Cross-working-condition bearing fault diagnosis method based on deep subdomain adversarial network
The cross-condition bearing fault diagnosis method using deep subdomain adversarial networks, which trains the DSA-MPCNN model with LMMD distance and neighborhood discriminator, solves the problems of accuracy and generalization ability of traditional methods in cross-condition bearing fault diagnosis, and achieves efficient and accurate bearing fault diagnosis.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-01-21
- Publication Date
- 2026-04-28
AI Technical Summary
Traditional bearing fault diagnosis methods based on manual judgment are difficult to achieve high accuracy. Deep learning methods have limited generalization ability across operating conditions and the cost of acquiring fault data is high. Differences in the distribution of existing data lead to inaccurate diagnosis.
A cross-condition bearing fault diagnosis method based on deep subdomain adversarial networks is adopted. By preprocessing the data of key bearing components, the DSA-MPCNN model is trained using the Local Maximum Mean Difference (LMMD) distance and the neighborhood discriminator to achieve cross-condition fault diagnosis.
It improves the accuracy and generalization ability of bearing fault diagnosis, with an average accuracy of over 91%, which is better than existing methods and reduces data acquisition costs.
Smart Images
Figure CN121935667A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of mechanical equipment condition monitoring technology, specifically relating to a cross-condition bearing fault diagnosis method based on deep subdomain adversarial networks. Background Technology
[0002] Transmission systems are widely used in important production fields such as road transportation, aerospace, shipbuilding, and military industry, and bearings are one of the most basic and critical components in these systems. As a key component supporting the connection, operation, and power transmission of equipment, maintaining the healthy condition of bearings is crucial for equipment operation. However, in real-world working environments, bearings are constantly in operation and often fail due to improper installation, environmental corrosion, and alternating loads. If bearing failures are not detected in time, they can not only affect the normal operation of equipment but, in severe cases, even endanger the lives of operators, causing serious production accidents.
[0003] The vibration and impact of a localized bearing failure are extremely weak compared to the vibration during transmission system operation. This makes it difficult to achieve high accuracy in traditional manual judgment-based diagnosis of the operating status of key bearing components, hindering intelligent condition monitoring of these components. The development of artificial intelligence and deep learning technologies has provided new impetus for intelligent fault diagnosis. However, while fault diagnosis methods based on deep learning networks such as CNN, ResNet, and Transformer and their improved versions have achieved some success, data-driven deep learning fault diagnosis methods still face data challenges. As the carrier of fault characteristics, data plays a decisive role in data-driven fault diagnosis methods. Deep learning fault diagnosis models require a large amount of labeled fault data for training, and training data and test data are usually assumed to have the same distribution. However, the amount of labeled data collected in actual engineering is very small, and the complex and variable operating conditions of transmission systems mean that actual monitoring data and model training data do not necessarily follow the same data distribution conditions. This severely limits the diagnostic and generalization capabilities of deep learning fault diagnosis methods constrained by the assumption that training and test data follow the same distribution. Furthermore, obtaining fault data for critical bearing components often requires building corresponding fault test benches and artificially creating various faults to obtain valuable fault samples. This makes data acquisition for fault diagnosis of critical bearing components difficult and costly. However, a large amount of similar fault data for critical bearing components currently exists, derived from simulation models, laboratory data, and publicly available datasets. Rationally utilizing these data sources and new technologies to improve model generalization performance is of great significance for the further development of deep learning fault diagnosis models. Summary of the Invention
[0004] In view of the above problems, the present invention provides a cross-condition bearing fault diagnosis method based on deep subdomain adversarial networks, which solves the technical problems that the traditional signal processing-based fault diagnosis method in the prior art cannot meet the needs of bearing local fault monitoring, and that a large amount of transmission system vibration interference leads to inaccurate deep learning intelligent fault diagnosis based on vibration signals and limited generalization ability.
