Small sample domain generalization fault diagnosis method based on orthogonal feature decoupling

Abstract

The invention relates to a small sample domain generalization fault diagnosis method based on orthogonal feature decoupling, and belongs to the technical field of mechanical fault diagnosis. According to the method, the problems of poor diagnosis model generalization ability and low decision reliability caused by domain offset and scarce labeled samples of the rotating machinery under complex and changeable working conditions are solved. According to the technical scheme, the method comprises the steps of collecting and preprocessing vibration data, decoupling features through an orthogonal projection module to separate fault essence from working condition interference, enhancing feature domain invariance in combination with a domain discriminator, optimizing small sample adaptation through a task self-adaptive meta-learning framework, and obtaining the vibration data. And synchronously outputting a fault category and predicting uncertainty through an evidence deep learning classifier. According to the method, efficient adaptation and high-reliability diagnosis can be realized under the condition that the target domain is invisible and only a small number of samples are used, and the cross-domain generalization ability and the engineering practical value of the model are remarkably improved.

Description

Technical Field

[0001] This invention belongs to the field of mechanical fault diagnosis technology, and relates to a small sample domain generalization fault diagnosis method based on orthogonal feature decoupling. Background Technology

[0002] Rotating machinery, as core equipment in modern industrial production, directly affects the continuity and safety of the production system. Key components such as bearings, gearboxes, and motors are prone to various faults during long-term operation, making fault diagnosis technology a crucial aspect of equipment health management. In recent years, with the rapid development of deep learning technology, data-driven fault diagnosis methods have made significant progress. However, two major bottlenecks remain to be addressed in real-world industrial scenarios.

[0003] On the one hand, rotating machinery typically operates under complex and variable conditions. Factors such as load fluctuations, speed adjustments, and temperature changes cause fault data to exhibit significant multi-source domain distribution characteristics. Different operating conditions correspond to different data distribution domains, while traditional deep learning models heavily rely on the fundamental assumption that training and test data are distributed in the same way. When the model is deployed to an unseen target domain, its generalization ability drops sharply due to domain shift. Figure 1 This paper clearly demonstrates the fundamental difference between traditional domain adaptation methods and the few-sample domain generalization method involved in this invention in fault diagnosis applications, highlighting the technical challenges in scenarios where the target domain is not visible.

[0004] On the other hand, obtaining high-quality fault samples presents significant challenges. Fault data collection often requires system downtime, and data labeling heavily relies on the knowledge of specialized engineers and sophisticated testing equipment. This results in only a small number of labeled samples being available in most real-world scenarios—a situation known as small sample conditions. Models trained with limited samples are highly susceptible to overfitting and struggle to learn stable and discriminative fault characteristics from the data.

[0005] Existing fault diagnosis methods have significant shortcomings in addressing the aforementioned challenges. First, regarding feature decoupling, traditional methods lack strict orthogonal constraint mechanisms, leading to information redundancy between domain-invariant common features and domain-specific unique features, failing to completely separate the essential fault features from the interference of operating condition differences. Second, adaptation strategies for few-sample learning are relatively rigid. Traditional meta-learning methods, such as Model-Agnostic Meta-Learning (MAML), typically employ fixed inner layer learning rates and adaptation steps, unable to dynamically adjust based on the specific differences between one or five example tasks, resulting in a trade-off between adaptation efficiency and final generalization performance. Finally, existing models lack decision reliability. Most models only output fault category labels, lacking a quantitative assessment of the uncertainty of the prediction results themselves. In industrial applications, a single misjudgment can lead to unplanned downtime, production interruptions, or even safety accidents, making the need for decision reliability extremely urgent.

[0006] In summary, there is an urgent need in industrial settings for a domain generalization method capable of achieving high-precision and high-reliability fault diagnosis using only a very small number of target domain samples when target domain data is not visible. This method needs to effectively decouple features, intelligently adapt to small-sample tasks, and provide uncertainty quantification to enhance its practical value in real-world, complex industrial environments. Figure 2 The overall training steps of the fault diagnosis model of this invention are shown, laying the foundation for subsequent detailed explanations. Summary of the Invention

[0007] In view of this, the purpose of this invention is to provide a small sample domain generalization fault diagnosis method based on orthogonal feature decoupling.

