Bearing unbalanced open-set fault diagnosis method based on deep fusion of simulated measurement data
By employing a deep fusion method of simulated test data, and utilizing physical statistical feature consistency fusion and adaptive open set networks, the problems of data imbalance and atypical fault identification in rolling bearing fault diagnosis were solved, achieving a more efficient fault diagnosis effect.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-16
- Publication Date
- 2026-03-27
AI Technical Summary
In existing technologies, rolling bearing fault diagnosis models suffer from insufficient ability to identify data imbalances and atypical faults, especially the distribution differences between simulation and measured data and the lack of ability to identify atypical faults that have not been observed.
A deep fusion method of simulation data is adopted. The simulation data is adjusted to match the actual measurement data through a physical statistical feature consistency fusion strategy. An adaptive open set network is constructed to learn domain-invariant feature representations. Fault diagnosis is performed using prototype metric loss and adaptive rejection threshold.
It effectively solves the problem of data imbalance, improves the diagnostic performance of the model in real-world scenarios, can identify atypical faults that have not been observed, and improves the reliability and accuracy of diagnosis.
Smart Images
Figure CN121740441A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of bearing fault diagnosis technology, and in particular to a bearing unbalanced open set fault diagnosis method based on deep fusion of simulation data. Background Technology
[0002] Rolling bearings are critical components for power transmission and load bearing in mechanical systems, widely used in important industries such as aviation, aerospace, and energy. As vulnerable components in mechanical systems, approximately one-third of mechanical failures are caused by bearing failure. The health of rolling bearings has a significant impact on the stable operation of mechanical systems. Once a bearing fails, timely warnings and corresponding measures should be taken; otherwise, it can easily lead to equipment damage or even safety accidents. Rolling bearing fault diagnosis technology can promptly identify the fault point and locate the cause of the fault, which is of great significance for ensuring the safe operation of equipment.
[0003] Artificial intelligence technologies, represented by deep learning, have been widely applied to rolling bearing fault diagnosis due to their powerful adaptive feature extraction capabilities. However, the diagnostic performance of deep learning models is limited by the imbalanced data in real-world industrial settings. Existing diagnostic methods to address data imbalance primarily focus on the algorithmic and data layers. At the algorithmic level, cost-sensitive learning, for example, alleviates class imbalance by adjusting the weights of the loss function. However, this method heavily relies on the quality of minority class samples, making it prone to overfitting and limiting its generalization ability when there are very few fault samples. At the data layer, generative adversarial networks (GANs) attempt to artificially generate fault samples to balance the dataset. However, the training process of GANs is unstable, the generated samples lack clear physical interpretability, and they are essentially imitations of existing minority samples, making it difficult to generate effective new samples outside the distribution.
[0004] To fundamentally address the problem of scarce fault samples, researchers have begun to introduce physics-based dynamic simulation models to generate large amounts of simulated fault data. These methods can provide comprehensive knowledge of typical faults and are expected to overcome the limitation of scarce real fault samples, potentially enabling a fundamental rebalancing of datasets. However, two key bottlenecks exist in practical applications:
[0005] (1) Significant distribution differences exist between simulation and measured data. Idealized simulation data cannot reproduce the complex noise, transmission path attenuation, and measurement errors in the actual test environment, resulting in significant distribution differences between simulation and measured data in terms of amplitude, statistical distribution, and other characteristics. Directly mixing the two for training will severely impair the diagnostic performance of the model in real-world scenarios.
[0006] (2) Lack of ability to identify atypical faults not previously observed. Existing simulation models can typically only simulate typical faults such as those of the outer race, inner race, and rolling elements. For atypical faults with complex or coupled mechanisms, such as cage faults or compound faults, there is a lack of effective modeling methods. This causes models trained based on simulation data to completely lose their ability to distinguish atypical fault modes not observed during training. Summary of the Invention
[0007] The purpose of this invention is to address the problems in the background technology, such as the significant distribution differences between simulation and measured data and the lack of ability to identify atypical faults that have not been observed, by proposing a bearing unbalanced open set fault diagnosis method based on deep fusion of simulation and measured data.
[0008] The technical solution of this invention: A bearing imbalance open set fault diagnosis method based on deep fusion of simulation data, comprising the following steps:
[0009] S1. Data Acquisition and Fusion Enhancement: Acquire simulated vibration data of rolling bearings under known typical fault conditions, as well as measured vibration data containing a small number of known typical fault samples and a large number of normal samples; use a physical statistical feature consistency fusion strategy to adjust the simulated vibration data so that the multi-dimensional physical statistical feature distribution of the adjusted data matches the measured vibration data, and generate an enhanced training dataset.