[0005] This invention provides a cross-condition bearing fault diagnosis method based on deep subdomain adversarial networks, the specific steps of which are as follows: Step 1. Preprocess the fault data of key bearing components; obtain source domain samples and target domain samples; The key components of a bearing include the inner raceway, outer raceway, cage, and rollers. Step 2. Divide the source domain samples and target domain samples to obtain the model training set and model test set; perform feature transformation on the data in the model training set and model test set to obtain the corresponding time-frequency distribution map; The training set consists of a portion of the target domain samples and all the source domain samples, while the test set consists of the remaining target domain samples. Step 3. Train a DSA-MPCNN cross-domain fault diagnosis model based on a deep subdomain adversarial adaptive parallel convolutional network that combines LMMD distance and a domain discriminator using the time-domain features, empirical features, and time-frequency distribution map of the model training set; Step 4. For the trained DSA-MPCNN cross-domain fault diagnosis model, test the model using the test dataset to obtain a DSA-MPCNN cross-domain fault diagnosis model that meets the requirements; use the DSA-MPCNN cross-domain fault diagnosis model that meets the requirements to perform cross-working-condition bearing fault diagnosis.
[0006] Optionally, the DSA-MPCNN cross-domain fault diagnosis model includes a feature extraction module, a domain adaptive feature alignment module, and a fault classification module.
[0007] Optionally, the feature extraction module includes a time-frequency domain feature extraction channel, a time-domain feature extraction channel, and an empirical feature extraction channel.
[0008] Optionally, the time-frequency domain feature extraction channel takes the time-frequency distribution map as input and dynamically extracts time-frequency domain fault feature information by combining the attention mechanism of the convolution module.
[0009] Optionally, the temporal feature extraction channel uses the model training set as input and combines a multi-head attention mechanism to dynamically extract temporal fault feature information.
[0010] Optionally, the empirical feature extraction channel extracts the root mean square value, variance, kurtosis, margin, permutation entropy, and centroid frequency of the model training set.
[0011] Optionally, in step 3, a fault classification loss trained based on source domain samples is used. 、 The domain discrimination loss based on the domain discriminator and the distribution difference distance loss based on the LMMD distance distribution metric are used as loss functions in the training process of the DSA-MPCNN cross-domain fault diagnosis model.
[0012] Optionally, the fault classification loss is trained based on samples from the source domain. The expression is:
[0013] in, Represents samples from the source domain; Indicates the first i The true labels of samples from each source domain; This represents the total number of samples in the source domain; This represents the feature extractor of the feature extraction module; This represents a fault classifier; This represents the fault classification loss function; Indicates the first i Fault data of samples from each source domain.
[0014] Optionally, the domain discrimination loss is based on the domain discriminator. The expression is:
[0015] in, and These represent the source domain samples and the target domain samples, respectively. Indicates the first i The real domain of each source domain sample This represents the total number of samples in the source domain and the target domain. Indicates the first i Diagnostic results for samples from individual source domains; This represents a domain discriminator. This represents the sub-domain discriminant loss function; Indicates the first i Fault data of samples from individual source domains; This represents the feature extractor of the feature extraction module.
[0016] Optionally, the distribution difference distance loss is based on the LMMD distance distribution metric. The expression is:
[0017] in, and These represent the source domain samples and the target domain samples, respectively. This represents the set of joint distributions whose marginal distributions are the data distributions of the source domain samples and the target domain samples. Expressing expectations, Represents the distance function; The joint distribution represents the set of joint distributions whose marginal distributions are the data distributions of the sample sets in the source domain and the target domain. This represents sample data from the source domain. This represents sample data from the target domain.
[0018] Compared with the prior art, the present invention has at least the following beneficial effects: (1) The bearing fault diagnosis method under different working conditions of the present invention utilizes a large amount of existing data and a small amount of fault data to reduce the difference in the distribution of fault data characteristics under different working conditions, and can characterize common fault information under different working conditions, thereby improving the generalization ability of fault diagnosis. (2) The cross-condition bearing fault diagnosis method of the present invention has high accuracy in cross-condition bearing fault diagnosis based on the deep subdomain adversarial adaptive parallel convolution model. The average accuracy can reach more than 91%, which is higher than the existing fault diagnosis methods. Attached Figure Description
[0019] The accompanying drawings are for illustrative purposes only and are not intended to limit the scope of the invention.
