single-channel monitoring signals (such as current, vibration, etc.) can usually be obtained, making it impossible to rely on multi-source data or domain expert knowledge for auxiliary analysis. Against this backdrop, data-driven univariate time series classification methods have become the core approach to solving fault diagnosis problems. Existing feature-based fault diagnosis methods extract descriptive statistical features (such as mean, variance, entropy, etc.) from time series data and combine them with classifiers to identify fault types. Their process is transparent and highly interpretable, and they have been widely used in industrial equipment health management.

[0003] In recent years, research on industrial equipment fault diagnosis has increased significantly. For example, existing patent document CN117726320A provides a method and apparatus for industrial equipment fault diagnosis. This method first constructs an industrial equipment fault diagnosis knowledge graph; then constructs a user-item interaction matrix, updating the values ​​of each element in the user-item interaction matrix based on fault triples; subsequently, it constructs a knowledge recommendation model based on graph neural networks and multi-task learning; finally, it trains and validates the knowledge recommendation model based on graph neural networks and multi-task learning. This existing technology introduces knowledge graphs and recommendation systems into the field of fault diagnosis. By extracting feature information from the industrial equipment fault diagnosis knowledge graph through the knowledge recommendation model based on graph neural networks and multi-task learning, it embeds the information into learning and knowledge reasoning. By training the user-item interaction matrix, it predicts the user's target, providing industrial equipment maintenance personnel with fast, accurate, and personalized maintenance methods when handling equipment faults. Existing technology CN118296452A discloses an industrial equipment fault diagnosis method based on Transformer model optimization, which solves the shortcomings of low accuracy, weak generalization, and weak anti-interference ability of fault diagnosis models. The process includes the following steps: acquiring industrial equipment operating data; preprocessing the industrial equipment operating data; constructing an industrial equipment fault diagnosis model; training the industrial equipment fault diagnosis model; acquiring the operating data of the industrial equipment to be diagnosed; and diagnosing the industrial equipment fault. This invention uses a hawk-swarm optimization algorithm to optimize the structure of the Transformer model, ensuring that the optimal number of encoder layers, attention heads, and algorithm iterations are maintained during model operation; increasing the model's robustness and generalization ability, and improving the accuracy of industrial equipment fault diagnosis. The prior art, document CN118034261A, discloses an artificial intelligence-based industrial equipment fault diagnosis system and method. It includes a fault detection unit: setting basic target operating data and establishing a standard range for the operating data; then analyzing the range of changes in the operating data to determine if a fault exists in the basic target. This document analyzes the range of changes in operating data to determine if a fault exists in the industrial equipment circuit and matches corresponding fault solutions, reducing the time required for matching fault diagnosis solutions and enabling real-time monitoring and rapid fault location. The binary search method for fault node maintenance maximizes time savings in fault diagnosis and location. Prioritized diagnosis using marked fault ranges allows for prioritization based on past fault data. Based on artificial intelligence analysis, this effectively improves the efficiency of industrial equipment fault diagnosis.

[0004] However, existing technologies still have the following significant drawbacks:

[0005] Feature redundancy and insufficient scalability: Mainstream feature extraction tools (such as HCTSA, Catch22, TSFresh, etc.) generate a large number of features through preset statistical formulas, but these feature sets generally contain highly redundant statistical patterns (such as repeated time-domain statistics). For example, among the more than 7,700 features extracted by HCTSA, more than 60% are significantly correlated, making subsequent classifiers susceptible to interference from redundant features. At the same time, existing frameworks lack flexible feature expansion mechanisms, making it difficult to introduce new low-redundancy features.

[0006] Limited feature representation capabilities: Traditional methods typically perform global statistics directly on the original time series or extract local features through a sliding window, but fail to effectively combine complementary information from global and local patterns. For example, while TSFresh supports multi-scale feature extraction, its window division method is fixed and cannot adaptively enhance time series patterns at different granularities. Furthermore, existing methods ignore the structural relationships within the sequence (such as the spatial distribution of time points and phase changes), making it difficult for features to capture the deep discriminative information of fault signals.

[0007] Feature selection strategies are often simplistic: Existing methods mostly employ embedded feature selection (such as LASSO regression and random forest importance scoring) or single filtering metrics (such as mutual information and variance thresholding), without considering the interaction effects and stability between features. For example, TSFresh selects features through hypothesis testing, but its independence assumption often fails in real-world industrial data, leading to the accidental deletion of important features. Furthermore, existing methods lack dynamic adjustment mechanisms and cannot generate robust feature subsets tailored to the characteristics of different datasets.

[0008] Insufficient classifier diversity: Current feature-based fault diagnosis algorithms mostly employ a single classification model (such as SVM or random forest) or simply integrate homogeneous classifiers, resulting in limited model generalization ability. For example, although FreshP improves classification performance through rotating forests, its base classifiers are all decision trees, making it difficult to cover the discrimination boundaries under different data distributions.

[0009] The aforementioned shortcomings severely limit the accuracy and reliability of fault diagnosis for industrial equipment, especially when processing high-noise, small-sample, or non-stationary operating condition data, where the performance of existing methods deteriorates significantly. Therefore, there is an urgent need for a new fault diagnosis framework that can balance feature diversity, expressive power, and model robustness, while adapting to the stringent constraints of industrial scenarios. Summary of the Invention

[0010] The technical problem to be solved by the present invention

[0011] To address the shortcomings of existing methods described in the background section, the present invention aims to solve the following technical problems:

[0012] 1. Construct an extensible feature enhancement framework to overcome the redundancy bottleneck of traditional feature sets, support the flexible introduction of new low-redundancy features, and dynamically optimize feature combinations to adapt to the needs of different industrial scenarios.

[0013] 2. Achieve multi-view collaborative enhancement of global and local modes. Through an innovative sequence mapping mechanism, simultaneously capture the macroscopic trends and microscopic fluctuations of time-series signals to improve the feature discrimination capability.

[0014] 3. Design an adaptive feature selection strategy, based on the dual indicators of feature stability and diversity, to dynamically select a subset of high-value features and avoid overfitting or information loss caused by traditional single measurement.

[0015] 4. Construct a heterogeneous ensemble classifier that integrates classification models based on different principles (such as ridge regression and extreme random trees) and improves the model's generalization ability to complex failure modes through diversity enhancement mechanisms (such as hard voting and weighted fusion).

[0016] The technical solution adopted by the present invention to solve the above-mentioned technical problems is as follows:

[0017] This invention provides a method for fault diagnosis of industrial equipment based on multi-view integration features, characterized by comprising the following steps:

[0018] Step S1: Obtain a time series dataset of the operational fault status of industrial equipment. The time series dataset contains a sample set S of univariate time series data. L =(s1,s2,…,s n ) and the label set L = (l1, l2, ..., l n The training set Train = (S) consists of... L (L) and a sample set S of univariate time series T =(s n+1 ,s n+2 ,…,s n+m Test set consisting of (S) T ), where n and m are the number of time series samples contained in the training set and the test set, respectively, and the time series samples s i =(s1,s2,…,s σ All are univariate time series with no missing values, equal intervals, and equal length, where 1 ≤ i ≤ n + m, and σ is the length of the time series sample.

