A semi-supervised multi-view svdd industrial equipment fault detection method based on dictionary learning

Abstract

The application relates to a semi-supervised multi-view SVDD industrial equipment fault detection method based on dictionary learning, comprising the following steps: collecting multi-source sensor vibration signals of an industrial rotating machine bearing and carrying out pretreatment; based on the pretreated signals, a semi-supervised multi-view double-dictionary SVDD model is constructed; the semi-supervised multi-view double-dictionary SVDD model is initialized; based on the initialized model, global iterative optimization is carried out, and view weight smoothing update is carried out; according to the weight update result, multi-view weighted fusion abnormal decision is carried out, and the final classification of samples is obtained. The application can simultaneously utilize labeled normal samples, a small amount of labeled abnormal samples and a large amount of unlabeled samples, adaptively fuse multi-view information, and realize high-precision and high-robustness bearing fault detection.

Description

Technical Field

[0001] This invention relates to the fields of industrial equipment fault detection and machine learning technology, and in particular to a semi-supervised multi-view SVDD method for industrial equipment fault detection based on dictionary learning. Background Technology

[0002] Rotating machinery is one of the most critical types of equipment in industrial production, widely used in industries such as power, petroleum, chemical, and metallurgy. Bearings, as key components of rotating machinery, directly determine the safety and reliability of the entire equipment through their operating condition. Statistics show that approximately 30% of rotating machinery failures are caused by bearing failures. Therefore, early detection of bearing failures is of paramount importance for ensuring production safety, reducing maintenance costs, and preventing major accidents.

[0003] Currently, data-driven fault detection methods have become the mainstream research direction, but existing technologies still suffer from the following core shortcomings: 1. Severe label scarcity: In industrial scenarios, equipment is in normal operation most of the time, with sufficient and easily obtainable normal samples; however, the number of fault samples is extremely small, and labeling requires professional personnel and significant time costs. Pure supervised learning methods require a large number of labeled samples to train an effective model, making them unsuitable for real-world industrial scenarios; traditional single-class SVDD methods can only utilize normal samples for modeling, completely wasting a small number of valuable labeled abnormal samples and a large number of unlabeled samples, resulting in limited detection accuracy. 2. Insufficient utilization of multi-view information: Industrial equipment is typically equipped with multiple sensors that can collect vibration signals from different locations such as the drive end, fan end, and base. These signals reflect the operating status of the equipment from different angles and have strong complementarity. However, most existing multi-view fault detection methods adopt simple averaging or hard fusion strategies, which cannot adaptively distinguish the importance of different views. Interference from weak or noisy views can severely dilute the effective information of strong views, leading to a decrease in model robustness. 3. Limitations of Dictionary Learning Methods: Dictionary-based fault detection methods extract features by learning sparse representations of data, offering good interpretability. However, most existing methods only use comprehensive dictionaries for feature reconstruction, requiring complex optimization problems during the encoding process, resulting in low computational efficiency and high model complexity. For example, Chinese patent CN111027594A uses a stacked normal and abnormal dictionary structure for step-by-step detection, but this method requires a large number of abnormal samples to construct an effective abnormal dictionary. However, abnormal samples are scarce in industrial scenarios, failing to meet its training requirements. Furthermore, this method is a single-view approach and cannot utilize complementary information from multiple sensor sources. 4. Coarse Processing of Unlabeled Samples: Existing semi-supervised anomaly detection methods typically employ fixed category assumptions for unlabeled samples (e.g., assuming all are normal), lacking dynamic verification and correction mechanisms. When the unlabeled samples contain many abnormal samples, it leads to severe model boundary shifts and a significant decrease in detection accuracy. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of the prior art and provide a semi-supervised multi-view SVDD industrial equipment fault detection method based on dictionary learning, which solves the problems of scarce labels, insufficient utilization of multi-view information, and coarse processing of unlabeled samples in industrial scenarios.