[0008] To achieve the above objectives, the present invention provides the following technical solution: Firstly, this invention proposes a multi-source domain small-sample domain generalization fault diagnosis method based on orthogonal feature decoupling, including: Vibration data of key components of rotating machinery are collected, and after data preprocessing, they are input into the trained fault diagnosis model to obtain the fault diagnosis results of rotating machinery. The method is divided into a meta-training phase and a meta-testing phase, and the specific process is as follows: I. Meta-training Phase S1: Accelerometers are used to collect time-domain vibration signals of key mechanical components under multiple operating conditions. First, the 1024-point time-domain signal is converted into 512-dimensional frequency domain features by FFT transformation. Then, the mean and standard deviation are calculated based on the source domain data to complete the standardization. Finally, the data preprocessing is completed.

[0009] S2: The model training sample set is built based on the preprocessed multi-source domain labeled data. Unseen working condition data is used as the target domain and does not participate in any training process.

[0010] S3: A 1D convolutional neural network adapted to the characteristics of frequency domain feature sequences is used to extract high-dimensional features from the source domain preprocessed data. After convolution, fully connected and other layer operations, the source domain fusion features are obtained.

[0011] S4: Decouple the source domain fusion features using the orthogonal projection module. Construct the orthogonal projection matrix through QR decomposition to obtain domain-invariant common features (representing the essence of the fault) and domain-specific unique features (representing the differences in operating conditions). Simultaneously calculate the orthogonal constraint loss to minimize feature redundancy and quantify the decoupling effect through the CSFR index.

[0012] S5: Introduce a gradient reversal layer and a dual-domain discriminator to perform domain invariance enhancement training on common features, making it difficult for the common features to be distinguished by the domain discriminator. At the same time, calculate the domain discrimination loss of common features and unique features.

[0013] S6: Based on the task-adaptive meta-learning framework, a meta-task containing a support set and a query set is sampled from the source domain data. Differentiated inner layer learning rates and adaptation steps are set for 1-shot / 5-shot tasks respectively. Inner layer adaptation is completed using the source domain support set. At the same time, Gaussian noise enhancement is applied to the source domain support set and query set to alleviate small sample overfitting.

[0014] S7: Using the decoupled common features as input, the evidence vector and Dirichlet distribution parameters are generated through the evidence deep learning classifier. The fault category confidence and prediction uncertainty are output simultaneously, and the classification loss and uncertainty loss are calculated respectively.

[0015] S8: The total loss function is constructed by fusing classification loss, domain discrimination loss, orthogonal constraint loss and uncertainty loss. The Adam optimizer is used to iteratively optimize the model network parameters through backpropagation. After 400 rounds of training and early stopping strategy screening, meta-training is completed and the general parameters of the model are saved.

[0016] II. Meta-testing Phase S9: The preprocessing method (FFT + normalization) for reusing unlabeled data in the target domain during the meta-training stage is used, and then the data is divided into 1-shot / 5-shot target domain support sets (1 / 5 samples per class) and target domain query sets according to the task type.

[0017] S10: Load the general parameters of the model obtained from meta-training, and use the inner adaptation strategy of the meta-training stage to complete the model parameter adaptation using the target domain support set. The adaptation process does not affect the general parameters of meta-training, and finally the adapted model parameters are obtained.

[0018] S11: Input the target domain query set into the adapted model, and output the fault category and corresponding prediction uncertainty results through the model to complete the fault diagnosis of rotating machinery in the target domain.

[0019] In a second aspect, the present invention proposes an electronic device, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the multi-source domain small sample domain generalization fault diagnosis method based on orthogonal feature decoupling described in the first aspect of the present invention.

[0020] Thirdly, the present invention proposes a computer-readable storage medium storing a computer program thereon, wherein the computer program, when controlled by a processor, implements the multi-source domain small sample domain generalization fault diagnosis method based on orthogonal feature decoupling described in the first aspect of the present invention.

