A method for fault diagnosis of cross-domain rotating machinery based on physical feature enhancement
By employing a method based on physical feature enhancement and elastic weight solidification, the problems of data scarcity and small sample fit in cross-domain diagnosis of rotating machinery are solved, achieving efficient fault diagnosis and improving diagnostic accuracy and robustness.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SICHUAN UNIVERSITY OF SCIENCE AND ENGINEERING
- Filing Date
- 2026-05-08
- Publication Date
- 2026-06-02
AI Technical Summary
Existing methods for diagnosing rotating machinery faults suffer from problems such as scarce source domain data, insufficient fault samples, and catastrophic forgetting during small sample migration in the target domain when adapting across domains, leading to a decrease in diagnostic accuracy.
A cross-domain fault diagnosis method based on physical feature enhancement is adopted. Enhanced samples that meet physical consistency constraints are generated through multi-dimensional physical feature extraction and difference-aware interpolation enhancement. A lightweight fine-tuning strategy with elastic weight solidification is adopted to achieve efficient model adaptation in the target domain.
It significantly improves the accuracy and generalization ability of rotating machinery fault diagnosis, especially showing stronger robustness and adaptability in cross-operating conditions and small sample scenarios, which is superior to traditional methods.
Smart Images
Figure CN122133041A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of intelligent rotating machinery fault diagnosis technology, and more specifically, to a cross-domain rotating machinery fault diagnosis method based on physical feature enhancement. Background Technology
[0002] Rotating machinery, as a key power equipment in the industrial field, directly affects the safety and stability of the entire production process. Because this type of equipment operates under harsh conditions of high speed and heavy load for extended periods, critical components are highly susceptible to damage. Once a malfunction occurs, it can lead not only to unplanned downtime and significant economic losses, but also potentially serious accidents resulting in personal injury or death. Therefore, research on fault diagnosis technology for rotating machinery is of great practical significance for ensuring the safe and reliable operation of equipment.
[0003] In real-world industrial scenarios, the operating conditions of rotating machinery are often not constant. Changes in operating parameters such as speed and load can lead to significant changes in the data distribution of vibration signals, a phenomenon known as the "cross-domain" problem. Traditional fault diagnosis methods typically assume that training and test data follow the same distribution. When a diagnostic model trained under a specific operating condition (source domain) is directly applied to a new operating condition (target domain), the diagnostic accuracy drops sharply due to the difference in data distribution. To address this issue, transfer learning technology has been introduced into the field of fault diagnosis. By fine-tuning the pre-trained model in the source domain, it adapts the model to the data characteristics of the target domain.
[0004] However, existing transfer learning methods still suffer from the following technical shortcomings when applied to rotating machinery fault diagnosis: First, the quality and completeness of the source domain data directly affect the transfer effect. In actual engineering, rotating machinery is in normal operation most of the time, and fault samples, especially those of certain specific fault types, are extremely scarce, leading to a serious class imbalance problem in the source domain dataset. Traditional oversampling methods, such as simple random copying or SMOTE linear interpolation, ignore the inherent physical laws of rotating machinery vibration signals, and the generated samples often violate the physical consistency constraints of the fault mechanism, making it difficult to effectively improve the model's generalization ability. Second, existing model fine-tuning strategies are insufficiently adaptable when the target domain samples are extremely limited. When only a small number of labeled samples are available under new operating conditions, standard fine-tuning methods uniformly adjust all parameters, which can easily lead to the model overfitting the small number of samples and forgetting the general fault feature knowledge learned in the source domain, i.e., the "catastrophic forgetting" problem. How to effectively retain the valuable fault diagnosis knowledge contained in the source domain while performing cross-domain adaptation is the core challenge currently facing small-sample cross-domain diagnosis.