[0020] Figure 1 This is a flowchart of the cross-condition bearing fault diagnosis method based on deep subdomain adversarial networks according to the present invention. Figure 2 This is a schematic diagram of the DSA-MPCNN cross-domain fault diagnosis model of the present invention; Figure 3 This is a schematic diagram of the feature alignment mechanism of the present invention; Figure 4 This is a schematic diagram illustrating the fault classification principle of the present invention. Detailed Implementation
[0021] To better understand the above-described objectives, features, and advantages of the present invention, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, unless otherwise specified, the embodiments of the present invention and the features thereof can be combined with each other. Furthermore, the present invention can be implemented in other ways different from those described herein; therefore, the scope of protection of the present invention is not limited to the specific embodiments disclosed below.
[0022] A specific embodiment of the present invention, such as Figures 1-4This paper presents a cross-condition bearing fault diagnosis method based on deep subdomain adversarial networks, with the following specific steps: Step 1. Preprocess the fault data of key bearing components; obtain source domain samples and target domain samples.
[0023] Specifically, variational mode decomposition and bandpass filtering signal processing algorithms are applied to preprocess the fault data of key bearing components, and signal enhancement is performed on the fault data; the preprocessed fault data of key bearing components is then divided into source domain samples. and target domain samples The labels of the source domain samples are known.
[0024] Furthermore, the key components of a bearing include the inner raceway, the outer raceway, the cage, and the rollers.
[0025] Step 2. Divide the source domain samples and target domain samples to obtain the model training set and model test set; perform feature transformation on the data in the model training set and model test set to obtain the corresponding time-frequency distribution map.
[0026] Specifically, a portion of the target domain samples and all source domain samples constitute the model training set, while the remaining target domain samples constitute the test set.
[0027] Preferably, the target domain sample comprises 15% of the target domain sample.
[0028] Furthermore, signal processing algorithms such as wavelet transform are used to perform feature transformation on the data in the model training set and the model test set to obtain the time-frequency distribution maps of the corresponding model training set and model test set.
[0029] Furthermore, the model training set and time-frequency distribution map are used as inputs to the cross-domain fault diagnosis model.
[0030] Step 3. Use the time-domain features, empirical features, and time-frequency distribution map of the model training set to train a DSA-MPCNN cross-domain fault diagnosis model based on a deep subdomain adversarial adaptive parallel convolutional network that combines the local maximum mean difference (LMMD) distance and the neighborhood discriminator.
[0031] See Figure 2 The DSA-MPCNN cross-domain fault diagnosis model includes a feature extraction module, a domain adaptive feature alignment module, and a fault classification module.
[0032] Specifically, the feature extraction module includes a time-frequency domain feature extraction channel, a time-domain feature extraction channel, and an empirical feature extraction channel.
[0033] Furthermore, the time-frequency domain feature extraction channel takes the time-frequency distribution map as input and dynamically extracts time-frequency domain fault feature information by combining the attention mechanism of the convolution module, which serves as the input data for the domain adaptive feature alignment module; Furthermore, the temporal feature extraction channel uses the model training set as input and combines a multi-head attention mechanism to dynamically extract temporal fault feature information, which is then used as input data for the domain adaptive feature alignment module. Furthermore, the empirical feature extraction channel extracts the root mean square value, variance, kurtosis, margin, permutation entropy, and centroid frequency of the model training set as input data for the domain adaptive feature alignment module.
[0034] Furthermore, in the feature extraction process, the source domain samples and target domain samples, after being processed by the feature extraction structure, yield source domain features respectively. and target domain features This serves as the input data for the domain-adaptive feature alignment module.
[0035] It is understandable that the source domain sample features and the target domain sample features both include the corresponding time-frequency domain fault feature information, time-domain fault feature information, root mean square value, variance, kurtosis, margin, permutation entropy, and centroid frequency.
[0036] Specifically, see Figure 3 The domain-adaptive feature alignment module includes a neighborhood discriminator and an LMMD distance distribution metric, used to extract shared fault features between the source and target domains, which serve as input to the fault classification module. During training, the neighborhood discriminator and the feature extractor in the feature extraction module undergo adversarial training using an adaptive adversarial training method. Through adversarial training of the feature extractor and the neighborhood discriminator, the differences in data distribution under different operating conditions are gradually reduced. The ultimate goal is to enable the neighborhood discriminator to extract shared features from data under different operating conditions, achieving feature alignment of the overall data distribution.