[0021] The beneficial effects of this invention are as follows: This invention achieves a breakthrough in the field of rotating machinery fault diagnosis by integrating advanced technologies such as orthogonal feature decoupling, task-adaptive meta-learning, and evidence-based deep learning. Its beneficial effects are significant and multi-layered. First, this invention strictly ensures that target domain data does not participate in model training throughout the entire process, fully conforming to the actual constraint that the target domain is invisible in industrial settings. This effectively avoids distortion of model generalization performance caused by data leakage, resulting in higher reliability and practicality of the diagnostic results. Figure 1 As shown, traditional domain adaptation methods rely on target domain data, while the small-sample domain generalization paradigm of this invention fundamentally solves the distribution shift problem.

[0022] The orthogonal feature decoupling mechanism, through the synergistic effect of QR decomposition and orthogonal constraint loss, achieves a complete separation between domain-invariant common features and domain-specific unique features. This orthogonalization process minimizes information redundancy between features, protecting the essential characteristics of faults from interference by differences in operating conditions, thereby improving the purity and discriminative power of feature representation. Figure 4 As shown, the residual value of the unique characteristic of the working condition, CSFR, is significantly reduced, verifying the effectiveness of the decoupling effect.

[0023] Regarding few-shot adaptation, the task-adaptive meta-learning framework addresses the dynamic needs of one-shot and five-shot tasks by differentiating the inner learning rate and adaptation steps, thus resolving the rigidity issue of traditional meta-learning strategies. In the meta-testing phase, only a very small number of target domain support set samples (e.g., 1 or 5 samples per class) are required to quickly complete model adaptation, resulting in high adaptation efficiency and strong generalization ability. Accuracy exceeds 95% in multiple tasks, reaching 96.77% for A3 and 97.69% for A4.

[0024] The reliability of decision-making is significantly enhanced. The evidence-based deep learning classifier not only outputs fault category predictions but also simultaneously provides quantification of prediction uncertainty. This dual-output mechanism provides interpretable references for industrial decision-making, achieving an accuracy rate of up to 95.87% for low-uncertainty samples, effectively reducing the risk of misjudgment and improving system safety. Figure 5 As shown, the diagonal values ​​of the confusion matrix are close to 1.00, indicating that the prediction results are highly reliable.

[0025] The method exhibits excellent domain generalization ability, thanks to the introduction of the gradient inversion layer (GRL) and the domain discriminator, which enhances the domain invariance of common features, allowing the model to maintain stable performance even under unseen target domain conditions. Furthermore, the method has broad engineering applicability, suitable for various rotating machinery such as bearings, gearboxes, and motors. It requires no target domain training data, achieving rapid adaptation with only a small number of support set samples, significantly reducing the cost and barriers to industrial applications.

[0026] Other advantages, objectives, and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination, or may be learned from practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description

[0027] To make the objectives, technical solutions, and advantages of the present invention clearer, the preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings, wherein: Figure 1 A schematic diagram comparing traditional domain adaptation and small-sample domain generalization in rotating machinery fault diagnosis. Figure 2 This is a flowchart of the training steps for the rotating machinery fault detection model in an embodiment of the present invention; Figure 3 This is a flowchart illustrating an embodiment of the present invention; Figure 4 Summarize the CSFR values ​​of each experimental group under different tasks; Figure 5 This is a schematic diagram of the confusion matrix between the predicted labels and the true labels of unlabeled source domain data by the diagnostic model in one of the experimental verification phases of this invention. Detailed Implementation

[0028] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0029] The accompanying drawings are for illustrative purposes only and are schematic diagrams, not actual pictures. They should not be construed as limiting the invention. To better illustrate the embodiments of the invention, some parts in the drawings may be omitted, enlarged, or reduced, and do not represent the actual product dimensions. It is understandable to those skilled in the art that some well-known structures and their descriptions may be omitted in the drawings.

[0030] In the accompanying drawings of the embodiments of the present invention, the same or similar reference numerals correspond to the same or similar components. In the description of the present invention, it should be understood that if terms such as "upper," "lower," "left," "right," "front," and "rear" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, they are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, the terms used to describe positional relationships in the drawings are only for illustrative purposes and should not be construed as limiting the present invention. For those skilled in the art, the specific meaning of the above terms can be understood according to the specific circumstances.