[0019] Step S2: Perform feature extraction and enhancement on the time series dataset, including:

[0020] Step S2.1: Extract statistical features from the original time series sample set S to obtain the feature vector FV. RThe original time series sample set S includes the training set and the univariate time series sample set S. L With the test set univariate time series sample set S T ;

[0021] Perform c groups of dilation mappings at different scales on the original time series sample set S to obtain the dilated sample set S. s1 ,…,S sc For the sample set S s1 ,…,S sc Perform statistical feature extraction to obtain the feature vector FV Rs1 ,…,FV Rsc ;

[0022] Step S2.2: Perform a first-order difference transformation on the original time series sample set S to obtain the first-order difference time series sample set S. D ;

[0023] For the first-order difference time series sample set S D Perform statistical feature extraction to obtain the feature vector FV D ;

[0024] For the first-order difference time series sample set S D Perform dilation mapping at different scales in group c to obtain the dilated sample set S. Ds1 ,…,S Dsc For the sample set S Ds1 ,…,S Dsc Perform statistical feature extraction to obtain the feature vector FV Ds1 ,…,FV Dsc ;

[0025] Step S2.3: Convert the feature vector FV R FV Rs1 ,…,FV Rsc FV D FV Ds1 ,…,FV Dsc Concatenate them to form a high-dimensional initial feature vector set FV;

[0026] Step S3: Perform multi-view integrated feature selection on the initial feature vector set FV, including:

[0027] Step S3.1: Perform multi-metric scoring and generate initial view on the high-dimensional feature set:

[0028] M independent feature evaluation indicators are used to score each feature in the initial feature vector set FV. Each feature evaluation indicator generates a corresponding score matrix, resulting in M ​​score matrices.

[0029] Based on the scoring matrix corresponding to the i-th feature evaluation index, sort each feature in descending order and select the top K1, K2, ..., K1 features. T Given several high-scoring features, we obtain the initial view feature set FV corresponding to the i-th feature evaluation index. 1i ,FV 2i ,…,FV Ti , where i is a positive integer less than M;

[0030] The i increases sequentially from 1 to M, where i = 1, 2, ..., M, thus obtaining the initial view feature set corresponding to the first feature evaluation index. The initial view feature set FV corresponding to the second type of feature evaluation index 12 ,FV 22 ,…,FV T2 ..., the initial view feature set corresponding to the Mth feature evaluation index

[0031] The number of features K1, K2, ..., K T Based on the preset convergence threshold list thresholds = [threshold1, threshold2, ..., threshold] T Determine the number of candidate features K (K = 1, 2, ..., N). feature ), calculate using formula Select in sequence those that satisfy rate(K)≤threshold i The smallest eigenvalue is K i , 1≤i≤T, where rate(K) is the relative rate of change of the stability-diversity score (SD-score), N feature The number of features in the initial feature vector set FV;

[0032] For each candidate feature quantity K, based on the scoring matrix corresponding to each feature evaluation index j, the top K high-scoring features are selected from the initial feature vector set FV to obtain a feature subset. The j increases sequentially from 1 to M, where j = 1, 2, ..., M, thus obtaining the feature subsets corresponding to all feature evaluation indicators. Calculate the feature subset corresponding to the number of candidate features K. Pearson correlation coefficient absolute value matrix Each sub-block of the absolute value matrix A(K) of the Pearson correlation coefficient All are K×K dimensional square matrices. Representing a feature subset With feature subset The absolute value matrix of the Pearson correlation coefficients between them 1≤p≤M, 1≤q≤M;

[0033] The stability-diversity score (SD-score) corresponding to the number of candidate features K is expressed by the formula. Calculation, where Indicates a pair of child blocks Summing all elements;

[0034] Step S3.2: Generate the intersection-optimized extended view feature set:

[0035] For the initial view feature set FV corresponding to the first type of feature evaluation index 11 ,FV 21 ,…,FV T1 The initial view feature set FV corresponding to the second feature evaluation index 12 ,FV 22 ,…,FV T2 ..., the initial view feature set FV corresponding to the Mth feature evaluation index 1M ,FV 2M ,…,FV TM For each of the initial view feature sets corresponding to different feature evaluation indicators, the intersection of feature indices is taken to generate T common feature subsets. in

[0036] The common feature subset Concatenate into a common feature vector FV interse and the common feature vector FV intersection Initial view feature sets FV corresponding to all feature evaluation metrics 11 ,FV 21 ,…,FV T1 FV 12 ,FV 22 ,…,FV T2 , ..., FV 1M ,FV 2M ,…,FV TM The features are stitched together to generate N = T·M extended view feature sets FV'. V1 ,FV' V2 ,…,FV' VN ;

[0037] Step S3.3: Redundancy removal and output of the final multi-view feature set:

[0038] For each of the extended view feature sets FV' V1 ,FV' V2 ,…,FV' VNPearson correlation coefficient analysis was performed on the features, and redundant features with correlation coefficients greater than a preset threshold were removed to obtain the final multi-view feature set FV'1, FV'2, ..., FV' N ;

[0039] Step S4, Heterogeneous Integration Classification and Diagnosis:

[0040] Classifier training: using the training set Train = (S L L) The training set multi-view feature set is obtained through steps S2 and S3. Feature set in each view The classifier CL is trained on the above. i , 1≤i≤N;

[0041] Hard voting integration: using the test set Test = (S T The test set multi-view feature set is obtained through steps S2 and S3. Feature set of each view Input classifier CL i The multi-view prediction result L is obtained. i , 1≤i≤N;

[0042] The classification results L1, L2, L, L are integrated from all views using a majority voting rule. N Output the final fault label L T ;

[0043] Step S5, Result Output: Output the diagnostic label L T The data is transmitted to the industrial equipment control terminal via the communication module, triggering fault warnings or maintenance commands.

[0044] Furthermore, the first-order difference transform of the original time series sample set S in step S2.2 can be implemented by a tracking differentiator with an adjustable filter factor k. The function implementation process of the tracking differentiator with an adjustable filter factor is as follows:

[0045] The formulas for calculating the first column x1[1] of matrix x1 and the first column x2[1] of matrix x2 are as follows:

[0046]

[0047] Wherein, time series sample s is the time series sample input to the tracking differentiator in the original time series sample set S, s[i] is the i-th value in time series sample s, 1≤i≤σ, where σ is the length of time series sample s;

[0048] For i = 1, 2, ..., σ increases sequentially:

[0049] y i= x1[i] - s[i] + khx2[i];

[0050] If |y i |>k 2 h 2 r, then otherwise,

[0051] If |a|>rhk, then fh i = -r·sign(a), otherwise,

[0052] The formulas for calculating the (i+1)th column x1[i+1] of matrix x1 and the (i+1)th column x2[i+1] of matrix x2 are as follows:

[0053]

[0054] By performing a differential transformation on the univariate time series using a tracking differentiator with an adjustable filter factor, the resulting differential time series is: Wherein, r is a preset speed factor parameter, h is a preset step size parameter, and k is an adjustable filter factor.

[0055] Furthermore, the iterative optimization process of the adjustable filter factor k of the tracking differentiator is as follows:

[0056] (a) Setting optimization parameters: number of candidate solutions nc = 10, maximum number of iterations inter max =10. Filter factor search range [b L ,b R ], where the search range is [b L ,b R The default value for ] is [1,3];

[0057] (b) Define the objective function f(k):

[0058] i. For the training set time series S L Generate a set of differential sequences using the filter factor k

[0059] ii. Regarding The following sampling rates are sorted in ascending order by row. rate =4 Extracting quantile features

[0060] iii. Calculate the feature set The contour score (S-score) of label L is used as the objective function value f(k), and the filter factor k is iteratively optimized by maximizing the contour score (S-score).

