Fault diagnosis method based on residual generative adversarial network and domain adaptation
By using residual generative adversarial networks and domain adaptation mechanisms, high-quality simulation data is generated and fused with real data, solving the problems of scarce labeled samples and insufficient multimodal information fusion in motor fault diagnosis, and realizing high-precision cross-domain fault diagnosis.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-26
- Publication Date
- 2026-05-15
AI Technical Summary
Existing motor fault diagnosis methods rely on a large number of high-quality labeled samples, which are difficult to obtain. Traditional hardware detection methods are highly invasive and depend on expert experience. Most deep learning methods ignore multimodal information fusion, resulting in insufficient diagnostic accuracy.
By employing residual generative adversarial networks and domain adaptation mechanisms, a multimodal feature extractor is constructed by fusing high-quality simulation data with real data, thereby achieving data augmentation and cross-domain fault diagnosis.
Under conditions of scarce labeled samples, the accuracy and robustness of fault diagnosis are improved, high-quality augmented data is generated to expand the training set, multimodal features are effectively integrated, and the diagnostic accuracy of the model under cross-domain conditions is improved.
Smart Images
Figure CN122045912A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of fault diagnosis technology, specifically relating to a fault diagnosis method based on residual generative adversarial networks and domain adaptation, which is particularly suitable for achieving high-precision fault diagnosis under conditions of scarce labeled samples. Background Technology
[0002] Electric motors often operate under various conditions and harsh environments, including variable speed, variable load, high temperature, extreme cold, and strong vibration, making them prone to various types of faults. These faults not only affect the motor's production efficiency, reliability, and safety, but can also cause equipment damage, and even lead to serious safety accidents and economic losses. Therefore, developing efficient and reliable intelligent fault diagnosis and monitoring technologies is of great significance for improving equipment lifespan and overall operational reliability.
[0003] Currently, motor fault diagnosis methods mainly include traditional hardware testing, physical model-based methods, and data-driven methods. Traditional hardware testing requires periodic disassembly of the motor or installation of internal testing devices, a complex process that relies on expert experience and has a degree of subjectivity, making it difficult to apply widely. Physical model-based methods depend on the motor's physical characteristics and operating mechanisms; while performing well in specific fault assessments, they require high model accuracy and prior knowledge. In recent years, with the development of artificial intelligence technology, data-driven fault diagnosis methods have gradually become a research hotspot. Methods such as neural networks, support vector machines, sparse representations, and deep learning can automatically identify fault types and severity based on training data.
[0004] However, existing methods still have the following problems: (1) Most deep learning fault diagnosis models rely on a large number of high-quality labeled samples, but in actual engineering, it is difficult to obtain such samples due to cost constraints. (2) Traditional hardware testing methods are invasive and rely on expert experience, which is not conducive to their promotion; (3) Existing fault diagnosis methods often focus on single-dimensional features and ignore the fusion of multimodal information, which can easily lead to the loss of key features and affect the accuracy of diagnosis. Summary of the Invention
[0005] To address the aforementioned problems in existing technologies, the present invention aims to provide a fault diagnosis method based on residual generative adversarial networks and domain adaptation. This method effectively utilizes simulation data and a small amount of real data, achieves data augmentation through residual generative adversarial networks, and combines a multimodal domain adaptation mechanism to improve the model's fault classification capability in the target domain.
[0006] This invention provides the following technical solution: a fault diagnosis method based on residual generative adversarial networks and domain adaptation, comprising the following steps: Step 1: Collect the actual raw signal data of the faulty equipment and preprocess it. Extract the time domain features of the preprocessed data and perform Fourier transform on the data to generate a spectrum. Step 2: Establish a finite element model of the synchronous motor, simulate various fault behaviors, generate tagged simulation signal data, and perform the same preprocessing operation on the simulation data as in Step 1. Step 3: Construct a diagnostic model that includes a residual generative adversarial network module and a domain adaptive network module; Step 4: Based on the simulation data generated in Step 2 and the real data obtained in Step S1, jointly train and optimize the diagnostic model; Step 5: Use the trained diagnostic model to diagnose and classify faults in the target device.