[0031] The method of this invention is divided into a meta-training phase and a meta-testing phase, and the specific process is as follows: I. Meta-training Phase S1. Vibration data of key components of rotating machinery under multiple different operating conditions are collected using sensors, and the vibration data is preprocessed. Specifically, accelerometers are used to collect the time-domain vibration signals of key components such as bearings and gearboxes, with a sampling frequency of 12kHz and a sample length of 1024 points. Data preprocessing includes two steps: FFT transformation and standardization. (1) FFT Transform: Perform a Fast Fourier Transform (FFT) on the 1024-point time-domain signal to convert the time-domain signal into frequency-domain features. Due to the conjugate symmetry of the FFT result of the real-number signal, only the first half (0~6kHz) of the amplitude spectrum is retained, resulting in 512-dimensional frequency-domain features, as shown in the formula: Frequency domain features can highlight the characteristic frequencies corresponding to the fault and reduce time domain noise interference. Based on expert knowledge and experience, the collected vibration data are manually labeled to obtain the fault type and its label. The fault type is usually also called the fault category. (2) Standardization: The mean is calculated based only on the source domain data. with standard deviation The formula is: To avoid the impact of dimensional differences on model training, the target domain data is reused in subsequent preprocessing. and This ensures consistency in preprocessing.

[0032] The gearbox fault dataset used in this example was collected from an experimental test bench under three different operating conditions (1200 RPM, 1800 RPM, and 2400 RPM) and three different motor loads (0 HP, 4 HP, and 8 HP). The test bench consisted of a motor, a planetary gearbox, a parallel shaft gearbox, and a magnetic actuator. Seven gearbox faults were simulated: healthy (H), inner race fault (IRF), outer race fault (ORF), ball fault (BF), gear eccentricity fault (GE), tooth root crack fault (GRC), and gear tooth breakage fault (GTB). Each fault type had 200 original fault samples of 1024 points each. The gearbox fault dataset was used to construct a partial-domain adaptive fault diagnosis task for partially labeled data. After labeling, the fault types and their labels for the gearbox fault dataset were obtained, as shown in Table 1.

[0033] Table 1. Fault types and their labels in the gearbox fault dataset.

[0034] S2. Based on the preprocessed vibration data, a training sample set is obtained. The source domain data consists of labeled vibration data, while the target domain data does not participate in model training. Specifically, vibration data under multiple known working conditions are used as the source domain (e.g., Condition_0,1,3,6,7); unseen working condition data are used as the target domain (e.g., Condition_5,8). All target domain data is unlabeled and does not participate in any training process. The training sample set contains only labeled data for model meta-training.

[0035] After empirical experiments and annotation, the source domain data and target domain data of the gearbox fault dataset are shown in Table 2.

[0036] Table 2 shows the source and target domain data of the gearbox fault dataset.

[0037] In this embodiment of the invention, the rotating machinery fault detection model is based on a deep neural network and includes a domain fault feature extractor, an orthogonal projection module, a domain discriminator, and an evidence classifier.

[0038] For example, the network parameters of the rotating machinery fault detection model are shown in Table 3.

[0039] Table 3 Network parameters of the rotating machinery fault detection model

[0040] S3. High-dimensional features are extracted from the source domain data using the feature encoding module to obtain the source domain data features. The feature encoding module adopts a 1D convolutional neural network to adapt to the sequence characteristics of frequency domain features. The specific structure is as follows: Input layer: (Number of channels: 1; Sample length: 512); Convolutional layer 1: 32 output channels, kernel size 5, stride 2, padding=2, followed by BatchNorm, ReLU and Dropout (0.3). Convolutional layer 2: 64 output channels, kernel size 5, stride 2, padding=2, followed by BatchNorm, ReLU and Dropout (0.3). Fully connected layer 1: Input dimension 64×128, output dimension 256, followed by BatchNorm and Dropout (0.3). Fully connected layer 2: Output dimension 128, obtaining source domain fusion features. .