[0005] To achieve the above objectives, the present invention provides the following solution: A semi-supervised multi-view SVDD method for industrial equipment fault detection based on dictionary learning includes: Vibration signals from multiple sensors on industrial rotating machinery bearings are collected and preprocessed. Based on the preprocessed signal, a semi-supervised multi-view dual-dictionary SVDD model is constructed; wherein, the semi-supervised multi-view dual-dictionary SVDD model is configured with an objective function and outlier sample constraints; Initialize the semi-supervised multi-view dual-dictionary SVDD model; Based on the initialized model, global iterative optimization is performed. Hypothesis-validation-correction is carried out on unlabeled samples during the iteration, and the view weights are smoothly updated through softmax. Based on the weight update results, perform multi-view weighted fusion anomaly decision-making to obtain the final classifier; The final classifier is used to detect faults in industrial equipment.

[0006] Optionally, multi-source sensor vibration signals from industrial rotating machinery bearings are collected and preprocessed, including: Vibration signals from multiple sources of industrial rotating machinery bearings were collected. The raw signals underwent non-overlapping segmentation and wavelet denoising processing. Multi-view features of each sample were extracted to construct training and testing sets. The training set was divided into a set of labeled normal samples. Labeled abnormal sample set and unlabeled sample set The multi-view features include: time-domain statistical features, frequency-domain amplitude features, and wavelet packet time-frequency domain features.

[0007] Optionally, the semi-supervised multi-view dual-dictionary SVDD model includes: The objective function includes: reconstruction error term, sparse regularization term, encoding consistency term, SVDD hypersphere constraint term, and cross-view constraint term. Figure 1 The consistency term and view weight term, the expression of the objective function is: in, Weights for each view, and The first Comprehensive and analytical dictionaries for each view. These are the sparse coding matrices for the corresponding samples. The first A matrix of labeled normal, labeled abnormal, and unlabeled samples for each view. For the first The radius of the SVDD hypersphere in each view As slack variables, for Norm, V For the number of views, For sparse regularization weights, For coding consistency weights, The penalty coefficient for slack variables, F for F Norm, This represents the number of labeled positive samples. The number of labeled negative samples. This represents the number of unlabeled samples. i For sample index, For cross-view Figure 1 Consistency weight, For the first View indexes, For the first One view index; Constraints: Normal sample constraints: Abnormal sample constraints: Initial constraints for unlabeled samples: Non-negativity of slack variables: Dictionary atomic constraints: View weight constraints: in, For the first The center of the SVDD hypersphere in one view For the comprehensive dictionary One atom, For the labeled positive samples under the v-th view, For the first v Labeled negative samples under each view No. v Unlabeled samples under each view For labeled positive samples, For labeled negative samples, These are unlabeled samples.

[0008] Optionally, the initialization of the semi-supervised multi-view dual-dictionary SVDD model includes: Initialize the weights of each view ,satisfy and ; Initialize the first A comprehensive dictionary of views and analysis dictionary A random matrix with unit F norm. For the first The dictionary size for each view; Initialize the center of the hypersphere The mean and radius of the labeled normal sample codes are given. Initialize the hyperparameters to encode the maximum distance from the initial sample to the center of the sphere. , , , , , To smooth the weight coefficients, initialize the number of iterations and the convergence threshold.