[0005] Therefore, how to innovate collaboratively from the two dimensions of data augmentation and model transfer, expand scarce fault samples while ensuring the consistency of physical laws, and achieve efficient model adaptation under the condition of a small number of target domain samples, has become a technical problem that urgently needs to be solved by those skilled in the art. Summary of the Invention
[0006] The purpose of this invention is to address the problems of insufficient source domain fault samples and catastrophic forgetting during small sample migration in rotating machinery fault diagnosis under cross-operating conditions. It proposes a cross-domain fault diagnosis method based on physical feature enhancement and elastic weight solidification. This method effectively expands scarce fault features through a sample enhancement strategy under physical consistency constraints and employs elastic weight solidification fine-tuning technology to achieve accurate and efficient transfer of source domain knowledge to the target domain, providing reliable technical support for intelligent fault diagnosis of rotating machinery under varying operating conditions and small sample scenarios.
[0007] The technical solution for achieving the present invention is: a method for fault diagnosis of cross-domain rotating machinery based on physical feature enhancement, comprising the following steps:
[0008] Step S1: Acquire the rotating machinery vibration signals in the source domain and the target domain, wherein the source domain is the reference working condition and the target domain is the new working condition to be diagnosed;
[0009] Step S2: Extract multi-domain physical features from the source domain vibration signal and construct a multi-dimensional physical feature vector;
[0010] Step S3: Based on the physical feature vector, a difference-aware interpolation enhancement method is used to generate enhanced fault samples that meet physical consistency constraints, forming an enhanced source domain dataset;
[0011] Step S4: Pre-train a convolutional neural network as a feature extractor on the augmented source domain dataset to obtain the initial diagnostic model f. θ0 ;
[0012] Step S5: Transfer the initial diagnostic model f θ0 Deployed to the target domain, and utilizing labeled samples collected from the target domain, the model is fine-tuned using a lightweight fine-tuning strategy with fixed elastic weights to obtain a diagnostic model f adapted to the target domain. θ1 ;
[0013] Step S6: Utilize the diagnostic model f θ1 Fault classification and diagnosis are performed on real-time vibration signals in the target domain.
[0014] Furthermore, in step S2, the multi-domain physical feature extraction includes the following sub-steps:
[0015] Step S21: Extract the fault characteristic frequency f of the vibration signal d and its amplitude Ad ;
[0016] Step S22: Extract sideband information from the envelope spectrum and calculate sideband modulation features:
[0017]
[0018] Where N s For the number of sidebands, A i Let be the amplitude of the i-th sideband;
[0019] Step S23: Extract the center frequency f of the resonant demodulation band of the vibration signal. c and bandwidth B;
[0020] Step S24: Extract the impact response attenuation coefficient λ and natural frequency f of the vibration signal. n ;
[0021] Step S25: Extract f d A d , N s , R s , f c , B, λ, f n After dimensional unification and normalization, it is combined with the fault type code c to construct a multidimensional physical feature vector p, with each sample corresponding to a multidimensional physical feature vector.
[0022] Furthermore, in step S3, the difference-aware interpolation enhancement method includes the following sub-steps:
[0023] Step S31: Count the number of samples of each category in the source domain vibration signal, calculate the sample entropy of each category, and identify the category with sample entropy lower than the preset threshold as a scarce fault category;
[0024] Step S32: For each rare fault category, randomly select two samples x from that category. i x j and its corresponding multidimensional physical feature vector p i p j And calculate the Mahalanobis distance d between the two samples. M ;
[0025] Step S33: Based on the Mahalanobis distance described in step S32 and the current number of samples in this category, dynamically determine the interpolation weight α:
[0026] α=1 / (1+(d M / d0) m ×(n i / N) n )
[0027] Where d0 is the baseline distance parameter, m is the distance sensitivity index, and n i N is the current number of samples for this scarce fault category, N is the target number of samples, and n is the scarcity sensitivity index.
[0028] Initial samples are generated through linear interpolation.