[0041] S4. Decouple the source domain data features using an orthogonal projection module to obtain domain-invariant common features and domain-specific unique features. The orthogonal projection module constructs an orthogonal projection matrix through QR decomposition and minimizes the information redundancy of common and unique features by combining orthogonal constraint loss. Specific implementation steps: S41. Initialize the random matrix ,satisfy ; S42, to Performing QR decomposition yields an orthogonal matrix Q and an upper triangular matrix R, satisfying... ,and ; S43. Divide Q into blocks with a common projection basis. With unique projection base ,and ; S44. The decoupling characteristic formula obtained from the projection operation is: , ; In the formula, As a common feature, As a unique feature, For a common projection basis, It is a unique projection base.

[0042] S45. The formula for calculating the orthogonal constraint loss is:

[0043] In the formula, N is the number of samples in the source domain. As a common feature, As a unique feature, the two types of features exhibit the strongest orthogonality and the lowest information redundancy when the loss is minimized.

[0044] S46. Calculate the Condition-Specific Residual Value (CSFR) to quantify the decoupling effect: CSFR measures the amount of condition-specific information remaining in the common features after decoupling. The smaller the value, the less condition interference and the purer the essential fault features in the common features. Specifically, for each fault category k and each source domain d, first calculate the common features of that fault category in that source domain. variance and original fusion features variance Next, calculate the average of the variance ratios for all combinations of (source domain d, fault k), which is the CSFR. The formula is:

[0045] In the formula, where For operating conditions, Number of fault categories This represents the characteristic variance under different operating conditions for the same fault category. A smaller CSFR indicates that the unique characteristics of the operating condition are more pronounced. The less residue in the material, the better the peeling effect.

[0046] In this embodiment, (5 source domains) (7 fault categories) The calculated CSFR value of task A1 of this method is only 0.28, which verifies the effectiveness of orthogonal decoupling.

[0047] S5. Introduce a gradient inversion layer and a domain discriminator, perform domain invariance enhancement training on common features, and calculate the domain discriminant loss; Design of the gradient inversion layer (GRL): identity mapping during forward propagation. Gradient inversion during backpropagation Domain discriminators include common feature domain discriminators. Discriminator with unique feature domain All of them use a 2-layer fully connected network and are trained only on source domain data.

[0048] The formula for calculating the domain discrimination loss is:

[0049]

[0050] Where d is the domain label of the source domain data. As a common feature, It is a unique feature.

[0051] GRL forces common features to be difficult to distinguish by the domain discriminator, thereby learning domain-invariant features across source domains and laying the foundation for target domain generalization.

[0052] S6. Based on the task-adaptive meta-learning framework, differentiated inner layer learning rates and adaptation steps are designed for 1-shot / 5-shot tasks. Inner layer adaptation is performed using the source domain support set, and a Gaussian noise enhancement strategy is combined to alleviate overfitting in small samples. The core process of the task-adaptive meta-learning framework is: Meta-task sampling: Randomly sampling meta-tasks from the source domain data. , For the support set (1-shot / 5-shot samples), For the query set (5 samples / class); inner adaptation (supports set optimization): dynamically adjusts the learning rate and number of steps according to the task type: 1-shot task: inner learning rate Adapted to 2 steps, updated parameters. 5-shot task: Inner layer learning rate Adapted to 4 steps, parameters updated. Outer layer update (query set optimization): using meta-learning rate Update meta-parameters by query set loss Gaussian noise enhancement: Enhancement is performed only on the source domain support set / query set, avoiding the involvement of target domain data: Support set enhancement: , (Weak enhancement, preserving core fault features); Query set enhancement: , (Moderate enhancement to simulate changes in operating conditions).

[0053] S7. Utilize an evidence-based deep learning classifier, taking common features as input, to output the fault category prediction result and prediction uncertainty, and guide model training through classification loss and uncertainty loss (this classifier is trained only on source domain data): The formula for generating evidence vectors is:

[0054] in (7 represents the number of fault categories), the Softplus function guarantees... ; Dirichlet distribution parameters: To ensure the legality of the distribution; The formula for calculating category confidence is:

[0055] For evidence and; The formula for predicting uncertainty is: The larger the value of U, the lower the reliability of the prediction. Loss Calculation and Training Guidelines: The classification loss formula is:

[0056] in For the digamma function, One-hot encoding of labeled samples in the source domain is used to optimize classification accuracy; The formula for uncertainty loss is: By penalizing high-uncertainty predictions, it directly participates in the optimization of model parameters, guiding the model to actively reduce prediction uncertainty during training, and achieving collaborative training of "high-precision classification + high-reliability decision-making".