[0009] Optionally, global iterative optimization includes: Perform the following single-round alternating convex optimization independently for each view, simultaneously completing the dynamic correction of unlabeled samples, iterating until convergence; The single-wheel alternating convex optimization includes: Step 1, Initial Hypothesis Anchoring: At the start of this iteration, all unlabeled samples are subject to normal sample constraints. Complete the initial hypothesis setting; among which, For the first Unlabeled samples of each view; Step 2, Encoding Layer Verification: Fixed Solve separately The closed-form solution; based on sparse coding, the reconstruction error of unlabeled samples and cross-view coding consistency deviation are calculated to complete the first round of verification of the initial hypothesis and screen out suspected abnormal samples; Step 3, Feature Layer Enhancement: Fixing Solve Closed-form solution; fixed Solve The closed-form solution; through step-by-step iterative updates of the dual dictionary, the feature reconstruction and encoding mapping capabilities of normal samples are enhanced, the feature differences of suspected abnormal samples are amplified, and the second round of verification is completed; Step 4, Decision-making level adjustment: Fixed Solving for the center of a hypersphere using the Lagrange duality method and radius Based on the distance from the sample to the hypersphere, a final determination is made for suspected anomalous samples. For high-confidence anomalous samples, their optimization constraints are modified to anomalous sample constraints. The criteria for identifying high-confidence outlier samples are: after the sample is coded using the analysis dictionary, it reaches the center of the hypersphere. The distance is greater than 1.5 times the radius of the hypersphere. Furthermore, the reconstruction error is greater than three times the average reconstruction error of normal samples; samples that meet both conditions are judged as high-confidence anomalous samples, thus completing the correction of the initial erroneous hypothesis.

[0010] Optionally, smoothing view weight updates via softmax includes: The comprehensive loss for each view is calculated, and temporary weights are obtained based on the exponential function and softmax normalization. The view weights are then updated using a smoothing formula. The comprehensive loss includes reconstruction loss, sparsity loss, encoding consistency loss, and SVDD loss. After the view weights are updated, determine whether the objective function meets the convergence threshold: if it does, terminate the iteration; if it does not, the iteration count t = t + 1, and return to the global iterative optimization to start the next round of iteration.

[0011] Optionally, the multi-view weighted fusion anomaly decision includes: Calculate the classification decision function for each view: in, For the v-th view targeting the sample to be detected The classification decision function value, The sample to be tested; Based on the classification decision and the updated view weights, the final classification is obtained through fusion: in, The final classification decision value after weighted fusion of multiple views; like If it is determined to be a normal sample; It was determined to be an abnormal sample.

[0012] The beneficial effects of this invention are as follows: 1. It solves the problem of scarce labels in industrial scenarios, requiring only a small number of labeled samples to train a high-precision model, significantly reducing labeling costs; 2. Adaptive fusion of multi-view information was achieved, effectively suppressing the interference of weak view noise and significantly improving the robustness of the model; 3. The dual-dictionary collaborative learning framework is adopted, which balances feature extraction capability and computational efficiency, and the inference speed is more than 3 times faster than the traditional dictionary learning method. 4. A closed-loop correction mechanism for unlabeled samples was designed to solve the model drift problem caused by the initial hypothesis bias. Even when the unlabeled samples contain 20% anomalies, the detection accuracy can still be maintained above 95%. 5. Experiments on the CWRU bearing fault diagnosis dataset show that the fault detection accuracy of the present invention reaches 96.8%, which is 14.5% higher than the traditional SVDD and 5.1% higher than the existing multi-view SVDD method. At the same time, the false detection rate is reduced to 2.3% and the false negative rate is reduced to 1.2%. Attached Figure Description

[0013] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0014] Figure 1 This is a schematic diagram of a semi-supervised multi-view SVDD industrial equipment fault detection method based on dictionary learning according to an embodiment of the present invention; Figure 2 This is a flowchart illustrating the global iterative optimization process in an embodiment of the present invention. Detailed Implementation

[0015] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0016] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0017] like Figure 1 As shown, this embodiment proposes a semi-supervised multi-view SVDD (Semi-supervised Multi-view Dictionary Detection) method for industrial equipment fault detection based on dictionary learning, including: Vibration signals from multiple sensors on industrial rotating machinery bearings are collected and preprocessed. Based on the preprocessed signal, a semi-supervised multi-view dual-dictionary SVDD model is constructed; wherein, the semi-supervised multi-view dual-dictionary SVDD model is configured with an objective function and outlier sample constraints; Initialize the semi-supervised multi-view dual-dictionary SVDD model; Based on the initialized model, global iterative optimization is performed. Hypothesis-validation-correction is carried out on unlabeled samples during the iteration, and the view weights are smoothly updated through softmax. Based on the weight update results, perform multi-view weighted fusion anomaly decision-making to obtain the final classifier; The final classifier is used to detect faults in industrial equipment.