[0037] Since the data distributions of the source and target domains differ, this invention achieves better transfer of knowledge from the source domain to the target domain through feature alignment of the overall data distribution, thereby improving the model's generalization ability and performance in the target domain. Furthermore, traditional domain adaptive transfer only focuses on minimizing global distribution differences during feature alignment, neglecting local data alignment. This invention treats each fault category as a sub-domain and measures the data distribution differences of these sub-domains under different operating conditions based on the LMMD distance. It then aligns corresponding sample data of the same fault type under different operating conditions, further reducing the data distribution differences across different operating conditions.
[0038] Further, see Figure 4 The fault classification module includes a source domain fault classifier and a target domain fault classifier. First, it is based on the features of the source domain... The source domain fault classifier is trained using sample labels corresponding to the aforementioned features, enabling it to accurately identify features in the source domain. And perform correct classification. Secondly, construct a target domain fault classifier, sharing parameters with the source domain fault classifier. The goal of feature alignment is to extract fault features shared by the source and target domains; therefore, during training, when target domain features... and source domain characteristics As they get closer, the target domain fault classifier, which shares fault features, is able to classify faults in the target domain. Classify correctly.
[0039] Furthermore, the model training set obtained in step 2 is used to train the DSA-MPCNN cross-domain fault diagnosis model.
[0040] Specifically, the loss function in the training process of the DSA-MPCNN cross-domain fault diagnosis model consists of three parts: (1) Fault classification loss based on source domain samples. (2) Domain discrimination loss based on domain discriminator (3) Distribution difference distance loss based on LMMD distance distribution metric .
[0041] Furthermore, fault classification loss The loss represents the ability of the cross-domain fault diagnosis model to identify and classify input samples. The smaller the loss, the more accurate the fault classification. Therefore, the first optimization objective of the DSA-MPCNN cross-domain fault diagnosis model is to minimize the fault classification loss trained on samples from the source domain. The expression is: (1) in, Represents samples from the source domain. Indicates the first i The true labels of samples from each source domain n s This represents the total number of samples in the source domain. This represents the feature extractor of the feature extraction module. Represents a fault classifier. This represents the fault classification loss function; Indicates the first i Fault data of samples from each source domain.
[0042] Furthermore, the domain-adaptive feature alignment module includes a domain discriminator and an LMMD distance distribution metric, aiming to learn shared features across different domains. First, according to the domain adaptation theory based on adversarial training, if the domain discriminator cannot correctly distinguish features between the source and target domains, then that feature is a domain-shared feature. Therefore, the second optimization objective of the domain-adaptive feature alignment module is to maximize the domain discrimination loss based on the domain discriminator. The expression is: (2) in, and These represent the source domain samples and the target domain samples, respectively. Indicates the first i The real domain of each source domain sample This represents the total number of samples in the source domain and the target domain. Indicates the first i Diagnostic results for samples from individual source domains; This represents a domain discriminator. This represents the sub-domain discriminant loss function.
[0043] Furthermore, the distribution difference distance between features learned from the same fault type in different domains is directly measured using the LMMD distance. Therefore, the third optimization objective of the DSA-MPCNN cross-domain fault diagnosis model of this invention is to minimize the distribution difference distance between data in the same sub-domain of the source and target domains, i.e., the distribution difference distance loss based on the LMMD distance distribution metric. The expression is: (3) in, This represents the set of joint distributions whose marginal distributions are the data distributions of the source domain samples and the target domain samples. Expressing expectations, Represents the distance function; The joint distribution represents the set of joint distributions whose marginal distributions are the data distributions of the sample sets in the source domain and the target domain. This represents sample data from the source domain. This represents sample data from the target domain.
[0044] Furthermore, combining the above three loss functions, the expression for the overall loss function of the cross-domain fault diagnosis model is as follows: (4).
[0045] Furthermore, in order to train the cross-domain fault diagnosis model, this invention introduces a gradient reversal layer between the neighborhood discriminator and the feature extractor. When calculating the model gradient to update the model parameters, the parameter update direction of the neighborhood discriminator is optimized in the direction of increasing the neighborhood discrimination loss.