[0007] Furthermore, in step 3, the residual generative adversarial network module includes a generator and a discriminator. The generator is used to receive the simulation signal and learn to generate the residual signal, and superimpose the residual signal with the input simulation signal to generate an enhanced simulation signal that approximates the real signal in statistical characteristics.
[0008] Furthermore, in step 3, the domain adaptive network module includes a dual-branch feature extractor for extracting time-domain and frequency-domain features, a domain discriminator containing a gradient inversion layer, and a classifier; the feature extractor is used to fuse multimodal features of enhanced simulation data and real data, the domain discriminator is used to drive the feature extractor to learn domain-invariant features, and the classifier is used to determine the fault type based on the learned features.
[0009] Furthermore, the training of the residual generative adversarial network module adopts an adversarial training mechanism, wherein the optimization objective of the generator is to minimize the discriminator's ability to distinguish between the enhanced simulated signal and the real signal, and the optimization objective of the discriminator is to maximize its ability to distinguish between the two signals.
[0010] Furthermore, the generator employs an encoder-decoder architecture based on a deep convolutional network and includes residual connections.
[0011] Furthermore, in the domain adaptive network module, the gradient inversion layer reverses the gradient direction during the backpropagation process of model training, so as to enable the feature extractor to extract domain-invariant features that are discriminative to both the source and target domains.
[0012] Furthermore, the specific process of step 4 is as follows: 4.1. Enhance the input simulation data using a generator to generate enhanced simulation data; 4.2 Input the enhanced simulation data and labeled real data into the domain adaptive network module, extract the time domain and frequency domain features of the enhanced simulation data and real data respectively, and then fuse them; 4.3 Based on the fused features, the diagnostic model is jointly optimized using a multi-objective approach, combining the domain discrimination loss generated by the domain discriminator, the task classification loss generated by the classifier, and the center loss used to optimize intra-class clustering.
[0013] Furthermore, in step 4.3, the total loss function of the multi-objective joint optimization is composed of three terms: task classification loss, domain discrimination loss, and center loss, which are weighted and summed with preset weight coefficients.
[0014] Furthermore, in step 1, the acquired raw signal data includes at least one of phase current, phase voltage, rotational speed, and excitation current.
[0015] Furthermore, in step 2, the fault types simulated by the finite element model of the synchronous motor include at least one of the following: single-phase short circuit, two-phase short circuit, excitation current fault, excitation current disconnection, phase loss, and normal state.
[0016] By employing the above-described technology, the beneficial effects of the present invention compared to the prior art are as follows: 1) In this invention, in terms of data generation, a residual generative adversarial network is used. The generator takes the simulation data as input, learns and adds residual signals that reflect the noise and distribution characteristics of the actual working conditions, thereby generating an enhanced signal that approximates the real measurement data in terms of statistical distribution. This significantly reduces the distribution difference between the simulation data and the real data, and generates high-quality enhanced data with physical authenticity. 2) In this invention, in terms of feature extraction, a multimodal feature extraction mechanism is adopted, which effectively integrates the time-domain statistical features and frequency-domain structural features of the signal, thereby enhancing the discriminative ability of the extracted features; 3) In this invention, in terms of model generalization, a domain adaptive network containing a gradient reversal layer is introduced, and task classification loss, domain discrimination loss and center loss are jointly optimized to drive the feature extractor to learn domain-invariant features that are insensitive to the data source. This mechanism enables the model to be jointly trained using a large amount of labeled augmented simulation data and a small amount of labeled real data, so that high-precision fault classification can still be achieved under the condition that the labeled samples in the target domain are scarce. 4) In this invention, the data augmentation module and the domain adaptive diagnosis module are jointly trained and optimized in the overall architecture. This method not only expands the training set by generating high-quality augmented data, but also further improves the diagnostic robustness and accuracy of the model under cross-domain conditions. Attached Figure Description