[0018] Specifically, this embodiment proposes a semi-supervised multi-view SVDD method for industrial equipment fault detection based on dictionary learning. The data used is the CWRU bearing fault diagnosis dataset, which is a training set composed of a mixture of positive and negative samples. This invention introduces a dictionary learning method to enhance the feature extraction capability of the data samples. The core idea is to construct a unified objective function that couples dual dictionary learning and semi-supervised SVDD. This invention incorporates the constraint of labeled anomaly samples into the multi-view SVDD framework, adopts an alternating convex search framework for variable-based iterative optimization, designs a closed-loop "hypothesis-verification-correction" strategy for unlabeled samples and a smooth view weight update strategy, and finally fuses the multi-view decision results to achieve anomaly detection. The method includes the following steps: Step S1: Data Preprocessing: Collect vibration signals from multiple sources of industrial rotating machinery bearings. Perform non-overlapping segmentation and wavelet denoising on the raw signals, extract multi-view features for each sample, and construct training and testing sets. The training set is divided into a set of labeled normal samples. Labeled abnormal sample set and unlabeled sample set The multi-view features include: time-domain statistical features, frequency-domain amplitude features, and wavelet packet time-frequency domain features.

[0019] Step S2: Construct the semi-supervised multi-view dual-dictionary SVDD objective function: The objective function also includes a reconstruction error term, a sparse regularization term, a coding consistency term, an SVDD hypersphere constraint term, and a cross-view term. Figure 1 Consistency term and view weight term, the expressions are as follows: Constraints: Normal sample constraints: Abnormal sample constraints: Initial constraints for unlabeled samples: Non-negativity of slack variables: Dictionary atomic constraints: ( For the comprehensive dictionary (atoms) View weight constraints: in, The first A matrix of labeled normal, labeled abnormal, and unlabeled samples for each view; These are the sparse coding matrices for the corresponding samples; For the first The center of the SVDD hypersphere in one view The radius is ; These are slack variables used to tolerate sample bias and prevent overfitting; for Norms are used to promote feature row sparsity and achieve feature selection; cross-view Figure 1 Coherence terms Used to ensure that different views encode the same unlabeled sample consistently.

[0020] Step S3: Model Initialization: Initialize the weights of each view. ,satisfy and ; Initialize the first A comprehensive dictionary of views and analysis dictionary A random matrix with unit F norm ( For the first (Dictionary size of each view); initialize the hypersphere center. The mean and radius of the labeled normal sample codes are given. Initialize the hyperparameters to encode the maximum distance from the initial sample to the center of the sphere. (Sparse regularization weights) (Coding consistency weight) (Slack variable penalty coefficient) (Cross-view) Figure 1 Consistency weight) (Smoothing weight coefficients), initial iteration count t=0 and convergence threshold .

[0021] Step S4: As Figure 2 As shown, global iterative optimization: Perform the following single-round alternating convex optimization independently for each view, simultaneously completing the dynamic correction of unlabeled samples, iterating until convergence: S4.1 Initial Hypothesis Anchoring: At the start of this iteration, all unlabeled samples are subject to normal sample constraints. Complete the initial hypothesis setting.

[0022] S4.2 Coding Layer Verification: Fixed Solve separately The closed-form solution is obtained; the reconstruction error of unlabeled samples and the cross-view coding consistency deviation are calculated based on sparse coding to complete the first round of verification of the initial hypothesis and screen out suspected abnormal samples.

[0023] S4.3 Feature Layer Enhancement: Fixed Solve Closed-form solution; fixed Solve The closed-form solution is obtained; through step-by-step iterative updates of the dual dictionary, the feature reconstruction and encoding mapping capabilities of normal samples are enhanced, the feature differences of suspected abnormal samples are amplified, and the second round of verification is completed.