[0061] The contour score (S-score) is defined as follows: Where n is the number of time series samples contained in the training set, and SC(i) represents the feature set. The silhouette coefficient of the feature vector corresponding to the i-th time series sample. Where η(i) represents the minimum distance between clusters, i.e., the feature set The minimum Euclidean distance between the feature vector corresponding to the i-th time series sample and the feature vectors corresponding to all time series samples with different labels is defined as follows: The average distance within a cluster, i.e., the feature set. The average Euclidean distance between the feature vector corresponding to the i-th time series sample and the feature vectors corresponding to all time series samples with the same label;

[0062] (c) Iterative optimization:

[0063] i. Initialize the candidate solution set {k1,k2,…,k nc}∈[b L ,b R ];

[0064] ii. According to k i The objective function f(k) takes the value f(k) i Sort the candidate solutions from largest to smallest, 1≤i≤nc, and select the k that makes the objective function f(k) the largest, second largest, and third largest respectively. i As the optimal solution k α suboptimal solution k β and the third optimal solution k δ ;

[0065] iii. Update convergence coefficients and random vectors Where t is the current iteration number. A vector whose modulus takes random values ​​between [0,1]. It is a unit vector;

[0066] iv. Update candidate solution positions:

[0067]

[0068] in Representative vector The i-th term, k(t-1) is the value of the adjustable filter factor at iteration number t-1, k(t) is the value of the adjustable filter factor at iteration number t, k(0) = k α ;

[0069] v. Repeat the iteration until the maximum number of iterations is reached. max The optimal solution k of the adjustable filter factor of the output tracking differentiator adj =k(inter max ).

[0070] Furthermore, the feature evaluation metrics in step S3.1 include mutual information (MI) and analysis of variance (ANOVA).

[0071] Furthermore, the specific implementation steps of the dilation mapping in steps S2.1 and S2.2 are as follows:

[0072] (a) Bidirectional expansion mapping:

[0073] For the input time series s i =(s1,s2,…,s σ Global mode enhancement is performed based on the shuffle rate R, where R is an integer greater than 1. Specific operations include:

[0074] Forward downsampling: from s i =(s1,s2,…,s σ Starting from the i-th data point, 1≤i≤R, data points are selected sequentially every R data points to generate a subsequence s in the original order. fdi ;

[0075] Reverse downsampling: from s i =(s1,s2,…,s σ Starting from the i-th data point from the end of the sequence, where 1 ≤ i ≤ R, data points are selected in reverse order every R data points, and a subsequence s is generated in reverse order. bdi ;

[0076] As i increases from 1 to R, the forward downsampling and reverse downsampling operations are performed sequentially to obtain s. fd1 ,s fd2 ,…,s fdR With s bd1 ,s bd2 ,…,s bdR ;

[0077] s fd1 ,s fd2 ,…,s fdR Concatenate them in order to form s fd , will s bdR ,…,s bd2 ,s bd1 Concatenate them in order to form s bd ;

[0078] (b) Bidirectional cross mapping:

[0079] i. Input time series s i =(s1,s2,…,s σ The sequence is uniformly divided into R subsequences S1, S2, ..., S... R The length of each subsequence is or This indicates rounding down. This indicates rounding up to the nearest integer, ensuring that the length difference between subsequences does not exceed 1. σ represents the time series sample s. i =(s1,s2,…,s σ ) length,

[0080] ii. Concatenate the subsequences in a forward-interleaving order, such that S1[j], S2[j], ..., S R [j] are concatenated into s in sequence fj S1[j],S2[j],…,S R [j] represents the subsequences S1, S2, ..., Sj respectively. R The j-th value;

[0081] Reverse cross-joining: Interleaving subsequences in reverse order, combining S... R [j],…,S2[j],S1[j] are concatenated in order to form s bj

[0082] As j increases from 1 to By sequentially performing the forward cross-joining and reverse cross-joining operations, s is obtained. f1 ,s f2 ,…,s fu With s b1 ,s b2 ,…,s bu , will s f1 ,s f2 ,…,s fu Concatenate them in order to form s f , will s bu ,…,s b2 ,s b1 Concatenate them in order to form s b ;

[0083] (c) Output a set of sequences {s} that have undergone dilation mapping. fd ,s bd ,s f ,s b}, used for subsequent feature extraction.

[0084] Furthermore, the range of the shuffling rate R is as follows: Where σ is the length of the time series sample, and the maximum value of the shuffling rate R is 16.

[0085] Furthermore, statistical features are extracted from the original sequence, the differential sequence, and the sequence after the original sequence and the differential sequence are dilated and mapped. The statistical features include mean, variance, kurtosis, skewness, and quantile features.

[0086] Furthermore, the classifier in step S4 includes a Ridge Regression Classifier (RidgeCV). i ) and Extremely Random Tree Classifier (ExtraTrees) i ).

[0087] Furthermore, the industrial equipment mentioned in step S1 includes one or more of the following: general equipment, special equipment, or special equipment in the fields of machinery, electrical and electronic, computer networks, automobiles, chemicals, or energy. The equipment operating status information includes one or more of the following: temperature, pressure, speed, torque, sound, vibration, or electrical signals.

[0088] The present invention also provides an industrial equipment fault diagnosis system based on multi-view integration features, characterized in that: the system has a program module corresponding to the above steps, and executes the steps in the above-mentioned industrial equipment fault diagnosis method based on multi-view integration features when running.

[0089] The present invention has the following beneficial technical effects:

[0090] Compared with existing technologies, the "Industrial Equipment Fault Diagnosis Method Based on Multi-View Integrated Features" proposed in this invention has the following significant advantages: 1. Breakthrough improvement in feature representation capability: Through dilatation mapping, the signal is decomposed into multiple scales while preserving the original temporal structure, simultaneously enhancing global trends (such as the overall equipment degradation process) and local details (such as instantaneous fault pulses), making the feature set cover a more comprehensive range of fault modes; this invention introduces a multi-view feature generation mechanism to generate a multi-view feature set with strong complementarity and low redundancy, significantly improving classification performance. 2. Dynamic adaptive feature optimization system: this invention proposes a stability-diversity score-driven feature selector, based on the performance of features on different subsets (stability) and mutual information between features (diversity), dynamically generating robust multi-view feature subsets through adaptive filters and set intersection operations; this invention adopts feature interaction analysis based on Pearson correlation coefficient to quantify the synergistic and conflict relationships between features, avoiding the problems of accidental deletion or redundant accumulation caused by neglecting feature correlation in traditional methods. 3. Performance advantages of heterogeneous integrated classifiers: This invention integrates heterogeneous classifiers such as ridge regression (good at linear patterns) and extreme random trees (capturing nonlinear relationships), and outputs the final fault label through hard voting, which significantly improves the classification robustness under complex working conditions.

[0091] In summary, this invention addresses the key bottlenecks of traditional fault diagnosis methods in industrial fault diagnosis through three core technologies: multi-view feature enhancement, adaptive feature selection, and heterogeneous classifier integration. It provides an innovative solution for high-precision, low-cost equipment health management. This invention balances feature diversity, expressive power, and model robustness while adapting to the stringent constraints of industrial scenarios. Its application significantly improves the accuracy and reliability of industrial equipment fault diagnosis, exhibiting stable performance, especially when processing high-noise, small-sample, or non-stationary operating condition data. Attached Figure Description

[0092] Figure 1 This is a flowchart of an industrial equipment fault diagnosis method based on multi-view integration features in an embodiment of the present invention;

[0093] Figure 2 This is a flowchart illustrating the implementation of the tracking differentiator function based on the optimization algorithm for the filter factor in this embodiment of the invention.