[0010] S2. Adaptive Open Set Network Training: Construct an adaptive open set network using known samples from the augmented training dataset and the actual test data as training data to train the network; the training process jointly optimizes the classification loss, domain adaptation loss, and prototype metric loss to enable the network to learn domain-invariant feature representations and learn a prototype vector for each known typical fault category; calculate the adaptive rejection threshold corresponding to each known category;
[0011] S3. Open set fault diagnosis: The trained adaptive open set network is used to extract the features of the test samples to be diagnosed and calculate the Euclidean distance between the features and the prototype vectors of each category; the judgment is made based on the Euclidean distance and the adaptive rejection threshold to distinguish between known fault categories and unknown faults.
[0012] Preferably, in S1, the simulated vibration data is generated by a five-degree-of-freedom nonlinear dynamic model of the rolling bearing.
[0013] Preferably, in S1, the physical statistical feature consistency fusion strategy is as follows: the simulated vibration data sample is subjected to amplitude scaling and noise injection controlled by learnable parameters to obtain adjusted data; time-domain physical statistical features are extracted to form feature vectors; the parameters are iteratively optimized with the goal of minimizing the average Mahalanobis distance between the feature vectors of the adjusted data and the feature vectors of the measured data.
[0014] Preferably, in S2, the adaptive open set network includes a feature extractor, a classifier, and a prototype learning module; the domain adaptation loss adopts the local maximum mean difference loss; the prototype metric loss includes an intra-class aggregation term that makes features of samples of the same class closer to their corresponding prototype vectors, and an inter-class separation term that makes features of samples of different classes keep their distance from their non-corresponding prototype vectors.
[0015] Preferably, in S2, the adaptive rejection threshold is calculated for each known category k according to the following formula:
[0016] T k =μ k +λσ k ;
[0017] Among them, T k Let μ be the rejection threshold for the k-th class. k and σ k denoted as the mean and standard deviation of the distance between the features of all measured samples belonging to class k in the training set and the prototype vectors of class k, respectively, and λ is an adjustable coefficient.
[0018] Preferably, in S3, if the distance is less than the threshold, it is determined to be the corresponding known fault category; otherwise, it is determined to be an unknown fault.
[0019] Preferably, in S3, known typical fault categories include outer ring faults, inner ring faults, and rolling element faults; unknown faults include cage faults and compound faults.
[0020] Compared with existing technologies, this invention has the following beneficial technical effects: it constructs a complete diagnostic process from data generation, fusion enhancement, model training to open-set decision-making. Compared with existing technologies, the diagnostic method described in this invention forms a complete solution through the organic combination of three levels. First, at the data level, a physical statistical feature consistency fusion strategy is proposed, which effectively utilizes the complete fault knowledge of simulation data and overcomes the distribution difference between simulation data and measured data, fundamentally alleviating the data imbalance problem caused by the scarcity of real fault samples. Second, at the model level, an adaptive open-set network with fusion domain adaptation and prototype measurement is designed. Through joint optimization, fine-grained feature alignment between simulation and measured data is achieved, and a prototype feature space with a compact intra-class and separable inter-class structure is learned. Finally, at the decision-making level, a prototype-based class adaptive rejection threshold mechanism is proposed, providing a reliable basis for identifying atypical faults that did not appear during training. Attached Figure Description
[0021] Figure 1 This is a flowchart illustrating the implementation of the method of the present invention;
[0022] Figure 2 For the dynamic simulation model of rolling bearings;
[0023] Figure 3 A schematic diagram illustrating the consistency fusion strategy for physical and statistical characteristics of simulation test data;
[0024] Figure 4 A comparison chart of probability density functions for simulation, measured, and augmented data;
[0025] Figure 5 This is a schematic diagram of an adaptive open set network. Detailed Implementation
[0026] Example 1; as Figure 1 As shown, the present invention proposes a bearing imbalance open set fault diagnosis method based on deep fusion of simulation test data, which includes the following steps:
[0027] S1. Data Acquisition and Fusion Enhancement: Acquire simulated vibration data of rolling bearings under known typical fault conditions, as well as measured vibration data containing a small number of known typical fault samples and a large number of normal samples; use a physical statistical feature consistency fusion strategy to adjust the simulated vibration data so that the multi-dimensional physical statistical feature distribution of the adjusted data matches the measured vibration data, and generate an enhanced training dataset.