[0024] S4.4 Decision-making level correction: Fixed Solving for the center of a hypersphere using the Lagrange duality method and radius Based on the distance from the sample to the hypersphere, a final determination is made for suspected anomalous samples. For high-confidence anomalous samples, their optimization constraints are modified to anomalous sample constraints. The criteria for identifying high-confidence outlier samples are: after the sample is coded using the analysis dictionary, it reaches the center of the hypersphere. The distance is greater than 1.5 times the radius of the hypersphere. Furthermore, the reconstruction error is greater than three times the average reconstruction error of normal samples; samples that meet both conditions are judged as high-confidence anomalous samples, thus completing the correction of the initial erroneous hypothesis.

[0025] Step S5: Smoothly update view weights: Calculate the overall loss for each view. (Including reconstruction loss, sparsity loss, encoding consistency loss, and SVDD loss), temporary weights are obtained based on the exponential function and softmax normalization. Then update the view weights using a smoothing formula: This strategy maintains the stability of weight updates while quickly responding to changes in the performance of each view, adaptively enhancing strong views and suppressing noise in weak views.

[0026] Step S6 Multi-view weighted fusion anomaly decision: Calculate the classification decision function for each view: The fusion yields the final classifier: Decision-making rule: If If it is determined to be a normal sample; It was determined to be an abnormal sample.

[0027] When the method of this invention is applied to the fault detection of rolling bearings, the dataset selected is the internationally recognized CWRU bearing fault diagnosis dataset. This dataset was collected by the Bearing Data Center of Case Western Reserve University in the United States and is one of the most authoritative public datasets in the field of mechanical fault diagnosis.

[0028] In this embodiment, drive-end bearing data with a sampling frequency of 12kHz from the CWRU dataset is used, including one normal state and nine fault states. The nine fault states are inner ring fault, outer ring fault, and rolling element fault, each with three fault sizes (0.007 inches, 0.014 inches, and 0.021 inches). First, the acquired raw vibration signals are preprocessed, and each long time-series signal is divided into 2048 non-overlapping sampling points / samples, generating a total of 3200 samples, with 320 samples for each state.

[0029] Secondly, multi-view features are extracted for each sample, constructing a total of 3 views: See Figure 1 16-dimensional time-domain statistical features, including mean, variance, root mean square, peak value, kurtosis, margin, impulse factor, waveform factor, peak factor, kurtosis factor, skewness, peak-to-peak value, standard deviation, maximum value, minimum value, and mean square error; See Figure 2 : 64-dimensional frequency domain amplitude features, which convert the time-domain signal into a frequency-domain signal through Fast Fourier Transform (FFT) and take the amplitude of the first 64 frequency points as features; View 3: 8-dimensional wavelet packet time-frequency domain features. The time-domain signal is decomposed into 3 levels using the db4 wavelet to obtain 8 sub-bands. The energy of each sub-band is calculated as a feature.

[0030] Then, the 3200 samples were randomly divided into a training set and a test set. The training set comprised 80%, totaling 2560 samples (256 samples of one normal state and 256 samples of each of the nine fault states); the test set comprised 20%, totaling 640 samples (64 samples of one normal state and 64 samples of each of the nine fault states). The training set was strictly divided according to a real-world semi-supervised industrial scenario: 10% of the normal samples (26 samples) and 1% of the samples from each of the nine fault states (a total of 23 samples, 2-3 samples per fault state) were randomly selected as labeled samples; the remaining 2511 samples were all unlabeled samples, perfectly reflecting the actual working conditions in industrial scenarios where "normal samples are easy to label, but fault and abnormal samples are scarce and costly to label." The test set retained all labels and was used to evaluate model performance.