[0094] Figure 3 This is a schematic diagram of the expansion mapping (bidirectional expansion mapping, bidirectional cross mapping) process in an embodiment of the present invention;

[0095] Figure 4 The following are scatter plots comparing the accuracy of the present invention with existing methods in the embodiments of the present invention: (a) is a scatter plot comparing the accuracy of the present invention with the Catch22 method; (b) is a scatter plot comparing the accuracy of the present invention with the Signatures method; (c) is a scatter plot comparing the accuracy of the present invention with the TSFresh method; (d) is a scatter plot comparing the accuracy of the present invention with the TSFel method; (e) is a scatter plot comparing the accuracy of the present invention with the FreshP method; and (d) is a scatter plot comparing the accuracy of the present invention with the TD-MVDC method.

[0096] Figure 5 This is a graph showing the critical difference in accuracy between the method of this invention and other existing methods. Detailed Implementation

[0097] Combined with appendix Figure 1 -Appendix Figure 5 The implementation of the industrial equipment fault diagnosis method based on multi-view integration features described in this invention is explained as follows:

[0098] Specific Implementation Method 1: The flowchart of the industrial equipment fault diagnosis method based on multi-view integration features described in this invention is attached. Figure 1 As shown, the specific steps include:

[0099] Step S1: Obtain a time series dataset of the operational fault status of industrial equipment. The time series dataset contains a sample set S of univariate time series data.L =(s1,s2,…,s n ) and the label set L = (l1, l2, ..., l n The training set Train = (S) consists of... L (L) and a sample set S of univariate time series T =(s n+1 ,s n+2 ,…,s n+m Test set consisting of (S) T ), where n and m are the number of time series samples contained in the training set and the test set, respectively, and the time series samples s i =(s1,s2,…,s σ All are univariate time series with no missing values, equal intervals, and equal length, where 1 ≤ i ≤ n + m, and σ is the length of the time series sample.

[0100] The industrial equipment mentioned includes one or more of the following: general equipment, special equipment, or special equipment in the fields of machinery, electrical and electronic, computer networks, automobiles, chemicals, or energy. The equipment operating status information includes one or more of the following: temperature, pressure, speed, torque, sound, vibration, or electrical signals.

[0101] Step S2: Perform feature extraction and enhancement on the time series dataset, including:

[0102] Step S2.1: Extract statistical features from the original time series sample set S to obtain the feature vector FV. R The original time series sample set S includes the training set and the univariate time series sample set S. L With the test set univariate time series sample set S T ;

[0103] Perform c groups of dilation mappings at different scales on the original time series sample set S to obtain the dilated sample set S. s1 ,…,S sc For the sample set S s1 ,…,S sc Perform statistical feature extraction to obtain the feature vector FV Rs1 ,…,FV Rsc ;

[0104] Step S2.2: Perform a first-order difference transformation on the original time series sample set S to obtain the first-order difference time series sample set S. D ;

[0105] For the first-order difference time series sample set S D Perform statistical feature extraction to obtain the feature vector FV D ;

[0106] For the first-order difference time series sample set S D Perform dilation mapping at different scales in group c to obtain the dilated sample set S. Ds1 ,…,S Dsc For the sample set S Ds1 ,…,S Dsc Perform statistical feature extraction to obtain the feature vector FV Ds1 ,…,FV Dsc ;

[0107] The method involves extracting statistical features from the original sequence, the differential sequence, and the sequence obtained by dilation mapping of the original and differential sequences. The statistical features include mean, variance, kurtosis, skewness, and quantile features.

[0108] The first-order difference transform of the original time series sample set S can be implemented by a tracking differentiator with an adjustable filter factor k. The flowchart of the implementation of the tracking differentiator with an adjustable filter factor is attached. Figure 2 As shown, its specific functional implementation process is as follows:

[0109] The formulas for calculating the first column x1[1] of matrix x1 and the first column x2[1] of matrix x2 are as follows:

[0110]

[0111] Wherein, time series sample s is the time series sample input to the tracking differentiator in the original time series sample set S, s[i] is the i-th value in time series sample s, 1≤i≤σ, where σ is the length of time series sample s;

[0112] For i = 1, 2, ..., σ increases sequentially:

[0113] y i = x1[i] - s[i] + khx2[i];

[0114] If |y i |>k 2 h 2 r, then otherwise,

[0115] If |a|>rhk, then fh i = -r·sign(a), otherwise,

[0116] The formulas for calculating the (i+1)th column x1[i+1] of matrix x1 and the (i+1)th column x2[i+1] of matrix x2 are as follows:

[0117]

[0118] By performing a differential transformation on the univariate time series using a tracking differentiator with an adjustable filter factor, the resulting differential time series is: Wherein, r is a preset speed factor parameter, h is a preset step size parameter, and k is an adjustable filter factor;

[0119] The adjustable filter factor k of the tracking differentiator is iteratively optimized by an iterative optimization algorithm, and the iterative optimization process is as follows:

[0120] (a) Setting optimization parameters: number of candidate solutions nc = 10, maximum number of iterations inter max =10. Filter factor search range [b L ,b R ], where the search range is [b L ,b R The default value for ] is [1,3];

[0121] (b) Define the objective function f(k):

[0122] i. For the training set time series S L Generate a set of differential sequences using the filter factor k

[0123] ii. Regarding The following sampling rates are sorted in ascending order by row. rate =4 Extracting quantile features

[0124] iii. Calculate the feature set The contour score (S-score) of label L is used as the objective function value f(k), and the filter factor k is iteratively optimized by maximizing the contour score (S-score).

[0125] The contour score (S-score) is defined as follows: Where n is the number of time series samples contained in the training set, and SC(i) represents the feature set. The silhouette coefficient of the feature vector corresponding to the i-th time series sample. Where η(i) represents the minimum distance between clusters, i.e., the feature set The minimum Euclidean distance between the feature vector corresponding to the i-th time series sample and the feature vectors corresponding to all time series samples with different labels is defined as follows: The average distance within a cluster, i.e., the feature set. The average Euclidean distance between the feature vector corresponding to the i-th time series sample and the feature vectors corresponding to all time series samples with the same label;

[0126] (c) Iterative optimization:

[0127] i. Initialize the candidate solution set {k1,k2,…,k nc}∈[b L ,b R ];

[0128] ii. According to k i The objective function f(k) takes the value f(k) i Sort the candidate solutions from largest to smallest, 1≤i≤nc, and select the k that makes the objective function f(k) the largest, second largest, and third largest respectively. i As the optimal solution k α suboptimal solution k β and the third optimal solution k δ ;

[0129] iii. Update convergence coefficients and random vectors Where t is the current iteration number. A vector whose modulus takes random values ​​between [0,1]. It is a unit vector;

[0130] iv. Update candidate solution positions:

[0131]

[0132] in Representative vector The i-th term, k(t-1) is the value of the adjustable filter factor at iteration number t-1, k(t) is the value of the adjustable filter factor at iteration number t, k(0) = k α ;

[0133] v. Repeat the iteration until the maximum number of iterations is reached. max The optimal solution k of the adjustable filter factor of the output tracking differentiator adj =k(inter max );

[0134] The schematic diagrams of the dilation mapping in steps S2.1 and S2.2 are attached. Figure 3 As shown, the specific implementation steps are as follows:

[0135] (a) Bidirectional expansion mapping:

[0136] For the input time series s i =(s1,s2,…,s σ Global mode enhancement is performed based on the shuffle rate R, where R is an integer greater than 1. Specific operations include:

[0137] Forward downsampling: from s i =(s1,s2,…,s σ Starting from the i-th data point, 1≤i≤R, data points are selected sequentially every R data points to generate a subsequence s in the original order. fdi ;

[0138] Reverse downsampling: from s i =(s1,s2,…,s σ Starting from the i-th data point from the end of the sequence, where 1 ≤ i ≤ R, data points are selected in reverse order every R data points, and a subsequence s is generated in reverse order. bdi ;

[0139] As i increases from 1 to R, the forward downsampling and reverse downsampling operations are performed sequentially to obtain s. fd1 ,s fd2 ,…,s fdR With s bd1 ,s bd2 ,…,s bdR ;

[0140] s fd1 ,s fd2 ,…,s fdR Concatenate them in order to form s fd , will s bdR ,…,s bd2 ,s bd1 Concatenate them in order to form s bd ;

[0141] (b) Bidirectional cross mapping:

[0142] i. Input time series s i =(s1,s2,…,s σ The sequence is uniformly divided into R subsequences S1, S2, ..., S... R The length of each subsequence is or This indicates rounding down. This indicates rounding up to the nearest integer, ensuring that the length difference between subsequences does not exceed 1. σ represents the time series sample s. i =(s1,s2,…,s σ ) length,

[0143] ii. Concatenate the subsequences in a forward-interleaving order, such that S1[j], S2[j], ..., S R [j] are concatenated into s in sequence fj S1[j],S2[j],…,S R [j] represents the subsequences S1, S2, ..., Sj respectively. RThe j-th value;

[0144] Reverse cross-joining: Interleaving subsequences in reverse order, combining S... R [j],…,S2[j],S1[j] are concatenated in order to form s bj

[0145] As j increases from 1 to By sequentially performing the forward cross-joining and reverse cross-joining operations, s is obtained. f1 ,s f2 ,…,s fu With s b1 ,s b2 ,…,s bu , will s f1 ,s f2 ,…,s fu Concatenate them in order to form s f , will s bu ,…,s b2 ,s b1 Concatenate them in order to form s b ;

[0146] (c) Output a set of sequences {s} that have undergone dilation mapping. fd ,s bd ,s f ,s b}, used for subsequent feature extraction;

[0147] The range of the shuffling rate R is as follows: Where σ is the length of the time series sample, and the maximum value of the shuffling rate R is 16;

[0148] Step S2.3: Convert the feature vector FV R FV Rs1 ,…,FV Rsc FV D FV Ds1 ,…,FV Dsc Concatenate them to form a high-dimensional initial feature vector set FV;

[0149] Step S3: Perform multi-view integrated feature selection on the initial feature vector set FV, including:

[0150] Step S3.1: Perform multi-metric scoring and generate initial view on the high-dimensional feature set:

[0151] M independent feature evaluation metrics (at least two independent feature evaluation metrics, such as mutual information (MI) and analysis of variance (ANOVA)) are used to score each feature in the initial feature vector set FV. Each feature evaluation metric generates a corresponding score matrix, resulting in M ​​score matrices.

[0152] Based on the scoring matrix corresponding to the i-th feature evaluation index, sort each feature in descending order and select the top K1, K2, ..., K1 features. T Given several high-scoring features, we obtain the initial view feature set FV corresponding to the i-th feature evaluation index. 1i ,FV 2i ,…,FV Ti , where i is a positive integer less than M;

[0153] The i increases sequentially from 1 to M, where i = 1, 2, ..., M, thus obtaining the initial view feature set corresponding to the first feature evaluation index. The initial view feature set FV corresponding to the second type of feature evaluation index 12 ,FV 22 ,…,FV T2 ..., the initial view feature set corresponding to the Mth feature evaluation index

[0154] The number of features K1, K2, ..., K T Based on the preset convergence threshold list thresholds = [threshold1, threshold2, ..., threshold] T Determine the number of candidate features K (K = 1, 2, ..., N). feature ), calculate using formula Select in sequence those that satisfy rate(K)≤threshold i The smallest eigenvalue is K i , 1≤i≤T, where rate(K) is the relative rate of change of the stability-diversity score (SD-score), N feature The number of features in the initial feature vector set FV;

[0155] For each candidate feature quantity K, based on the scoring matrix corresponding to each feature evaluation index j, the top K high-scoring features are selected from the initial feature vector set FV to obtain a feature subset. The j increases sequentially from 1 to M, where j = 1, 2, ..., M, thus obtaining the feature subsets corresponding to all feature evaluation indicators. Calculate the feature subset corresponding to the number of candidate features K. Pearson correlation coefficient absolute value matrix Each sub-block of the absolute value matrix A(K) of the Pearson correlation coefficient All are K×K dimensional square matrices. Representing a feature subset With feature subset The absolute value matrix of the Pearson correlation coefficients between them 1≤p≤M, 1≤q≤M;

[0156] The stability-diversity score (SD-score) corresponding to the number of candidate features K is expressed by the formula. Calculation, where Indicates a pair of child blocks Summing all elements;

[0157] The feature evaluation metrics include mutual information (MI) and analysis of variance (ANOVA);

[0158] Step S3.2: Generate the intersection-optimized extended view feature set:

[0159] For the initial view feature set FV corresponding to the first type of feature evaluation index 11 ,FV 21 ,…,FV T1 The initial view feature set FV corresponding to the second feature evaluation index 12 ,FV 22 ,…,FV T2 ..., the initial view feature set FV corresponding to the Mth feature evaluation index 1M ,FV 2M ,…,FV TM For each of the initial view feature sets corresponding to different feature evaluation indicators, the intersection of feature indices is taken to generate T common feature subsets. in

[0160] The common feature subset Concatenate into a common feature vector FV interse and the common feature vector FV intersection Initial view feature sets FV corresponding to all feature evaluation metrics 11 ,FV 21 ,…,FV T1 FV 12 ,FV 22 ,…,FV T2 , ..., FV 1M ,FV 2M ,…,FV TM The features are stitched together to generate N = T·M extended view feature sets FV'. V1 ,FV' V2 ,…,FV' VN ;

[0161] Step S3.3: Redundancy removal and output of the final multi-view feature set:

[0162] For each of the extended view feature sets FV' V1,FV' V2 ,…,FV' VN Pearson correlation coefficient analysis was performed on the features, and redundant features with correlation coefficients greater than a preset threshold were removed to obtain the final multi-view feature set FV'1, FV'2, ..., FV' N ;

[0163] Step S4, Heterogeneous Integration Classification and Diagnosis:

[0164] Classifier training: using the training set Train = (S L L) The training set multi-view feature set is obtained through steps S2 and S3. Feature set in each view The classifier CL is trained on the above. i , 1≤i≤N;

[0165] The classifiers include Ridge Regression Classifier (RidgeCV). i ) and Extremely Random Tree Classifier (ExtraTrees) i );

[0166] Hard voting integration: using the test set Test = (S T The test set multi-view feature set is obtained through steps S2 and S3. Feature set of each view Input classifier CL i The multi-view prediction result L is obtained. i , 1≤i≤N;

[0167] The classification results L1, L2, L, L are integrated from all views using a majority voting rule. N Output the final fault label L T ;

[0168] Step S5, Result Output: Output the diagnostic label L T The data is transmitted to the industrial equipment control terminal via the communication module, triggering fault warnings or maintenance commands.

