Semi-supervised bearing fault diagnosis method based on threshold optimization and feature fusion
By combining semi-supervised domain adaptation and neural network optimization of pseudo-label confidence and entropy threshold, the problem of inconsistent data distribution in bearing fault diagnosis by deep learning models is solved, and efficient bearing fault identification and diagnosis is achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHANGCHUN UNIV OF TECH
- Filing Date
- 2025-12-09
- Publication Date
- 2026-06-12
AI Technical Summary
Deep learning models in bearing fault diagnosis are limited by the availability of labeled data and the consistency of data distribution, resulting in insufficient accuracy and reliability, especially when the data distribution changes in industrial environments.
By combining semi-supervised domain adaptation and neural network methods, the Chaotic Sparrow Search algorithm is used to optimize the false label confidence and entropy threshold. Feature extraction and fusion are performed through a domain-invariant feature network to improve the accuracy of false labels and the robustness of fault diagnosis.
It significantly improves the accuracy and robustness of bearing fault diagnosis, achieves efficient fault identification in complex environments, and demonstrates remarkable cross-domain fault diagnosis performance.
Smart Images
Figure CN122196648A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of fault diagnosis, and specifically discloses a semi-supervised bearing fault diagnosis method based on threshold optimization and feature fusion. Background Technology
[0002] As a core component of motor equipment, the stable operation of bearings is crucial to overall production safety. Therefore, the rapid and accurate identification of bearing faults and the development of efficient intelligent fault diagnosis technologies have become urgent needs in the industry.
[0003] The accuracy of deep learning models in bearing fault diagnosis is limited by the availability of labeled data and the consistency of data distribution. In industrial environments, acquiring labeled data is both time-consuming and expensive, and data distribution often changes over time due to environmental factors and equipment wear, affecting the accuracy and reliability of deep learning models in practical applications. To address these challenges, transfer learning has attracted significant attention in the field of fault diagnosis. This technology effectively alleviates the problem of inconsistent data distribution between training and test sets by leveraging cross-domain knowledge transfer, and is particularly suitable for situations where significant changes in operating conditions lead to changes in data distribution, making it highly valuable for bearing fault diagnosis.
[0004] Domain adaptation, as an important branch of transfer learning, can reduce the difference between the source and target domains by adjusting model parameters, thereby improving the model's performance in the target domain. However, poor quality pseudo-labels generated in the target domain can introduce significant bias during model training, thus affecting the accuracy of bearing fault diagnosis.
[0005] To address the shortcomings of existing methods, this invention combines semi-supervised domain adaptation with neural network methods, designing a method that utilizes the Chaotic Sparrow Search Optimization Algorithm (CSSOA) to optimize pseudo-label confidence. ) and entropy threshold ( This paper presents a bearing fault diagnosis method. Simultaneously, optimized pseudo-labels are used as actual labels in the target domain, enabling feature extraction and fusion between the source and target domains, ultimately achieving unbiased feature learning across domains. Summary of the Invention
[0006] This invention provides a semi-supervised bearing fault diagnosis method based on threshold optimization and feature fusion. This method can solve the problems of difficulty in obtaining dataset labels and inconsistent data distribution between training and test sets in the field of bearing fault diagnosis.
[0007] The core of this invention is to create a domain-invariant feature network to enhance fault diagnosis performance in the target domain through effective knowledge transfer. It mainly includes three key stages: pseudo-label generation, pseudo-label optimization, and the creation of a domain adaptation network.
[0008] S1: Use a recurrent neural network (RNN) to predict and generate pseudo-labels;
[0009] S2: Optimize the confidence of pseudo-tags using CSSOA ( ) and entropy threshold ( )parameter;
[0010] S3: Based on the above steps, construct a domain adaptation network based on Deep Belief Networks (DBN).
[0011] The specific process of step S1 is as follows:
[0012] S101: Perform wavelet denoising and Z-score normalization on the original data;
[0013] S102: Train an RNN model using source domain data and generate pseudo-labels for target domain data to capture the dynamic features of time series.
[0014] S103: Calculate the objective function of the RNN network and minimize the difference between the predicted and true labels, as shown in the following expression:
[0015]
[0016] in, This indicates the number of samples contained in the target domain. It is the set of weights and biases of all layers in an RNN. , They are the first Layer weight matrix and bias vector, It represents the total number of layers in the neural network. It is the model number Each sample belongs to The predicted probability of a class is determined by the parameter. Sure; It is the first Each sample belongs to The actual label of the class.