[0017] Figure 1This is a schematic diagram of the overall process of the fault diagnosis method in an embodiment of the present invention; Figure 2 In this embodiment of the invention, the finite element model of the synchronous motor is a single unit; Figure 3 This is a schematic diagram of the training set and test set in an embodiment of the present invention; Figure 4 This is a diagram illustrating a normal implementation scheme in an embodiment of the present invention; Figure 5 This is a schematic diagram illustrating an embodiment of the phase loss situation in this invention. Figure 6 This is a schematic diagram of an embodiment of a single-phase short circuit in this invention. Figure 7 This is a schematic diagram of an embodiment of the two-phase short circuit in this invention; Figure 8 This is a schematic diagram illustrating an embodiment of the excitation current disconnection situation in this invention. Figure 9 This is a schematic diagram of an implementation scheme for an excitation current fault condition in an embodiment of the present invention; Figure 10 This is a schematic diagram of the residual generative adversarial network structure in an embodiment of the present invention; Figure 11 This is a schematic diagram of the domain adaptive diagnostic network structure in an embodiment of the present invention; Figure 12 This is a schematic diagram of the confusion matrix of different diagnostic methods on the test set in an embodiment of the present invention; Figure 12 (a) In the process of training the model, a small amount of real data is used as the training dataset to train the confusion matrix obtained by the diagnostic method based on the convolutional neural network; Figure 12 (b) The confusion matrix is obtained by training an adversarial domain adaptation method using labeled original simulation data and unlabeled real data as the source and target domains; Figure 12 (c) The confusion matrix is obtained by training a difference-based domain adaptation method using labeled simulation data and unlabeled real data as the source and target domains; Figure 12 (d) The extended dataset is generated using the Vassatan Generative Adversarial Network. The generated extended dataset is used as the source domain and the real data is used as the target domain to train the confusion matrix obtained by the adversarial domain adaptation method. Figure 12 (e) To generate an extended dataset using a deep convolutional generative adversarial network, the generated extended dataset is used as the source domain and the real data is used as the target domain to train the confusion matrix obtained by the adversarial domain adaptation method. Figure 12 (f) is the confusion matrix obtained within the framework of the method proposed in this paper. Detailed Implementation
[0018] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.
[0019] Conversely, this invention encompasses any substitutions, modifications, equivalent methods, and solutions made within the spirit and scope of the invention as defined in the claims. Furthermore, to provide a better understanding of the invention, certain specific details are described in detail below. However, those skilled in the art will fully understand the invention even without these detailed descriptions.
[0020] Example: A fault diagnosis method based on residual generative adversarial networks and domain adaptation. For ease of explanation, this example uses electrical fault diagnosis in a motor as an example. See also... Figure 1 The method includes the following steps: S1. Collect multi-channel time-domain signals (such as phase current, phase voltage, speed, and excitation current) during motor operation, organize them according to fault categories, and convert them into frequency domain spectrum diagrams through Fourier transform.
[0021] S2. Construct a finite element model of the synchronous motor using Ansys Maxwell. The simulation model is as follows: Figure 2 As shown, the fault settings are as follows: Figures 4-9 As shown, different fault types are set, labeled simulation data is generated, and then converted into a frequency domain spectrum using Fourier transform.
[0022] S3. Construct a diagnostic model that includes a residual generative adversarial network and a domain adaptive network, wherein: (1) Residual Generative Adversarial Network: The generator in a Residual Generative Adversarial Network (RGAN) employs a deep convolutional encoder-decoder architecture, achieving domain adaptation from simulated signals to real signals through a residual learning mechanism. The network structure is as follows: Figure 10 As shown, the generator receives the simulated signal as input and learns to add appropriate noise components through nonlinear transformation, so that the generated signal approximates the real measured signal in statistical characteristics while maintaining the original fault characteristics.
[0023] Random noise input and splicing: (1); (2); in For simulation data, It is random noise. It is the signal learned by the generator.
[0024] The discriminator's loss: Binary cross-entropy loss function: (3); in The total number of samples, For the first The true label of each sample When the sample is a real sample, the value is 1. The value is 0 when generating samples. The model predicts the first The probability that a sample belongs to the positive class.
[0025] The total loss of the discriminator is the sum of the cross-entropy losses of the real data and the generated data. (4); in Represents the discriminator. Represents generator, These are real data. It is simulation data.
[0026] The generator's loss: Adversarial loss: This part of the loss makes the generator try to "fool" the discriminator, causing the discriminator to output a value close to 1 for the generated data. (5); The domain-adaptive network fuses bi-branch convolutional features and combines domain discrimination loss and center loss to optimize feature alignment.