[0029] x new = α·x i + (1-α)·x j ;
[0030] Step S34: Dynamically determine the disturbance intensity σ based on the differences in physical characteristics, using the following method:
[0031] σ=σ0×(1+γ×d M )×exp(-β×‖p i -p j ||1)
[0032] Where σ0 is the reference disturbance intensity, γ is the range amplification factor, and d M Let p be the Mahalanobis distance between the two samples, β be the feature difference decay coefficient, and ||p| ... i - p j ||1 represents the L1 norm difference of the physical feature vectors;
[0033] Step S35: Based on the perturbation intensity σ determined in step S34, add anisotropic Gaussian noise to the initial sample generated in step S33 to obtain the enhanced sample;
[0034] Step S36: Perform physical validity verification on the enhanced sample described in step S35. If the preset consistency threshold is met, retain the sample; otherwise, return to step S33 until the verification is passed or the preset maximum number of attempts is reached. If the verification still fails after reaching the maximum number of attempts, return to step S32.
[0035] Step S37: Extract the multidimensional physical feature vector of the sample retained in step S36, and determine whether the Mahalanobis distance with the original sample is less than a preset threshold. If it is less than the threshold, retain the sample; otherwise, return to step S33 until the verification is successful or the preset maximum number of attempts is reached. If the maximum number of attempts is reached and the verification is still unsuccessful, return to step S32.
[0036] Step S38: Repeat steps S32 to S37 until the number of scarce fault category samples reaches the target number of samples N, forming an enhanced source domain dataset.
[0037] Furthermore, in step S4, the pre-trained convolutional neural network includes the following sub-steps:
[0038] Step S41: Construct a one-dimensional convolutional neural network, which includes multiple convolutional layers, batch normalization layers, pooling layers, and a classifier;
[0039] Step S42: Supervised pre-training of the network using the augmented source domain dataset, with the optimization objective being the cross-entropy loss function, to obtain the initial diagnostic model f. θ0 ;
[0040] Step S43: After pre-training is completed, calculate the Fisher information matrix of each parameter in the model to quantify the importance of the parameters to the source domain task.
[0041] Furthermore, in step S5, the lightweight fine-tuning strategy for fixing the elastic weights includes the following sub-steps:
[0042] Step S51: Based on the source domain pre-trained model, calculate the Fisher information matrix of each parameter in the model as a measure of parameter importance;
[0043] Step S52: Based on the Fisher information, freeze all parameters of the feature extraction network and set the classifier parameters to a trainable state;
[0044] Step S53: Dynamically determine the regularization coefficient for each category based on the number of samples in each category of the target domain.
[0045] λ c = λ0×(1+η×(N max / n c ))
[0046] Where λ0 is the baseline regularization coefficient, η is the scarcity sensitivity factor, and N max n is the number of samples from the category with the largest number of samples in the target domain. c Let c be the number of samples in category c;
[0047] Step S54: Dynamically determine the adaptive learning rate lr for each parameter based on the Fisher information of each parameter. i :
[0048] lr i = lr0 / (1+κ×F i / F max )
[0049] Where lr0 is the baseline learning rate, κ is the sensitivity factor, and F... i For the Fisher information of the i-th parameter, F max The largest Fisher information among all parameters;
[0050] Step S55: Construct the joint optimization objective function:
[0051]
[0052] Where L few For small sample classification loss in the target domain, θ i and θ i,old These are the current value of the i-th parameter during the fine-tuning process and the initial value of the parameter after pre-training in the source domain, respectively.
[0053] Step S56: Using a small number of labeled samples from the target domain, with the goal of minimizing the joint optimization objective function, the classifier parameters are optimized using gradient descent and an adaptive learning rate, while the feature extraction network parameters are kept frozen, resulting in a diagnostic model f adapted to the target domain. θ1 .
[0054] The beneficial effects of this invention are:
[0055] 1. By constructing multidimensional physical feature vectors, the failure evolution law of rotating machinery is quantified into a computable feature representation, which lays a solid foundation for subsequent physical consistency constraints and overcomes the limitations of traditional methods that rely solely on data distribution and lack physical interpretability.