[0028] S2. Adaptive Open Set Network Training: Construct an adaptive open set network using known samples from the augmented training dataset and the actual test data as training data to train the network; the training process jointly optimizes the classification loss, domain adaptation loss, and prototype metric loss to enable the network to learn domain-invariant feature representations and learn a prototype vector for each known typical fault category; based on the trained network and training data, calculate the adaptive rejection threshold corresponding to each known category.
[0029] S3. Open set fault diagnosis: The trained adaptive open set network is used to extract the features of the test samples to be diagnosed and calculate the Euclidean distance between the features and the prototype vectors of each category; the judgment is made based on the Euclidean distance and the adaptive rejection threshold to distinguish between known fault categories and unknown faults.
[0030] Example 2; as Figures 2-5 As shown, the present invention proposes a bearing unbalance open set fault diagnosis method based on deep fusion of simulation test data. Compared with Embodiment 1, this embodiment details S1-S3.
[0031] In S1, the simulated vibration data is generated by a five-degree-of-freedom nonlinear dynamic model of the rolling bearing. This model simulates the contact force between the rolling elements and the raceway based on Hertzian contact theory, and simulates outer ring faults, inner ring faults, and rolling element faults by injecting local damage into the model. The physical statistical feature consistency fusion strategy is as follows: the simulated vibration data samples are subjected to amplitude scaling and noise injection controlled by learnable parameters to obtain adjusted data; multiple time-domain physical statistical features, including root mean square value, signal-to-noise ratio, kurtosis, peak-to-peak value, skewness, peak factor, impulse factor, and margin factor, are extracted to form a feature vector; the parameters are iteratively optimized with the goal of minimizing the average Mahalanobis distance between the feature vector of the adjusted data and the feature vector of the measured data.
[0032] In S2, the adaptive open-set network includes a feature extractor, a classifier, and a prototype learning module. The domain adaptation loss employs a local maximum mean difference loss to align the overall distribution of features between the simulated source domain and the measured target domain, as well as the conditional distribution of each known class. The prototype metric loss includes an intra-class aggregation term that encourages features of similar samples to approach their corresponding prototype vectors, and an inter-class separation term that encourages features of different classes to maintain a distance from their non-corresponding prototype vectors. The adaptive rejection threshold is calculated for each known class k according to the following formula:
[0033] T k =μ k +λσ k ;
[0034] Among them, T k Let μ be the rejection threshold for the k-th class. k and σ k denoted as the mean and standard deviation of the distance between the features of all measured samples belonging to class k in the training set and the prototype vectors of class k, respectively, and λ is an adjustable coefficient.
[0035] In S3, if the distance is less than the threshold, it is determined to be a known fault category; otherwise, it is determined to be an unknown fault. Known typical fault categories include outer ring faults, inner ring faults, and rolling element faults; unknown faults include cage faults and compound faults.
[0036] Example 3; as Figures 2-5 As shown, this invention proposes a bearing imbalance open set fault diagnosis method based on deep fusion of simulation test data. Compared with Embodiment 2, this embodiment details the execution process.
[0037] First, through Figure 2 The five-degree-of-freedom nonlinear dynamic simulation model of the rolling bearing shown simulates vibration signals of known typical fault states such as outer ring fault, inner ring fault, and rolling element fault, forming a simulation dataset. in The known K types of typical fault labels are used. Simultaneously, measured vibration signals of the bearings are collected from the equipment to be diagnosed, forming a measured dataset. During the training phase, only normal samples and a small number of known typical fault samples can be obtained from it.
[0038] To address the distribution discrepancies between simulation and measured data and to achieve effective data augmentation, the following measures were implemented: Figure 3 The physical statistical feature consistency fusion strategy is shown. Specifically, for each simulation sample x... s The adjusted sample is generated by performing amplitude scaling and noise injection using the following formula.
[0039]
[0040] Where A is the amplitude scaling factor and η is the noise modulation coefficient. Let N(0,1) be the standard deviation of the simulated sample, and let N(0,1) represent the standard normal distribution noise.
[0041] To measure and reduce distributional discrepancies, a set of time-domain statistical features with explicit physical meaning are defined, including: root mean square value, signal-to-noise ratio, kurtosis, peak-to-peak value, skewness, peak factor, impulse factor, and margin factor. For any signal x, its 8-dimensional feature vector F(x) is obtained through the feature extraction function F().
[0042] The core optimization objective of the fusion strategy is to minimize the distribution difference between the adjusted simulation data and the target measured data in the physical statistical feature space. Its optimization objective can be expressed as:
[0043] min A,η d M (F(D t ),F(x a ));
[0044] Where F() represents the signal feature vector extraction operator, d M () represents the average Mahalanobis distance between the adjusted sample and the target domain data features, which is calculated as follows:
[0045]
[0046] Where S is the joint characteristic covariance matrix.