[0046] Specifically, the model parameters are updated to the parameters of the feature extractor, fault classifier, and domain discriminator.
[0047] Specifically, the gradient reversal layer updates the parameters of the neighborhood discriminator in the same direction as the gradient during forward propagation; and updates the parameters of the neighborhood discriminator in the opposite direction of the gradient during backward propagation. The expression for the gradient reversal layer is: (5)
[0048] in, This represents the gradient reversal layer function. I represents the gradient reversal weighting coefficients, and I represents the gradient matrix. The output parameters of the domain adaptive feature alignment module.
[0049] Furthermore, combining equations (4) and (5), the final model loss function of the cross-domain fault diagnosis model in this invention can be obtained, and its expression is shown in equation (6): (6) in, These are the weighting coefficients for the distance metric loss. This represents the gradient reversal weighting coefficient.
[0050] Understandable, , and Let represent the parameters of the feature extractor, fault classifier, and domain discriminator, respectively. Then, the final model loss function can be rewritten as: (7) in, These represent the optimal parameter combinations that minimize the loss function.
[0051] Furthermore, the parameters of the feature extractor, fault classifier, and neighborhood discriminator in the cross-domain fault diagnosis model are updated using the stochastic gradient descent (SGD) algorithm. , and The update formulas are shown in equations (8), (9), and (10): (8) (9) (10) in, Indicates the learning rate; , and These represent the updated parameters of the feature extractor, fault classifier, and neighborhood discriminator, respectively.
[0052] Step 4. For the trained DSA-MPCNN cross-domain fault diagnosis model, test the model using the test dataset, calculate and check the diagnostic accuracy of the cross-domain fault diagnosis model on the target domain test samples, and apply it after the requirements are met.
[0053] To illustrate the effectiveness of the method proposed in this invention, the following detailed description of the above technical solution is provided through a specific embodiment: A fault dataset of planetary gear transmission mechanisms was collected, as shown in Table 1. Three speed condition datasets—OC1, OC2, and OC3—were selected for methodological explanation. N, B, and G represent healthy bearings, bearings with inner ring faults, and bearings with outer ring faults, respectively, and 12 and 22 represent progressively more severe faults. Six transfer fault diagnosis experiments were created in this experiment. For example, OC1→OC3 means that the OC1 dataset is considered the source domain and the OC3 dataset is considered the target domain; while OC3→OC1 means that the OC3 dataset is considered the source domain and the OC1 dataset is considered the target domain. As shown in Table 2, each transfer diagnosis task contains 6000 training set samples and 4000 test set samples. The training set samples contain 5000 source domain samples and 1000 target domain samples; all samples in the test set belong to the target domain.
[0054] Table 1 Description of fault data across speed operating conditions
[0055] Table 2 Details of Fault Data Migration and Usage Across Speed-Rate Operating Conditions
[0056] This paper compares the proposed cross-condition bearing fault diagnosis method (DSA-MPCNN) based on a deep subdomain adversarial adaptive parallel convolutional model with existing domain adversarial neural network (DANN), deep convolutional transfer learning (DTLN), and multi-adversarial domain adaptive network (DAMA) methods. All methods were trained and inferred using the data usage shown in Table 2, and the Adam optimization algorithm was used for network training and parameter updates. The cross-entropy loss function was used to calculate the loss during network training.
[0057] The diagnostic results of the proposed method and the comparative methods for fault diagnosis are summarized in Table 3. Table 3 shows that the average diagnostic accuracy of the proposed DSA-MPCNN method is 91.56%, the average diagnostic accuracy of the DANN method is 81.82%, the average diagnostic accuracy of the MADA method is 87.23%, and the average diagnostic accuracy of the DTLN method is 88.68%. It is evident that the average diagnostic accuracy of the proposed DSA-MPCNN method is 9.74% higher than the DANN method, 4.33% higher than the MADA method, and 2.88% higher than the DTLN method. Therefore, the fault diagnosis accuracy of the proposed DSA-MPCNN method is significantly better than the other three cross-condition fault diagnosis methods. Furthermore, the DSA-MPCNN method exhibits the highest diagnostic accuracy in almost all transfer diagnosis experiments, demonstrating the robustness of the proposed method.