[0169] Specific Implementation Method Two: The present invention also provides an industrial equipment fault diagnosis system based on multi-view integration features, characterized in that: the system has a program module corresponding to the above steps, and executes the steps in the above-mentioned industrial equipment fault diagnosis method based on multi-view integration features when running.

[0170] Specifically, it includes the following modules:

[0171] (a) Data acquisition and preprocessing module: used to acquire univariate time series data of the operating status of industrial equipment. The data includes a training set and a test set. The training set consists of standardized time series samples with equal intervals, equal lengths and no missing values ​​and their corresponding discrete fault labels. The test set consists of unlabeled standardized time series samples.

[0172] The module is also used to standardize the time series data so that its mean is 0 and its variance is 1;

[0173] (b) Multi-view feature extraction module, including:

[0174] i. Adaptive Parameter Optimization Unit: Based on a swarm intelligence optimization algorithm, using the contour score as the objective function, it dynamically adjusts the filter factor k of the tracking differentiator. adj Generate a set of differential sequences;

[0175] ii. Sequence Transformation Unit: Performs bidirectional dilation mapping and bidirectional cross mapping on the original sequence and the differential sequence, wherein: bidirectional dilation mapping generates a global trend enhancement sequence through forward downsampling and backward downsampling, and bidirectional cross mapping generates a local detail enhancement sequence by uniformly dividing the sequence and splicing it in forward and reverse order;

[0176] iii. Statistical feature extraction unit: Using a pre-set statistical feature library, the unit extracts mean, variance, and quantile features from the original sequence, differential sequence, and transformed sequence, and concatenates them into a high-dimensional initial feature vector set;

[0177] (c) Adaptive multi-view selection module, including:

[0178] i. Multi-metric scoring unit: Features are scored independently using feature evaluation metrics such as mutual information (MI) and analysis of variance (ANOVA) to generate a scoring matrix;

[0179] ii. Intersection Optimization Unit: Adaptively determine the number of features based on the Stability-Diversity Score (SD-score), generate a common feature subset by taking the intersection of the indices of features under each metric, and combine it with the initial view feature set to output a multi-view feature set;

[0180] (d) Heterogeneous integrated classification module, including:

[0181] i. Classifier training unit: Train a Ridge Regression Classifier (RidgeCV) and an Extra Random Tree Classifier on each view feature set in the training set, respectively;

[0182] ii. Hard Voting Integration Unit: Input the multi-view features of the test set into the trained ridge regression classifier and extreme random tree classifier respectively to obtain the multi-view prediction results, and perform majority voting (hard voting) to output the final fault diagnosis label;

[0183] (e) Results Interaction Module:

[0184] Used to transmit diagnostic tags to the control terminal of industrial equipment to trigger fault warnings or maintenance commands.

[0185] Verification has shown that the method proposed in this invention solves the technical problem raised in this invention, and practical application has verified the technical effects and practicality claimed in this invention.

[0186] The algorithm (method) proposed in this invention is the underlying technical core of this invention, and various products can be derived based on the algorithm.

[0187] Based on the algorithm (method) proposed in this invention, an industrial equipment fault diagnosis system based on multi-view integration features is developed using a programming language. This system has program modules corresponding to the steps of the above technical solution, and executes the steps in the above-mentioned industrial equipment fault diagnosis method based on multi-view integration features when running.

[0188] The developed system (software) computer program is stored on a computer-readable storage medium, and the computer program is configured to implement the steps of the above-described industrial equipment fault diagnosis method based on multi-view integration features when called by a processor. In other words, the invention is materialized on a carrier, becoming a computer program product.

[0189] An industrial equipment fault diagnosis device based on multi-view integration features is disclosed. The device includes at least one processor and a memory communicatively connected to the at least one processor. The memory stores instructions executable by the at least one processor. The instructions are executed by the at least one processor to enable the at least one processor to perform the aforementioned industrial equipment fault diagnosis method based on multi-view integration features, thereby realizing fault diagnosis of industrial equipment.

[0190] Various implementations of the systems and techniques described herein can be implemented in digital electronic circuit systems, integrated circuit systems, application-specific integrated circuits (ASICs), computer hardware, firmware, software, and / or combinations thereof. These various implementations may include: implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transferring data and instructions to the storage system, the at least one input device, and the at least one output device.

[0191] The computational programs (also referred to as programs, software, software applications, or code) of this invention include machine instructions of a programmable processor and can be implemented using high-level procedural and / or object-oriented programming languages, and / or assembly / machine languages. As used herein, the terms "machine-readable medium" and "computer-readable medium" refer to any computer program product, device, and / or apparatus (e.g., disk, optical disk, memory, programmable logic device PLD) for providing machine instructions and / or data to a programmable processor, including machine-readable media that receive machine instructions as machine-readable signals. The term "machine-readable signal" refers to any signal for providing machine instructions and / or data to a programmable processor.

[0192] Example

[0193] In this embodiment, existing methods compared to the method of the present invention include:

[0194] (1) TSFresh Pipeline with Rotation Forest Classifier (abbreviated as FreshP), the method of which is derived from the reference M. Middlehurst, A. Bagnall, “The FreshPRINCE: A Simple Transformation Based Pipeline Time Series Classifier,” in Proceedings of International Conference on Pattern Recognition and Artificial Intelligence 2022 (ICPRAI 2022), pp.150-161, May, 2022;

[0195] (2) Feature extraction algorithm based on 22 typical time series characteristics (22 Canonical Time-series Characteristics, abbreviated as: catch22), the method is derived from the reference CHLubba, SS Sethi, P. Knaute, et al., “catch22: CAnonical Time-series CHaracteristics Selected through highly comparative time-series analysis,” Data Mining Knowledge Discovery, vol.33, no.6, pp.1821–1852, Nov.2019;

[0196] (3) Time series feature extraction on basis of scalable hypothesis tests method (abbreviated as TSFresh), the method is derived from the reference M. Christ, N. Braun, J. Neuffer, and AWKempa-Liehr, “Time Series Feature Extraction on basis of Scalable Hypothesis tests (tsfresh–A Python package),” Neurocomputing, vol. 307, pp. 72–77, Sep. 2018;

[0197] (4) A generalized signature method for multivariate time series feature extraction (hereinafter referred to as Signatures). This method is derived from the reference published on arXiv, J. Morrill, A. Fermanian, P. Kidger, et al., “A generalized signature method for multivariate time series feature extraction,” arXiv, 2020. https: / / doi.org / 10.48550 / arXiv.2006.00873 ;

[0198] (5) A time series feature extraction method (TSFel), which is derived from the reference M. Barandas, D. Folgado, L. Fernandes, et al., “Tsfel: Time series feature extraction library,” SoftwareX, vol.11, pp.100456, Jan.2020;

[0199] (6) A fault diagnosis method for industrial equipment based on multi-view dilatation statistical features (abbreviated as: TD-MVDC) is disclosed in Chinese invention patent CN118644702A.

[0200] To verify the effectiveness of the method of this invention, a computer simulation classification comparison experiment was conducted on 112 datasets of the UCR open-source dataset, comparing the method of this invention with existing methods. Each dataset used a default training / test set split, and each sample was standardized. Classification accuracy was used as the performance metric for evaluation, and the significance of differences between different algorithms was measured using a Wilcoxon signed-rank test with Holm correction and a p-value of 0.02. The classification accuracy results are shown in Table 1, where the data unit is %.