[0031] Next, set the model's hyperparameters: dictionary size. =32, maximum number of iterations is 100, convergence threshold is... Historical weighting coefficient exist Selected from the middle; sparse regularization coefficient Coding consistency coefficient and cross-view Figure 1 Coherence coefficient exist Selected from; slack variable penalty coefficient exist The optimal hyperparameter combination is selected from the pool of samples. Five-fold cross-validation is used, with validation set detection accuracy, F1 score, or overall loss as evaluation metrics.

[0032] The parameter values ​​mentioned above are only preferred settings for this embodiment. In other industrial equipment fault detection scenarios, they can be adjusted according to sample size, feature dimensions, and equipment operating conditions.

[0033] Perform model training according to steps S1-S5 of this invention until the change in global loss is less than the convergence threshold. After training, save the comprehensive dictionary, analysis dictionary, hypersphere center, hypersphere radius, and view weights for each view.

[0034] Finally, the test set samples were input into the trained multi-view weighted fusion classifier for fault detection, and the results were compared with existing mainstream anomaly detection methods. The experimental results are shown in Table 1 below: Table 1 The experimental results show that the method proposed in this invention significantly outperforms existing methods in all evaluation metrics. Compared with the traditional single-view SVDD method, the accuracy is improved by 14.5%; compared with the existing best multi-view SVDD method mSVDD, the accuracy is improved by 5.1%, while the false positive rate is reduced by 2.5 percentage points and the false negative rate is reduced by 8.7 percentage points. This fully demonstrates the effectiveness and superiority of the method in industrial bearing fault detection, enabling high-precision fault detection with low false positive and low false negative rates even in scenarios where tags are scarce.

[0035] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A dictionary learning based semi-supervised multi-view SVDD industrial equipment fault detection method, characterized in that, include: Vibration signals from multiple sensors on industrial rotating machinery bearings are collected and preprocessed. Based on the preprocessed signal, a semi-supervised multi-view dual-dictionary SVDD model is constructed; wherein, the semi-supervised multi-view dual-dictionary SVDD model is configured with an objective function and outlier sample constraints; Initialize the semi-supervised multi-view dual-dictionary SVDD model; Based on the initialized model, global iterative optimization is performed. Hypothesis-validation-correction is carried out on unlabeled samples during the iteration, and the view weights are smoothly updated through softmax. Based on the weight update results, perform multi-view weighted fusion anomaly decision-making to obtain the final classifier; The final classifier is used to detect faults in industrial equipment.

2. The dictionary learning based semi-supervised multi-view SVDD industrial equipment fault detection method according to claim 1, characterized in that, The process involves collecting vibration signals from multiple sensors on industrial rotating machinery bearings and performing preprocessing, including: The multi-source sensor vibration signals of the bearings of industrial rotating machinery are collected, the original signals are subjected to non-overlapping segmentation and wavelet denoising processing, multi-view features of each sample are extracted, and a training set and a test set are constructed; the training set is divided into a normal sample set with labels , an abnormal sample set with labels , and an unlabeled sample set ; the multi-view features include time domain statistical features, frequency domain amplitude features, and wavelet packet time-frequency domain features.

3. The dictionary learning based semi-supervised multi-view SVDD industrial equipment fault detection method according to claim 1, characterized in that, The semi-supervised multi-view dual-dictionary SVDD model includes: The objective function includes: a reconstruction error term, a sparse regularization term, a coding consistency term, an SVDD hypersphere constraint term, a cross-view consistency term, and a view weight term. The expression of the objective function is: in, Weights for each view, and The first Comprehensive and analytical dictionaries for each view. These are the sparse coding matrices for the corresponding samples. The first A matrix of labeled normal, labeled abnormal, and unlabeled samples for each view. For the first The radius of the SVDD hypersphere in each view As slack variables, for Norm, V For the number of views, For sparse regularization weights, For coding consistency weights, The penalty coefficient for slack variables, F for F Norm, This represents the number of labeled positive samples. The number of labeled negative samples. This represents the number of unlabeled samples. i For sample index, For cross-view consistency weights, For the first View indexes, For the first One view index; Constraints: Normal sample constraints: Abnormal sample constraints: Initial constraints for unlabeled samples: Non-negativity of slack variables: Dictionary atomic constraints: View weight constraints: in, For the first The center of the SVDD hypersphere in one view For the comprehensive dictionary One atom, For the labeled positive samples under the v-th view, For the first v Labeled negative samples under each view No. v Unlabeled samples under each view For labeled positive samples, For labeled negative samples, These are unlabeled samples.