[0027] (2) Domain Adaptive Diagnostic Network: This embodiment proposes a domain adaptive diagnostic network, the network structure of which is as follows: Figure 11 As shown, after generating a large amount of enhanced simulation data through a deep residual connection generator, the generated labeled enhanced simulation data is realigned with a small amount of supervised real data, thereby improving the accuracy of the diagnostic model. The domain adaptive diagnostic network structure proposed in this embodiment mainly consists of a feature extractor, a domain discriminator, and a label classifier.
[0028] The neighborhood discrimination loss function for batch t is: (6); The number of samples in the source domain. The number of samples in the target domain. For the source domain Features of each sample For the first Features of each target domain sample.
[0029] Determine network loss: Loss function for batch t of tasks: (7); in For the sample In category The real labels on it Predict samples for the model For category The probability, To increase the total number of simulation data samples, This represents the total number of categories.
[0030] Central loss function: The initialization of the categorical feature centroids is achieved by random generation using a standard normal distribution, and can be represented as follows: During training, the class center points are updated according to formula (8): (8); in, To update the weights, To update to the current version Batch number The global center of the class To update to the current version Batch No. The global center of the class For the first Batch No. The feature mean of all samples in the class.
[0031] The formula for calculating the center loss of batch t is: (9); in For the first sample in the source domain Batch No. The feature vector of each sample For the first sample in the target domain Batch No. The feature vector of each sample. and These are indicator functions, where and They are respectively and The corresponding tags It is the first The feature center of the class The total number of categories, and These represent the number of samples in the source and target domain batches, respectively.
[0032] The overall loss function is: (10); in It is the weight of the domain loss. It is the weight of the center loss.
[0033] S4. The experiment was conducted in Jupyter Notebook software, with GPU acceleration for training. The optimizer used was Adam, the base learning rate for the feature extractor and classifier was 0.001, the learning rate for the domain discriminator was 0.0001, the training epochs were 50, and each experiment was repeated 5 times to avoid experimental randomness.
[0034] S5. Input the test set into the trained model, output the fault type prediction results, and evaluate the model performance based on accuracy, precision, recall, and F1 score. To verify the accuracy of the fault diagnosis method based on residual generative adversarial network and domain adaptation in this embodiment, this embodiment selects the training dataset of Northeast Electric Power University and the simulation data generated by the finite element model as the training dataset of the model, and selects the test dataset of Northeast Electric Power University for experimental verification. The Northeast Electric Power University dataset is shown in Table 1. Data for six health conditions were selected, including single-phase short circuit, two-phase short circuit, excitation current fault, excitation current disconnection, phase loss, and normal state. Each fault includes data in four dimensions: phase voltage, phase current, excitation current, and speed, for classification tasks. The fault diagnosis method based on residual generative adversarial network and domain adaptation proposed in this embodiment is used for diagnosis and classification. The accuracy of different diagnosis methods on the Northeast Electric Power University dataset is shown in Table 2. The proposed method achieves an average accuracy of 99.75%, compared to 53.54% for CNN models. The average accuracy in adversarial domain adaptive networks and differential domain adaptive networks is 80.34% and 88.78%, respectively, while WGAN and DCGAN achieve average accuracies of 91.76% and 93.28%, respectively. Compared to traditional diagnostic methods, the proposed method demonstrates improvements in accuracy, precision, recall, and F1 score. The effectiveness of residual generative adversarial networks and comprehensive diagnostic networks is also demonstrated. Table 1. Description of the dataset from Northeast Electric Power University health status Training samples Test samples Phase loss fault 20×4000 80×4000 Single-phase short circuit fault 20×4000 80×4000 Two-phase short circuit fault 20×4000 80×4000 Rotor current fault 20×4000 80×4000 Rotor current disconnection fault 20×4000 80×4000 health status 20×4000 80×4000 Table 2. Accuracy of different diagnostic methods on the Northeast Electric Power University dataset. method accuracy Accuracy Recall rate F1 score CNN 53.54% 59.14% 53.54% 53.64% DANN 83.34% 75.00% 83.34% 77.78% Domain Adaptation 88.78% 87.53% 87.08% 87.59% WGAN 91.76% 90.31% 90.83% 90.49% DCGAN 93.28% 92.71% 92.50% 92.80% Proposed 99.75% 99.75% 99.75% 99.75% .