[0201] Table 1

[0202]

[0203]

[0204]

[0205]

[0206]

[0207] The scatter plot showing the accuracy of the method of this invention compared with existing methods on 112 UCR datasets is shown below. Figure 4 As shown, points falling above / above / below the diagonal represent the accuracy of the method of the present invention on the dataset that is inferior to / equal to / superior to the accuracy of the prior art method, respectively. Here, W represents the number of datasets where the classification accuracy of the method is superior to the comparison method, T represents the number of datasets where the classification accuracy of the method is equal to the comparison method, and L represents the number of datasets where the classification accuracy of the method is inferior to the comparison method.

[0208] From Table 1 and Figure 4The results show that the method of the present invention not only outperforms existing methods in accuracy on most datasets, but also significantly outperforms them in accuracy on some datasets.

[0209] Figure 5 The graph shows the critical difference in accuracy between the method of this invention and existing methods. Methods with no statistically significant differences are connected by thick black lines. The numbers represent the average ranking of the method of this invention and existing methods on 112 UCR datasets. The method of this invention achieves the highest average ranking on the accuracy metric, demonstrating its superiority.

[0210] It should be understood that the various forms of processes shown above can be used to rearrange, add, or delete steps. For example, the steps described in this application can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution disclosed in this application can be achieved, they are all within the protection scope of this invention.