[0058] Table 3 Summary of Fault Diagnosis Results Across Speed Conditions
[0059] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.
Claims
1. A cross-condition bearing fault diagnosis method based on deep subdomain adversarial networks, characterized in that, The specific steps are as follows: Step 1. Preprocess the fault data of key bearing components; obtain source domain samples and target domain samples; The key components of a bearing include the inner raceway, outer raceway, cage, and rollers. Step 2. Divide the source domain samples and target domain samples to obtain the model training set and model test set; perform feature transformation on the data in the model training set and model test set to obtain the corresponding time-frequency distribution map; The training set consists of a portion of the target domain samples and all the source domain samples, while the test set consists of the remaining target domain samples. Step 3. Train a DSA-MPCNN cross-domain fault diagnosis model based on a deep subdomain adversarial adaptive parallel convolutional network that combines LMMD distance and a domain discriminator using the time-domain features, empirical features, and time-frequency distribution map of the model training set; Step 4. For the trained DSA-MPCNN cross-domain fault diagnosis model, test the model using the test dataset to obtain a DSA-MPCNN cross-domain fault diagnosis model that meets the requirements; use the DSA-MPCNN cross-domain fault diagnosis model that meets the requirements to perform cross-working-condition bearing fault diagnosis.
2. The cross-condition bearing fault diagnosis method according to claim 1, characterized in that, The DSA-MPCNN cross-domain fault diagnosis model includes a feature extraction module and a domain... Adaptive feature alignment module and fault classification module.
3. The cross-condition bearing fault diagnosis method according to claim 2, characterized in that, The feature extraction module includes a time-frequency domain feature extraction channel, a time-domain feature extraction channel, and an empirical feature extraction channel.
4. The cross-condition bearing fault diagnosis method according to claim 3, characterized in that, The time-frequency domain feature extraction channel takes the time-frequency distribution map as input and dynamically extracts time-frequency domain fault feature information by combining the attention mechanism of the convolution module.
5. The cross-condition bearing fault diagnosis method according to claim 3, characterized in that, The temporal feature extraction channel uses the model training set as input and combines a multi-head attention mechanism to dynamically extract temporal fault feature information.
6. The cross-condition bearing fault diagnosis method according to claim 3, characterized in that, The empirical feature extraction channel extracts the root mean square value, variance, kurtosis, margin, permutation entropy, and centroid frequency of the model training set.
7. The cross-condition bearing fault diagnosis method according to claim 1, characterized in that, In step 3, a fault classification loss trained based on source domain samples is used. 、 The domain discrimination loss based on the domain discriminator and the distribution difference distance loss based on the LMMD distance distribution metric are used as loss functions in the training process of the DSA-MPCNN cross-domain fault diagnosis model.
8. The cross-condition bearing fault diagnosis method according to claim 7, characterized in that, Fault classification loss based on source domain samples training The expression is: in, Represents samples from the source domain; Indicates the first i The true labels of samples from each source domain; This represents the total number of samples in the source domain; This represents the feature extractor of the feature extraction module; This represents a fault classifier; This represents the fault classification loss function; Indicates the first i Fault data of samples from each source domain.
9. The cross-condition bearing fault diagnosis method according to claim 7, characterized in that, Domain discrimination loss based on domain discriminator The expression is: in, and These represent the source domain samples and the target domain samples, respectively. Indicates the first i The real domain of each source domain sample This represents the total number of samples in the source domain and the target domain. Indicates the first i Diagnostic results for samples from individual source domains; This represents a domain discriminator. This represents the sub-domain discriminant loss function; Indicates the first i Fault data of samples from individual source domains; This represents the feature extractor of the feature extraction module.
10. The cross-condition bearing fault diagnosis method according to claim 7, characterized in that, Distribution Dispersion Loss Based on LMMD Distance Distribution Metric The expression is: in, and These represent the source domain samples and the target domain samples, respectively. This represents the set of joint distributions whose marginal distributions are the data distributions of the sample sets from the source domain and the target domain. Expressing expectations, Represents the distance function; The joint distribution represents the set of joint distributions whose marginal distributions are the data distributions of the sample sets in the source domain and the target domain. This represents sample data from the source domain. This represents sample data from the target domain.