[0035] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A fault diagnosis method based on residual generative adversarial networks and domain adaptation, characterized in that, Includes the following steps: Step 1: Collect the actual raw signal data of the faulty equipment and preprocess it. Extract the time domain features of the preprocessed data and perform Fourier transform on the data to generate a spectrum. Step 2: Establish a finite element model of the synchronous motor, simulate various fault behaviors, generate tagged simulation signal data, and perform the same preprocessing operation on the simulation data as in Step 1. Step 3: Construct a diagnostic model that includes a residual generative adversarial network module and a domain adaptive network module; Step 4: Based on the simulation data generated in Step 2 and the real data obtained in Step S1, jointly train and optimize the diagnostic model; Step 5: Use the trained diagnostic model to diagnose and classify faults in the target device.
2. The fault diagnosis method based on residual generative adversarial networks and domain adaptation according to claim 1, characterized in that, In step 3, the residual generative adversarial network module includes a generator and a discriminator. The generator is used to receive the simulation signal and learn to generate the residual signal. The residual signal is superimposed on the input simulation signal to generate an enhanced simulation signal that approximates the real signal in terms of statistical characteristics.
3. The fault diagnosis method based on residual generative adversarial networks and domain adaptation according to claim 2, characterized in that, In step 3, the domain adaptive network module includes a dual-branch feature extractor for extracting time-domain and frequency-domain features, a domain discriminator containing a gradient inversion layer, and a classifier. The feature extractor is used to fuse multimodal features of enhanced simulation data and real data. The domain discriminator is used to drive the feature extractor to learn domain-invariant features. The classifier is used to determine the fault type based on the learned features.
4. The fault diagnosis method based on residual generative adversarial networks and domain adaptation according to claim 2, characterized in that, The training of the residual generative adversarial network module adopts an adversarial training mechanism. The optimization objective of the generator is to minimize the discriminator's ability to distinguish between the enhanced simulated signal and the real signal, while the optimization objective of the discriminator is to maximize its ability to distinguish between the two signals.
5. The fault diagnosis method based on residual generative adversarial networks and domain adaptation according to claim 4, characterized in that, The generator employs an encoder-decoder architecture based on a deep convolutional network and includes residual connections.
6. The fault diagnosis method based on residual generative adversarial networks and domain adaptation according to claim 3, characterized in that, In the domain adaptive network module, the gradient inversion layer reverses the gradient direction during the backpropagation process of model training, so as to enable the feature extractor to extract domain-invariant features that are discriminative to both the source and target domains.
7. The fault diagnosis method based on residual generative adversarial networks and domain adaptation according to claim 3, characterized in that, The specific process of step 4 is as follows: 4.
1. Enhance the input simulation data using a generator to generate enhanced simulation data; 4.2 Input the enhanced simulation data and labeled real data into the domain adaptive network module, extract the time domain and frequency domain features of the enhanced simulation data and real data respectively, and then fuse them; 4.3 Based on the fused features, the diagnostic model is jointly optimized using a multi-objective approach, combining the domain discrimination loss generated by the domain discriminator, the task classification loss generated by the classifier, and the center loss used to optimize intra-class clustering.
8. The fault diagnosis method based on residual generative adversarial networks and domain adaptation according to claim 7, characterized in that, In step 4.3, the total loss function of the multi-objective joint optimization is composed of three terms: task classification loss, domain discrimination loss, and center loss, which are weighted and summed with preset weight coefficients.
9. The fault diagnosis method based on residual generative adversarial networks and domain adaptation according to claim 1, characterized in that, In step 1, the acquired raw signal data includes at least one of phase current, phase voltage, rotational speed, and excitation current.
10. The fault diagnosis method based on residual generative adversarial networks and domain adaptation according to claim 1, characterized in that, In step 2, the fault types simulated by the synchronous motor finite element model include at least one of the following: single-phase short circuit, two-phase short circuit, excitation current fault, excitation current disconnection, phase loss, and normal state.