[0057] S8. Construct the total loss function based on orthogonal constraint loss, domain discrimination loss, classification loss and uncertainty loss, optimize the model network parameters through backpropagation, and complete meta-training; The formula for the total loss function is:

[0058] The weight configuration was determined through hyperparameter tuning: For classifying losses, For the domain discriminator loss, For orthogonal constraint loss, The loss is due to uncertainty. The model is trained using the Adam optimizer. , Weight decay The training run consists of 400 rounds, with 150 source domain meta-tasks sampled per round. An early stopping strategy is employed (training stops if the loss does not decrease after 10 consecutive rounds). The general parameters after meta-training are saved. .

[0059] II. Meta-testing phase (the target domain only participates in adaptation and testing) S9. After preprocessing the unlabeled target domain data in step S1, divide the target domain support set (1-shot / 5-shot samples) and the target domain query set (test samples); specifically, the target domain data preprocessing reuses the methods from the meta-training stage. and To ensure data distribution consistency, the data is divided according to task type: the target domain support set for 1-shot tasks contains 1 sample per category, and the target domain query set for 5-shot tasks contains 5 samples per category; these are used to evaluate diagnostic performance.

[0060] S10, Loading the general parameters of the model after meta-training By utilizing the target domain support set, the inner adaptation strategy of step S6 is executed to obtain the adapted model parameters. The adaptation process only updates model parameters and does not involve gradient backpropagation to meta-training parameters, ensuring that target domain data does not affect meta-training results. The adaptation strategy is completely consistent with the meta-training stage (learning rate and number of steps for 1-shot / 5-shot), guaranteeing adaptation consistency.

[0061] S11. Input the target domain query set into the adapted model, output the fault category and prediction uncertainty, and obtain the fault diagnosis result.

[0062] After the rotating machinery fault detection model is trained, the model's network parameters have been optimized. When the model is applied to the actual detection or prediction stage, only the fault feature extractor and evidence classifier are used. The prediction process includes: Vibration data of key components of rotating machinery under multiple different operating conditions are collected using accelerometers, and the vibration data is preprocessed. Features are extracted using the optimized fault feature extractor; The extracted features are input into the optimized fault classifier to obtain the fault diagnosis results of the rotating machinery.

[0063] This invention provides an electronic device, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the rotating machinery fault diagnosis method based on partial domain adaptation described in the first aspect of this invention.

[0064] This invention provides a computer-storable medium storing a computer program thereon, which, when controlled by a processor, implements the rotating machinery fault diagnosis method based on partial domain adaptation described in the first aspect of this invention.

[0065] Experimental verification: To verify the validity of this invention, an experimental verification was conducted using a gearbox fault vibration dataset (DDS) collected from the DrivetrainDiagnostics Simulator (DDS).

[0066] In the Gearbox Fault Dataset (DDS), gearbox fault vibration signals were collected from an experimental test bench at three different speeds (i.e., 30 Hz and 40 Hz). The test bench consisted of an electric motor, a planetary gearbox, a parallel shaft gearbox, and a magnetic actuator. Seven gearbox faults were simulated: healthy (H), inner race fault (IRF), outer race fault (ORF), ball fault (BF), gear eccentricity fault (GE), tooth root crack fault (GRC), and gear tooth breakage fault (GTB). Each fault type had 200 original fault samples of 1024 points each.

[0067] Based on the available gearbox fault data, the deep learning-based rotating machinery fault diagnosis model proposed in this invention (Proposed) is compared with that in the literature. International conference on machine learning The MAML model proposed in "Model-Agnostic Meta-Learning for Fast Adaptation of Deep Networks" (2017), and related literature. Expert Systems With Applications This paper compares the proposed MetaTEDL model with that presented in "Meta transferevidence deep learning for trustworthy few-shot classification" (2025). To ensure fairness in the experiment, all the methods described above use the same feature extractor and classifier structure.