[0017] The specific process of step S2 is as follows:
[0018] S201: Confidence level ( ) and entropy threshold ( Set the initial search range;
[0019] S202: Using CSSOA, the optimal solution is sought in the parameter space, and the candidate threshold set is continuously refined to gradually approach the optimal confidence level. ) and entropy threshold ( );
[0020] S203: Use Tent chaos theory to enhance randomness and avoid getting trapped in local optima;
[0021] S204: Settings As the termination condition, where The number of times the threshold parameter is optimized. This represents the maximum number of iterations.
[0022] S205: Calculate the fitness value and select the optimal combination of threshold parameters, as shown in the following expression:
[0023]
[0024] Where, mean( ) represents all confidence thresholds within the set range ( The average of ) and mean( ) represents all entropy thresholds within a set range ( The average value of )
[0025] S206: Output new pseudo-tags via CSSOA , as the actual label of the target domain.
[0026] The specific process of step S3 is as follows:
[0027] S301: Utilize DBN to extract key features of the source and target domains and identify new fault types in the target domain;
[0028] S302: The Maximum Mean Discrepancy (MMD) algorithm is used to reduce inter-domain variability. The expression for the MMD algorithm is shown below:
[0029]
[0030]
[0031]
[0032]
[0033] in, and These represent the feature sets of the source domain and the target domain, respectively. and Indicates the number of samples included; function It is responsible for mapping the original data to a new feature space. Regenerating Hilbert space. It contains kernel functions. The expression is as follows:
[0034]
[0035] in, This represents the bandwidth of the kernel function;
[0036] S303: After feature extraction is completed, DBN performs feature fusion to identify the fault type;
[0037] S304: Feature classification using DBN, the objective function expression is as follows:
[0038]
[0039] in, This represents the actual label of the source domain. Indicates the predicted label of the source domain. Pseudo-tags representing the target domain, This represents the pseudo-label obtained after threshold optimization, with parameters... This represents the balance factor.
[0040] By integrating all the aforementioned objective functions, a comprehensive domain adaptation network is constructed, and the objective function expression is shown below:
[0041]
[0042] in, and This represents the weighting parameter.
[0043] In summary, compared with the prior art, the advantages of the present invention are as follows:
[0044] First, this invention uses CSSOA to optimize confidence and entropy thresholds, thereby improving the accuracy of pseudo-labels and the robustness of the fault diagnosis system.
[0045] Second, this invention utilizes DBN for feature extraction and fusion, which significantly improves the ability to identify complex faults and demonstrates excellent fault diagnosis results.
[0046] Third, this invention sets up multiple transfer tasks on a real bearing fault dataset, and shows significant fault classification performance on both the test set and the training set, achieving ideal cross-domain fault diagnosis results. Attached Figure Description
[0047] Figure 1This is a flowchart of the fault diagnosis method described in this invention;
[0048] Figure 2 This is the initial probability density estimation diagram described in this invention;
[0049] Figure 3 This is a flowchart of the threshold parameter optimization process described in this invention;
[0050] Figure 4 This is the fitness curve analysis diagram described in this invention;
[0051] Figure 5 This is the optimized probability density estimation diagram described in this invention;
[0052] Figure 6 This is an area curve analysis diagram under the receiver operation characteristic curve described in this invention. Detailed Implementation
[0053] The present invention will be further described below with reference to the embodiments and the accompanying drawings:
[0054] The semi-supervised bearing fault diagnosis method based on threshold optimization and feature fusion described in this invention is as follows: Figure 1 As shown, the detailed steps of this method are as follows:
[0055] S1: Start the PT700 bearing test platform.
[0056] S101: This platform integrates a three-phase asynchronous AC motor gearbox, frequency converter, and brake, among other devices.
[0057] S102: This platform uses vibration and current sensors to detect performance.
[0058] S2: Collect data.
[0059] S201: Data acquisition under a UC206 rolling bearing with 0.3 mm of artificially induced damage;
[0060] S202: Replace the faulty gear and collect data on five bearing conditions under normal conditions, inner and outer ring faults, cage and rolling element faults, and sampling frequency of 5,000 times per second.
[0061] S3: Split the dataset.