4. The semi-supervised multi-view SVDD industrial equipment fault detection method based on dictionary learning according to claim 3, characterized in that, The initialization of the semi-supervised multi-view dual-dictionary SVDD model includes: Initialize the weights of each view ,satisfy and ; Initialize the first A comprehensive dictionary of views and analysis dictionary A random matrix with unit F norm. For the first The dictionary size for each view; Initialize the center of the hypersphere The mean and radius of the labeled normal sample codes are given. Initialize the hyperparameters to encode the maximum distance from the initial sample to the center of the sphere. , , , , , To smooth the weight coefficients, initialize the number of iterations and the convergence threshold.

5. The semi-supervised multi-view SVDD industrial equipment fault detection method based on dictionary learning according to claim 3, characterized in that, Global iterative optimization includes: Perform the following single-round alternating convex optimization independently for each view, simultaneously completing the dynamic correction of unlabeled samples, iterating until convergence; The single-wheel alternating convex optimization includes: Step 1, Initial Hypothesis Anchoring: At the start of this iteration, all unlabeled samples are subject to normal sample constraints. Complete the initial hypothesis setting; Step 2, Encoding Layer Verification: Fixed Solve separately The closed-form solution; based on sparse coding, the reconstruction error of unlabeled samples and cross-view coding consistency deviation are calculated to complete the first round of verification of the initial hypothesis and screen out suspected abnormal samples; Step 3, Feature Layer Enhancement: Fixing Solve Closed-form solution; fixed Solve The closed-form solution; through step-by-step iterative updates of the dual dictionary, the feature reconstruction and encoding mapping capabilities of normal samples are enhanced, the feature differences of suspected abnormal samples are amplified, and the second round of verification is completed; Step 4, Decision-making level adjustment: Fixed Solving for the center of a hypersphere using the Lagrange duality method and radius Based on the distance from the sample to the hypersphere, a final determination is made for suspected anomalous samples. For high-confidence anomalous samples, their optimization constraints are modified to anomalous sample constraints. The criteria for identifying high-confidence outlier samples are: after the sample is coded using the analysis dictionary, it reaches the center of the hypersphere. The distance is greater than 1.5 times the radius of the hypersphere. Furthermore, the reconstruction error is greater than three times the average reconstruction error of normal samples; samples that meet both conditions are judged as high-confidence anomalous samples, thus completing the correction of the initial erroneous hypothesis.

6. The semi-supervised multi-view SVDD industrial equipment fault detection method based on dictionary learning according to claim 3, characterized in that, Smoothly updating view weights using softmax includes: The comprehensive loss for each view is calculated, and temporary weights are obtained based on the exponential function and softmax normalization. The view weights are then updated using a smoothing formula. The comprehensive loss includes reconstruction loss, sparsity loss, encoding consistency loss, and SVDD loss. After the view weights are updated, determine whether the objective function meets the convergence threshold: if it does, terminate the iteration; if it does not, the iteration count t = t + 1, and return to the global iterative optimization to start the next round of iteration.

7. The semi-supervised multi-view SVDD industrial equipment fault detection method based on dictionary learning according to claim 1, characterized in that, The multi-view weighted fusion anomaly decision includes: Calculate the classification decision function for each view: in, For the v-th view targeting the sample to be detected The classification decision function value, The sample to be tested; Based on the classification decision and the updated view weights, the final classification is obtained through fusion: in, The final classification decision value after weighted fusion of multiple views; like If it is determined to be a normal sample; It was determined to be an abnormal sample.

Images