[0068] To obtain stable experimental results and minimize the impact of randomness, each comparison method was tested five times, with 400 model training iterations per test. The fault classification accuracy was compared with the proposed method across eight tasks. Specific fault diagnosis accuracy results are shown in Table 4. The confusion matrix for diagnosis accuracy is also included. Figure 4 Show, Figure 4 In the diagram, the horizontal axis represents the true label (true Lalel) and the predicted Lalel represents the label predicted by the model.

[0069] Table 4. Cross-domain fault diagnosis results of the gearbox fault dataset.

[0070] The experimental results show that, under all tasks, the diagnostic accuracy of MAML, MetaTEDL and the method of this invention is as follows: MAML has an average accuracy of about 79.1%, MetaTEDL has an average accuracy of about 89.3%, while the method of this invention has an average accuracy of 94.7%, which is significantly better under all conditions. The core support for this advantage lies in the precise synergy between the orthogonal projection feature decoupling module and the task-adaptive meta-learning module proposed in this invention. On the one hand, the orthogonal projection feature decoupling module constructs an orthogonal projection matrix through QR decomposition, separating features into domain-invariant common features and domain-specific unique features. Combined with orthogonal constraint loss, it reduces interference residue, thereby making the essential fault features purer under multiple operating conditions. On the other hand, the task-adaptive meta-learning module customizes differentiated inner layer learning rates and adaptation steps for 1-shot / 5-shot tasks, solving the problem of rigidity in traditional meta-learning strategies. With the synergy of the two, high-purity domain-invariant features provide a high-quality foundation for meta-learning adaptation, while precise small-sample adaptation further releases the value of feature decoupling, ultimately achieving comprehensive performance leadership under various target domain operating conditions, directly demonstrating the role of the proposed module in improving diagnostic performance.

[0071] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A small-sample domain generalization fault diagnosis method based on orthogonal feature decoupling, characterized in that: Vibration data of key components of rotating machinery are collected, and after data preprocessing, they are input into a trained fault diagnosis model to obtain fault diagnosis results of rotating machinery. The training process of the fault diagnosis model includes S1 to S8, and the application process of the fault diagnosis model includes S9 to S11. S1: Collect vibration data of key components of rotating machinery under multiple different operating conditions using sensors and perform data preprocessing on the vibration data: S2: A training sample set is obtained based on the vibration data after data preprocessing. The training sample set only includes source domain data, while the target domain data consists entirely of unlabeled vibration data and does not participate in the training. S3: Use the feature encoding module to extract high-dimensional features from the source domain data to obtain the source domain data features; S4: The orthogonal projection module is used to decouple the source domain data features to obtain the common features that are invariant to the domain and the unique features that are specific to the domain. The orthogonal projection module constructs an orthogonal projection matrix through QR decomposition and combines orthogonal constraint loss to minimize the information redundancy of the common features and unique features. S5: Introduce a gradient reversal layer and a domain discriminator to enhance the domain invariance of common features during training and calculate the domain discriminant loss: S6: Based on the task-adaptive meta-learning framework, differentiated inner layer learning rates and adaptation steps are designed for small sample tasks. Inner layer adaptation is performed using the source domain support set and combined with Gaussian noise enhancement strategy to alleviate small sample overfitting. S7: Use an evidence-based deep learning classifier with common features as input to output fault category prediction results and prediction uncertainty, and calculate classification loss and uncertainty loss; S8: Construct a total loss function based on orthogonal constraint loss, domain discrimination loss, classification loss and uncertainty loss, and optimize the model network parameters through backpropagation to complete model training; S9: After preprocessing the unlabeled data of the target domain in S1, divide it into the target domain support set sample and the target domain query set; S10: Load the general parameters of the model after meta-training and use the inner adaptation strategy of S6 to obtain the adapted model parameters using the target domain support set; S11: Input the target domain query set into the model and after adaptation, output the fault category and prediction uncertainty to obtain the fault diagnosis result.