[0062] S301: Collected fault data Divided into labeled source domains and unlabeled target domains ;
[0063] S302: Using pseudo-labels Replace the target domain labels and estimate the conditional distribution of the target domain. ;
[0064] S303: Use a sliding window method to process data, and set the fixed sampling point of the window to 1024;
[0065] S304: Set the sampling frequency, sample size, fault type, and operating conditions of the dataset as shown in Table 1:
[0066]
[0067] S4: Divide migration tasks and assign labels.
[0068] S401: Based on the set source domain tags, generate and optimize the target domain pseudo tags;
[0069] S402: Utilizing confidence levels ( ) and entropy threshold ( Evaluate the quality of pseudo-tags to ensure that the source domain tag categories fully cover the target domain;
[0070] S403: Set up migration task and This ensures that the source domain covers all fault types of the data, while the fault types of the target domain may be the same as or different from those of the source domain.
[0071] S5: Generate pseudo tags.
[0072] S501: Wavelet denoising and Z-score normalization are used to process the original vibration signal to obtain a new dataset. ;
[0073] S502: Use source domain Training an RNN model;
[0074] S503: Target domain The input is fed into the trained RNN to generate predicted pseudo-labels. ;
[0075] S504: Optimize the weights and bias parameters of the RNN objective function, as shown in the following expression:
[0076]
[0077] in, This indicates the number of samples contained in the target domain. This represents the set of weights and biases for all layers of an RNN. , They are the first Layer weight matrix and bias vector, It represents the total number of layers in the neural network. The model is for the first Each sample belongs to The predicted probability of a class is determined by the parameter. Sure; It is the first Each sample belongs to The actual label of the class.
[0078] S6: Evaluate the initial pseudo-labels.
[0079] S601: Evaluate the initial pseudo-label distribution generated by RNN in the target domain using probability density estimation methods;
[0080] S602: Determine the consistency and deviation between the initial pseudo-labels and the actual data distribution. The results are as follows: Figure 2 As shown.
[0081] S7: Optimize pseudo-tags.
[0082] S701: Confidence level ( ) and entropy threshold ( Set the initial search range;
[0083] S702: Using CSSOA, the optimal solution is sought in the parameter space, and the candidate threshold set is continuously refined to gradually approach the optimal solution. ) and entropy threshold ( );
[0084] S703: Uses Tent chaos theory to enhance randomness and avoid getting trapped in local optima;
[0085] S704: Settings As the termination condition, where The number of times the threshold parameter is optimized. This represents the maximum number of iterations.
[0086] S705: Calculate the fitness value and select the optimal combination of threshold parameters. The specific expression is as follows:
[0087]
[0088] Where, mean( ) represents all confidence thresholds within the set range ( The average of ) and mean( ) represents all entropy thresholds within a set range ( The average value of )
[0089] S706: Output new pseudo-tags via CSSOA , as the actual label of the target domain.
[0090] S8: Evaluate the optimized pseudo-labels.
[0091] S801: CSSOA is compared with Bayesian Optimization (BO), Particle Swarm Optimization (PSO) and Sparrow Search Algorithm (SSA). The corresponding parameter settings and parameter values are shown in Table 2. The BO algorithm uses the radial basis function kernel as its core function.
[0092]
[0093] S802: Set the same number of iterations, sample size, optimization objective, and parameter range for the four optimization algorithms;
[0094] S803: Calculate the fitness functions of the four optimization algorithms, find the optimal solution with high confidence and low entropy, and the results are as follows. Figure 4 As shown;
[0095] S804: After CSSOA optimization, the optimized pseudo-tags and target domains are re-evaluated using probability density estimation methods. The results are as follows: Figure 5 As shown.
[0096] S9: Domain feature extraction and fusion.
[0097] S901: Utilize DBN to extract key features from the source and target domains to compensate for insufficient data coverage;
[0098] S902: In the feature extraction process, the MMD algorithm is introduced to minimize the distribution difference between the source domain and the target domain;
[0099] S903: After feature extraction is completed, DBN fuses features from different domains to identify new fault types;
[0100] S904: Calculate the classification loss of the DBN to enhance the model's adaptability to each domain feature and its classification accuracy. The expression for this loss function is as follows:
[0101]
[0102] in, Indicates the actual label of the source domain. Indicates the predicted label of the source domain. Pseudo-tags representing the target domain, This represents the pseudo-label obtained after threshold optimization, with parameters... This represents the balance factor.