[0047] The parameters A and η are iteratively optimized using the gradient descent method to make d M Converging to the minimum. After optimization, all enhanced signals... The statistical distribution of the enhanced signal is highly consistent with that of the measured signal. Compared with the original simulated signal, the distribution difference between the enhanced signal and the measured signal is significantly reduced, such as... Figure 4As shown, the augmented signal, together with the measured signal, forms a balanced augmentation dataset for subsequent network training.
[0048] Subsequently, as Figure 5 As shown, an adaptive open-set network is constructed, whose main body consists of a feature extractor, a classifier, and a prototype learning module. The augmented dataset is input into the network, and the feature extractor (a convolutional neural network containing three convolutional pooling layers) maps the data into high-dimensional features f = G. f (x). The classifier (containing three fully connected layers) outputs the closed-set classification probability. Furthermore, the network maintains a learnable prototype vector p for each known class k. k .
[0049] The total loss function L of the network total It consists of three parts:
[0050]
[0051] Where: Classification loss L clc Standard cross-entropy loss is used to ensure the ability to distinguish known classes;
[0052] Domain adaptation loss Local maximum mean difference loss is employed to align the overall distribution of the enhanced simulation data and the measured data, as well as the conditional distribution of each class, in the feature space, thereby learning domain-invariant features;
[0053] Prototype metric loss L pm It consists of intra-class aggregation terms and inter-class separation terms, aiming to shape a structured feature space. Intra-class aggregation term L intra Used to force the features of a sample to be close to the prototype of its corresponding class. Inter-class separating term L inter The requirement is that the distance between a sample and the prototypes of other classes is at least *m* greater than the distance to its own prototype. The detailed formula derivation is as follows:
[0054]
[0055] Where d j,k This represents the Euclidean distance between the extracted high-dimensional features and each known class prototype.
[0056] Minimize the total loss L using stochastic gradient descent (such as the Adam optimizer). total Iteratively update the network parameters of the feature extractor and classifier, as well as the prototype vectors of each known class.
[0057] After the adaptive open-set network is trained, all parameters are frozen. Known class samples from the actual training set are then input into the network again to obtain their feature representations, and the Euclidean distance between each sample and its corresponding class prototype is calculated. For each known category k, the adaptive rejection threshold T for that category is determined according to the following formula. k :
[0058] T k =μ k +λσ k ;
[0059] Where, μ k and σ k denoted as the mean and standard deviation of the distance between the features of all measured samples belonging to class k in the training set and the prototype vectors of class k, respectively, and λ is an adjustable coefficient.
[0060] Finally, for the test sample x to be diagnosed in the test set... t First, its features f are calculated using a feature extractor. i The classifier obtains a preliminary prediction result k of the known class. The distance d between the features of the test sample and the predicted class prototype is then calculated. i By comparing d i The rejection threshold T corresponding to category k k Perform an open-set decision. If d i ≤T k If d i >T k If so, the sample is determined to be an unseen atypical fault.
[0061] This specific implementation case uses real vibration data from rotating equipment. In this scenario, the training set contains 200 normal samples and 10 known typical fault samples (including outer ring faults, inner ring faults, and rolling element faults), forming an imbalance ratio of 20:1. The combined inner and outer ring fault, as an unseen atypical fault, only exists in the test set. All vibration signal samples have a uniform length of 4096 points, with no overlapping sampling.
[0062] For a comprehensive evaluation, five representative methods were selected for comparison, including Focalloss, ACGAN, and DT-ATFS for imbalanced scenarios, as well as ACGAN+MOS and DT-ATFS+OpenMax, which combine with existing open-set recognition methods. Three metrics were used for comprehensive evaluation: overall accuracy (OA), known class accuracy (Os), and unknown class recall (Uk), defined as follows:
[0063]
[0064] Among them, C k C represents the number of known class samples correctly identified by the model in tests within the target domain. u N represents the number of unseen atypical fault samples correctly identified by the model during testing in the target domain.k N represents the number of all known class samples in the target domain. u This represents the number of all unseen atypical fault samples in the target domain.
[0065] Table 1 shows the diagnostic performance comparison of each method on the specific implementation case test set. The experiment was repeated 10 times, and the data in the table are the average values. As can be seen from the table, the method of the present invention significantly outperforms all the comparison methods in terms of overall accuracy (OA), known class accuracy (Os), and unknown class recall (Uk). The above experimental results demonstrate that the present invention, through the collaborative design of the physical statistical feature consistency fusion strategy and the adaptive open set network, effectively solves the problem of the coexistence of data imbalance and open set fault identification, and significantly improves the comprehensive performance and reliability of the diagnostic model under actual complex working conditions.