2. The small-sample domain generalization fault diagnosis method based on orthogonal feature decoupling according to claim 1, characterized in that: In step S1, data preprocessing includes FFT transformation and standardization, and the specific process is as follows: (1) Perform a Fast Fourier Transform (FFT) on the vibration time-domain signal to convert the time-domain signal into frequency-domain features, retaining the frequency-domain amplitude spectrum as an effective feature. The formula is as follows: ,in For Fast Fourier Transform operations, For absolute value operations; (2) The vibration time-domain signal is standardized. The standardization process is calculated using the mean and standard deviation of the source domain data, and the formula is as follows: ,in The mean of the source domain data. The standard deviation of the source domain data.

3. The small-sample domain generalization fault diagnosis method based on orthogonal feature decoupling according to claim 1, characterized in that: In S4, the specific implementation of the orthogonal projection module includes: S41. Initialize the random matrix , where D is the dimension of the fused feature output by the feature encoding module; S42, to Performing QR decomposition yields an orthogonal matrix Q and an upper triangular matrix R, satisfying... ,and ; S43. Divide the orthogonal matrix Q into blocks with a common projection basis. With unique projection base Where C + U = D, and ; S44. Decoupling features are obtained through projection operations: , ,in The fused features output by the feature encoding module; S45. Calculate the orthogonal constraint loss: ; S46. Calculate the condition-specific residual value (CSFR) to quantify the feature decoupling effect. The formula is: ,in For the number of source domains, Number of fault categories Indicates the first One source domain, Indicates the first Class of faults, This indicates variance calculation. The original fused features are output by the feature encoding module.

4. The small-sample domain generalization fault diagnosis method based on orthogonal feature decoupling according to claim 1, characterized in that: In step S5, the domain discrimination loss includes the common feature domain discrimination loss and the unique feature domain discrimination loss. The common features are processed by the gradient inversion layer and then input into the domain discriminator. The gradient inversion layer performs an identity mapping during forward propagation and inverts the gradient during backward propagation. The domain discrimination loss is calculated using the following formula: , ,in For a discriminator with a shared feature domain, For a unique feature domain discriminator, d is the domain label. This is for gradient inversion layer operations.

5. The small-sample domain generalization fault diagnosis method based on orthogonal feature decoupling according to claim 1, characterized in that: In S6, the specific implementation of the task-adaptive meta-learning framework includes: S61, 1-shot task configuration: Inner learning rate The number of steps required is 2. S62, 5-shot task configuration: Inner learning rate The number of steps required is 4. S63, Gaussian noise enhancement: Supports set enhancement. Gaussian noise, query set enhancement employs Gaussian noise, the enhancement formula is: , .

6. The small-sample domain generalization fault diagnosis method based on orthogonal feature decoupling according to claim 1, characterized in that: In S7, the specific implementation of the evidence deep learning classifier includes: S71. Generating evidence vectors using a linear layer and the Softplus function: ; S72. Calculate the Dirichlet distribution parameters: ,in For category evidence parameters; S73, Output Category Confidence: Where K is the number of fault categories, For the k-th type of evidence parameters; S74. Calculate the uncertainty of prediction: ; S75. The classification loss uses Dirichlet log-likelihood loss: ,in , For the digamma function, One-hot encoding for the label; S76, Uncertainty Loss: .

7. The small-sample domain generalization fault diagnosis method based on orthogonal feature decoupling according to claim 1, characterized in that: In S8, the total loss function is: ,in , , , .

8. The small-sample domain generalization fault diagnosis method based on orthogonal feature decoupling according to claim 1, characterized in that: The feature encoding module uses a 1D convolutional neural network, which includes a convolutional layer, a batch normalization layer, a ReLU activation function, a Dropout layer and a fully connected layer connected in sequence. The kernel size of the convolutional layer is 5, the stride is 2, and the Dropout rate is 0.

3.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the multi-source domain small sample domain generalization fault diagnosis method based on orthogonal feature decoupling as described in any one of claims 1 to 8.

10. A computer-readable storage medium, characterized in that, It stores a computer program, which, when controlled by a processor, implements the multi-source domain small sample domain generalization fault diagnosis method based on orthogonal feature decoupling as described in any of claims 1 to 8.

Images