Claims

1. A multi-view ensemble feature based industrial equipment fault diagnosis method, characterized in that, Includes the following steps: Step S1: Obtain a time series dataset of the operational fault status of industrial equipment. The time series dataset contains a sample set S of univariate time series data. L =(s1,s2,…,s n ) and the label set L = (l1, l2, ..., l n The training set Train = (S) consists of... L (L) and a sample set S of univariate time series T =(s n+1 ,s n+2 ,…,s n+m Test set consisting of (S) T ), where n and m are the number of time series samples contained in the training set and the test set, respectively, and the time series samples s i =(s1,s2,…,s σ All are univariate time series with no missing values, equal intervals, and equal lengths, where 1 ≤ i ≤ n + m, and σ is the length of the time series sample. Step S2: Perform feature extraction and enhancement on the time series dataset, including: Step S2.1: Extract statistical features from the original time series sample set S to obtain the feature vector FV. R The original time series sample set S includes the training set and the univariate time series sample set S. L With the test set univariate time series sample set S T ; Perform c different scale dilation mappings on the original time series sample set S to obtain the dilated sample set S. s1 ,…,S sc For the sample set S s1 ,…,S sc Perform statistical feature extraction to obtain the feature vector FV Rs1 ,…,FV Rsc ; Step S2.2: Perform a first-order difference transformation on the original time series sample set S to obtain the first-order difference time series sample set S. D ; For the first-order difference time series sample set S D Perform statistical feature extraction to obtain the feature vector FV D ; For the first-order difference time series sample set S D Perform dilation mapping at different scales in group c to obtain the dilated sample set S. Ds1 ,…,S Dsc For the sample set S Ds1 ,…,S Dsc Perform statistical feature extraction to obtain the feature vector FV Ds1 ,…,FV Dsc ; Step S2.3: Convert the feature vector FV R FV Rs1 ,…,FV Rsc FV D FV Ds1 ,…,FV Dsc Concatenate them to form a high-dimensional initial feature vector set FV; Step S3: Perform multi-view integrated feature selection on the initial feature vector set FV, including: Step S3.1: Perform multi-metric scoring and generate initial view on the high-dimensional feature set: M independent feature evaluation indicators are used to score each feature in the initial feature vector set FV. Each feature evaluation indicator generates a corresponding score matrix, resulting in M ​​score matrices. Based on the scoring matrix corresponding to the i-th feature evaluation index, sort each feature in descending order and select the top K1, K2, ..., K1 features. T Given several high-scoring features, we obtain the initial view feature set FV corresponding to the i-th feature evaluation index. 1i ,FV 2i ,…,FV Ti , where i is a positive integer less than M; The i increases sequentially from 1 to M, where i = 1, 2, ..., M, to obtain the initial view feature set FV corresponding to the first feature evaluation index. 11 ,FV 21 ,…,FV T1 The initial view feature set FV corresponding to the second feature evaluation index 12 ,FV 22 ,…,FV T2 ..., the initial view feature set FV corresponding to the Mth feature evaluation index 1M ,FV 2M ,…,FV TM ; The number of features K1, K2, ..., K T Based on the preset convergence threshold list thresholds = [threshold1, threshold2, ..., threshold] T Determine the number of candidate features K (K = 1, 2, ..., N). feature ), calculate using formula Select in sequence those that satisfy rate(K)≤threshold i The smallest eigenvalue is K i , 1≤i≤T, where rate(K) is the relative rate of change of the stability-diversity score (SD-score), N feature The number of features in the initial feature vector set FV; For each candidate feature quantity K, based on the scoring matrix corresponding to each feature evaluation index j, the top K high-scoring features are selected from the initial feature vector set FV to obtain a feature subset. The j increases sequentially from 1 to M, where j = 1, 2, ..., M, thus obtaining the feature subsets corresponding to all feature evaluation indicators. Calculate the feature subset corresponding to the number of candidate features K. Pearson correlation coefficient absolute value matrix Each sub-block of the absolute value matrix A(K) of the Pearson correlation coefficient All are K×K dimensional square matrices. Representing a feature subset With feature subset The absolute value matrix of the Pearson correlation coefficients between them The stability-diversity score (SD-score) corresponding to the number of candidate features K is expressed by the formula. Calculation, where Indicates a pair of child blocks Summing all elements; Step S3.2: Generate the intersection-optimized extended view feature set: For the initial view feature set FV corresponding to the first type of feature evaluation index 11 ,FV 21 ,…,FV T1 The initial view feature set FV corresponding to the second feature evaluation index 12 ,FV 22 ,…,FV T2 ..., the initial view feature set FV corresponding to the Mth feature evaluation index 1M ,FV 2M ,…,FV TM For each of the initial view feature sets corresponding to different feature evaluation indicators, the intersection of feature indices is taken to generate T common feature subsets. in The common feature subset Concatenate into a common feature vector FV intersection and the common feature vector FV intersection Initial view feature sets FV corresponding to all feature evaluation metrics 11 ,FV 21 ,…,FV T1 FV 12 ,FV 22 ,…,FV T2 , ..., FV 1M ,FV 2M ,…,FV TM The features are stitched together to generate N = T·M extended view feature sets FV'. V1 ,FV' V2 ,…,FV' VN ; Step S3.3: Redundancy removal and output of the final multi-view feature set: For each of the extended view feature sets FV' V1 ,FV' V2 ,…,FV' VN Pearson correlation coefficient analysis was performed on the features, and redundant features with correlation coefficients greater than a preset threshold were removed to obtain the final multi-view feature set FV'1, FV'2, ..., FV' N ; Step S4, Heterogeneous Integration Classification and Diagnosis: Classifier training: using the training set Train = (S L L) The training set multi-view feature set is obtained through steps S2 and S3. Feature set FV in each view i L The classifier CL is trained on the above. i , 1≤i≤N; Hard voting integration: using the test set Test = (S T The test set multi-view feature set is obtained through steps S2 and S3. The feature set FV of each view i T Input classifier CL i The multi-view prediction result L is obtained. i , 1≤i≤N; The classification results L1, L2,..., L of all views are integrated using majority voting rule N , and the final fault label L is output T ; Step S5, result output: the diagnostic label L T Transmitted to the industrial equipment control terminal through the communication module, triggering the fault early warning or maintenance instruction. 2.The industrial equipment fault diagnosis method based on multi-view integrated features according to claim 1, characterized in that, In step S2.2, the first-order difference transform of the original time series sample set S can be implemented by a tracking differentiator with an adjustable filter factor k. The function of the tracking differentiator with an adjustable filter factor is as follows: The formulas for calculating the first column x1[1] of matrix x1 and the first column x2[1] of matrix x2 are as follows: Wherein, time series sample s is the time series sample input to the tracking differentiator in the original time series sample set S, s[i] is the i-th value in time series sample s, 1≤i≤σ, where σ is the length of time series sample s; For i = 1, 2, ..., σ increases sequentially: y i =x1[i]-s[i]+xx2[i]; if |y i | > k 2 h 2 r, then else, If |a| > rhk, then fh i = -r • sign(a), otherwise, The formulas for calculating the (i+1)th column x1[i+1] of matrix x1 and the (i+1)th column x2[i+1] of matrix x2 are as follows: By performing a differential transformation on the univariate time series using a tracking differentiator with an adjustable filter factor, the resulting differential time series is: Wherein, r is a preset speed factor parameter, h is a preset step size parameter, and k is an adjustable filter factor. 3.The industrial equipment fault diagnosis method based on multi-view integrated features according to claim 2, characterized in that, The iterative optimization process of the adjustable filter factor k of the tracking differentiator is as follows: (a) Setting optimization parameters: number of candidate solutions nc = 10, maximum number of iterations inter max =10. Filter factor search range [b L ,b R ], where the search range is [b L ,b R The default value for ] is [1,3]; (b) Define the objective function f(k): i. For the training set time series S L Generate a set of differential sequences using the filter factor k ii. to Following the ascending order of rows, the following down-sampling rate d rate = 4 extract quantile features iii. Computing the feature set The filter factor k is iteratively optimized by maximizing the silhouette score (S-score) with the silhouette score (S-score) of the labels L as the objective function value f(k). The contour score (S-score) is defined as follows: Where n is the number of time series samples contained in the training set, and SC(i) represents the feature set. The silhouette coefficient of the feature vector corresponding to the i-th time series sample. Where η(i) represents the minimum distance between clusters, i.e., the feature set The minimum Euclidean distance between the feature vector corresponding to the i-th time series sample and the feature vectors corresponding to all time series samples with different labels, where θ(i) represents the average distance within the cluster, i.e., the feature set. The average Euclidean distance between the feature vector corresponding to the i-th time series sample and the feature vectors corresponding to all time series samples with the same label; (c) Iterative optimization: i. Initialize the candidate solution set {k1,k2,…,k nc }∈[b L ,b R ]; ii. According to k i The objective function f(k) takes the value f(k) i Sort the candidate solutions from largest to smallest, 1≤i≤nc, and select the k that makes the objective function f(k) the largest, second largest, and third largest respectively. i As the optimal solution k α suboptimal solution k β and the third optimal solution k δ ; iii. Update convergence coefficients and random vectors Where t is the current iteration number. A vector whose modulus takes random values ​​between [0,1]. It is a unit vector; iv. Update the position of candidate solutions: in Representative vector The i-th term, k(t-1) is the value of the adjustable filter factor at iteration number t-1, k(t) is the value of the adjustable filter factor at iteration number t, k(0) = k α ; v. Repeat the iteration until the maximum number of iterations is reached. max The optimal solution k of the adjustable filter factor of the output tracking differentiator adj =k(inter max ). 4.The industrial equipment fault diagnosis method based on multi-view integrated features according to claim 3, characterized in that, The feature evaluation metrics in step S3.1 include mutual information (MI) and analysis of variance (ANOVA). 5.The industrial equipment fault diagnosis method based on multi-view integrated features according to claim 4, characterized in that, The specific implementation steps of the dilation mapping in steps S2.1 and S2.2 are as follows: (a) Bidirectional expansion mapping: For the input time series s i =(s1,s2,…,s σ Global mode enhancement is performed based on the shuffle rate R, where R is an integer greater than 1. Specific operations include: Forward downsampling: from s i =(s1,s2,…,s σ Starting from the i-th data point, 1≤i≤R, data points are selected sequentially every R data points to generate a subsequence s in the original order. fdi ; Reverse downsampling: from s i =(s1,s2,…,s σ Starting from the i-th data point from the end of the sequence, where 1 ≤ i ≤ R, data points are selected in reverse order every R data points, and a subsequence s is generated in reverse order. bdi ; As i increases from 1 to R, the forward downsampling and reverse downsampling operations are performed sequentially to obtain s. fd1 ,s fd2 ,…,s fdR With s bd1 ,s bd2 ,…,s bdR ; s fd1 ,s fd2 ,…,s fdR Concatenate them in order to form s fd , will s bdR ,…,s bd2 ,s bd1 Concatenate them in order to form s bd ; (b) Bidirectional cross mapping: i. Input time series s i =(s1,s2,…,s σ The sequence is uniformly divided into R subsequences S1, S2, ..., S... R The length of each subsequence is or This indicates rounding down. This indicates rounding up to the nearest integer, ensuring that the length difference between subsequences does not exceed 1. σ represents the time series sample s. i =(s1,s2,…,s σ ) length, ii. Concatenate the subsequences in a forward-interleaving order, such that S1[j], S2[j], ..., S R [j] are concatenated into s in sequence fj S1[j],S2[j],…,S R [j] represents the subsequences S1, S2, ..., Sj respectively. R The j-th value; Reverse cross-joining: Interleaving subsequences in reverse order, combining S... R [j],…,S2[j],S1[j] are concatenated in order to form s bj As j increases from 1 to By sequentially performing the forward cross-joining and reverse cross-joining operations, s is obtained. f1 ,s f2 ,…,s fu With s b1 ,s b2 ,…,s bu , will s f1 ,s f2 ,…,s fu Concatenate them in order to form s f , will s bu ,…,s b2 ,s b1 Concatenate them in order to form s b ; (c) output 1 set of expanded mapped sequence sets {s fd ,s bd ,s f ,s b} for subsequent feature extraction. 6.The industrial equipment fault diagnosis method based on multi-view integrated features according to claim 5, characterized in that, The range of the shuffling rate R is as follows: Where σ is the length of the time series sample, and the maximum value of the shuffling rate R is 16.

7. The industrial equipment fault diagnosis method based on multi-view integrated features according to claim 6, characterized in that, Statistical features are extracted from the original sequence, the differential sequence, and the sequence after the original sequence and the differential sequence are dilated. The statistical features include mean, variance, kurtosis, skewness, and quantile features.

8. The industrial equipment fault diagnosis method based on multi-view integrated features according to claim 7, characterized in that, The classifier in the step S4 comprises a Ridge regression classifier (RidgeCV i ) and an Extra Trees classifier (ExtraTrees i ). 9.The industrial equipment fault diagnosis method based on multi-view integrated features according to claim 8, characterized in that, The industrial equipment mentioned in step S1 includes one or more of the following: general equipment, special equipment, or special equipment in the fields of machinery, electrical and electronic, computer networks, automobiles, chemicals, or energy. The equipment operating status information includes one or more of the following: temperature, pressure, speed, torque, sound, vibration, or electrical signals.

10. A multi-view integrated feature based industrial equipment fault diagnosis system, characterized in that: The system has a program module corresponding to the steps of any one of the claims 1-9 above, and executes the steps in the above-described method for diagnosing industrial equipment faults based on multi-view integration features when it is run.

Images