[0066] Table 1 (Comparison Results of Algorithms):
[0067]
[0068] In summary, this invention constructs a complete diagnostic process from data generation, fusion enhancement, model training to open-set decision-making. Compared with existing technologies, the diagnostic method described in this invention forms a complete solution through the organic combination of three levels. First, at the data level, a physical statistical feature consistency fusion strategy is proposed, effectively utilizing the complete fault knowledge of simulation data and overcoming the distribution differences between simulation data and measured data, fundamentally alleviating the data imbalance problem caused by the scarcity of real fault samples. Second, at the model level, an adaptive open-set network with fusion domain adaptation and prototype measurement is designed, achieving fine-grained feature alignment between simulation and measured data through joint optimization, and learning a prototype feature space with a compact intra-class and separable inter-class structure. Finally, at the decision-making level, a prototype-based class adaptive rejection threshold mechanism is proposed, providing a reliable basis for identifying atypical faults not appearing during training.
[0069] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited thereto. Various changes can be made within the scope of knowledge possessed by those skilled in the art without departing from the spirit of the present invention.
Claims
1. A method for diagnosing bearing imbalance open-set faults through deep fusion of simulation test data, characterized in that, Includes the following steps: S1. Data Acquisition and Fusion Enhancement: Acquire simulated vibration data of rolling bearings under known typical fault conditions, as well as measured vibration data containing a small number of known typical fault samples and a large number of normal samples; use a physical statistical feature consistency fusion strategy to adjust the simulated vibration data so that the multi-dimensional physical statistical feature distribution of the adjusted data matches the measured vibration data, and generate an enhanced training dataset. S2. Adaptive Open Set Network Training: Construct an adaptive open set network, using known samples from the augmented training dataset and the actual test data as training data to train the network; The training process jointly optimizes the classification loss, domain adaptation loss, and prototype metric loss, enabling the network to learn domain-invariant feature representations and learn a prototype vector for each known typical fault category; and calculates the adaptive rejection threshold corresponding to each known category. S3. Open set fault diagnosis: The trained adaptive open set network is used to extract the features of the test samples to be diagnosed and calculate the Euclidean distance between the features and the prototype vectors of each category; the judgment is made based on the Euclidean distance and the adaptive rejection threshold to distinguish between known fault categories and unknown faults.
2. The bearing imbalance open set fault diagnosis method based on deep fusion of simulation test data according to claim 1, characterized in that, In S1, the simulated vibration data is generated by a five-degree-of-freedom nonlinear dynamic model of the rolling bearing.
3. The bearing imbalance open set fault diagnosis method based on deep fusion of simulation test data according to claim 1, characterized in that, In S1, the physical statistical feature consistency fusion strategy is as follows: the simulated vibration data sample is subjected to amplitude scaling and noise injection controlled by learnable parameters to obtain adjusted data; time-domain physical statistical features are extracted to form feature vectors; the parameters are iteratively optimized with the goal of minimizing the average Mahalanobis distance between the feature vectors of the adjusted data and the feature vectors of the measured data.
4. The bearing imbalance open set fault diagnosis method based on deep fusion of simulation test data according to claim 1, characterized in that, In S2, the adaptive open set network includes a feature extractor, a classifier, and a prototype learning module; the domain adaptation loss adopts the local maximum mean difference loss; the prototype metric loss includes an intra-class aggregation term that makes features of samples of the same class move closer to their corresponding prototype vectors, and an inter-class separation term that makes features of samples of different classes keep their distance from their non-corresponding prototype vectors.
5. The bearing imbalance open set fault diagnosis method based on deep fusion of simulation test data according to claim 1, characterized in that, In S2, the adaptive rejection threshold is calculated for each known category k according to the following formula: T k =μ k +λσ k ; Among them, T k Let μ be the rejection threshold for the k-th class. k and σ k denoted as the mean and standard deviation of the distance between the features of all measured samples belonging to class k in the training set and the prototype vectors of class k, respectively, and λ is an adjustable coefficient.
6. The bearing imbalance open set fault diagnosis method based on deep fusion of simulation test data according to claim 1, characterized in that, In S3, if the distance is less than the threshold, it is determined to be the corresponding known fault category; otherwise, it is determined to be an unknown fault.
7. The bearing imbalance open set fault diagnosis method based on deep fusion of simulation test data according to claim 1, characterized in that, In S3, known typical fault categories include outer ring faults, inner ring faults, and rolling element faults; unknown faults include cage faults and compound faults.