[0103] S10: Perform fault diagnosis.
[0104] S1001: Four fault diagnosis models are established based on four optimization algorithms, namely: Domain Adaptation Based on Bayesian Optimization and Deep Belief Network (BO-DADBN), Domain Adaptation Based on Particle Swarm Optimization and Deep Belief Network (PSO-DADBN), Domain Adaptation Based on Sparrow Search Algorithm and Deep Belief Network (SSA-DADBN), and the Domain Adaptation Based on Chaotic Sparrow Search Optimization Algorithm and Deep Belief Network (CSS-DADBN) proposed in this invention.
[0105] S1002: Using the features of the source and target domains as inputs to the fault diagnosis model, and minimizing the objective function, bearing fault classification is achieved. The expression of the objective function is as follows:
[0106]
[0107] in, and This represents the weighting parameter.
[0108] S1003: The classification performance of each model is evaluated using the area curve analysis method under the receiver operating characteristic curve;
[0109] S1004: Check the area under the curve of the method proposed in this invention to ensure that it is between 0 and 1. The closer the area value is to 1, the better the classification performance of the model.
[0110] S1005: Determine whether the area under the curve of the method proposed in this invention reaches the maximum compared to other methods. Visualize this using an area curve analysis plot under the receiver operation characteristic curve. The results are as follows: Figure 6 As shown.
Claims
1. A semi-supervised bearing fault diagnosis method based on threshold optimization and feature fusion, characterized in that, The fault diagnosis method covers three core stages: pseudo-label generation, pseudo-label optimization, and creation of a domain adaptation network, specifically including: S1: Use a recurrent neural network (RNN) to predict and generate pseudo-labels; S101: Perform wavelet denoising and Z-score normalization on the original data; S102: Train an RNN model using source domain data and generate pseudo-labels for target domain data; S103: Calculate the objective function of the RNN network and minimize the difference between the predicted and true labels, as shown in the following expression: in, This indicates the number of samples contained in the target domain. This represents the set of weights and biases for all layers of an RNN. , They are the first The weight matrix and bias vector of the layer, It represents the total number of layers in the neural network. It is the model number Each sample belongs to The predicted probability of a class is determined by the parameter. Sure, It is the first Each sample belongs to The true label of the class; S2: Optimize the confidence of pseudo-tags using the Chaotic Sparrow Search Optimization Algorithm (CSSOA). ) and entropy threshold ( )parameter; S201: Confidence level ( ) and entropy threshold ( Set the initial search range; S202: Using CSSOA, the optimal solution is sought in the parameter space, and the candidate threshold set is continuously refined to gradually approach the optimal confidence level. ) and entropy threshold ( ); S203: Use Tent chaos theory to enhance randomness and avoid getting trapped in local optima; S204: Settings As the termination condition, where The number of optimization attempts for the threshold parameter. This represents the maximum number of iterations. S205: Calculate the minimum fitness value using the model's fitness function in the target domain, and select the optimal combination of threshold parameters. The expression is as follows: Where, mean( ) represents all confidence thresholds within the set range ( The average of ) and mean( ) represents all entropy thresholds within a set range ( The average value of ) S206: Output new pseudo-tags via CSSOA , as the actual label of the target domain; S3: Based on the above steps, construct a domain adaptation network based on Deep Belief Networks (DBN); S301: Use DBN to extract key features of the source and target domains and identify new fault types in the target domain; S302: The Maximum Mean Discrepancy (MMD) algorithm is used to reduce inter-domain variability. The expression for the MMD algorithm is shown below: in, and Let them represent the feature sets of the source domain and the target domain, respectively. and Indicates the number of samples included; function Responsible for mapping the original data to a new feature space, reproducing the Hilbert space. It contains kernel functions. The expression is as follows: in, This represents the bandwidth of the kernel function; S303: After feature extraction is completed, DBN performs feature fusion to identify the fault type; S304: Feature classification using DBN, the objective function expression is as follows: in, Indicates the actual label of the source domain. Indicates the predicted label of the source domain. Pseudo-tags representing the target domain, This represents the pseudo-label obtained after threshold optimization, with parameters... Indicates the balance factor; S4: Integrate all the aforementioned objective functions to construct a comprehensive domain adaptation network. The objective function expression is as follows: in, and This represents the weighting parameter.