[0056] 2. To address the issue of insufficient samples for specific fault categories in the source domain, an innovative dynamic interpolation method based on Mahalanobis distance and physical feature differences was designed, along with a dual verification mechanism from the signal domain to the feature domain. This significantly improves the completeness of the source domain dataset and the generalization ability of the pre-trained model. 3. By quantifying parameter importance using the Fisher information matrix, and based on freezing the key feature extraction layer, a dual adaptive mechanism is introduced: dynamically adjusting the regularization coefficient according to the scarcity of samples for each category in the target domain, and simultaneously allocating differentiated learning rates based on parameter importance. This strategy, through fine-tuning at the parameter level, effectively prevents overfitting with small samples and avoids catastrophic forgetting of key knowledge in the source domain, achieving accurate and efficient adaptation across operating conditions, even with only a small number of target domain samples. 4. A complete technical closed loop of "physical augmentation - source domain pre-training - target domain adaptation" is formed, collaboratively innovating from two dimensions: data augmentation and model transfer. This achieves significant performance improvements compared to existing methods in rotating machinery fault diagnosis tasks with small sample sizes and across operating conditions, demonstrating good engineering application value. Attached Figure Description
[0057] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained from these drawings without creative effort.
[0058] Figure 1 This is a flowchart of a cross-domain rotating machinery fault diagnosis method based on physical feature enhancement provided by the present invention;
[0059] Figure 2 This is a flowchart of the difference-sensing interpolation enhancement method provided by the present invention. Detailed Implementation
[0060] Example:
[0061] like Figure 1 A cross-domain rotating machinery fault diagnosis method based on physical feature enhancement includes the following steps:
[0062] like Figure 1 A cross-domain rotating machinery fault diagnosis method based on physical feature enhancement includes the following steps:
[0063] Step S1: Acquire the rotating machinery vibration signals in the source domain and the target domain, wherein the source domain is the reference working condition and the target domain is the new working condition to be diagnosed;
[0064] Step S2: Extract multi-domain physical features from the source domain vibration signal and construct a multi-dimensional physical feature vector;
[0065] Step S3: Based on the physical feature vector, a difference-aware interpolation enhancement method is used to generate enhanced fault samples that meet physical consistency constraints, forming an enhanced source domain dataset;
[0066] Step S4: Pre-train a convolutional neural network as a feature extractor on the augmented source domain dataset to obtain the initial diagnostic model f. θ0 ;
[0067] Step S5: Transfer the initial diagnostic model f θ0 Deployed to the target domain, and utilizing labeled samples collected from the target domain, the model is fine-tuned using a lightweight fine-tuning strategy with fixed elastic weights to obtain a diagnostic model f adapted to the target domain. θ1 ;
[0068] Step S6: Utilize the diagnostic model f θ1 Fault classification and diagnosis are performed on real-time vibration signals in the target domain.
[0069] Furthermore, in step S2, the multi-domain physical feature extraction includes the following sub-steps:
[0070] Step S21: Extract the fault characteristic frequency f of the vibration signal d and its amplitude A d ;
[0071] Step S22: Extract sideband information from the envelope spectrum and calculate sideband modulation features:
[0072]
[0073] Where N s For the number of sidebands, A i Let be the amplitude of the i-th sideband;
[0074] Step S23: Extract the center frequency f of the resonant demodulation band of the vibration signal. c and bandwidth B;
[0075] Step S24: Extract the impact response attenuation coefficient λ and natural frequency f of the vibration signal. n ;
[0076] Step S25: Extract f d A d , N s , R s , f c , B, λ, f n After dimensional unification and normalization, it is combined with the fault type code c to construct a multidimensional physical feature vector p, with each sample corresponding to a multidimensional physical feature vector.
[0077] Furthermore, in step S3, the difference-aware interpolation enhancement method includes the following sub-steps:
[0078] Step S31: Count the number of samples of each category in the source domain vibration signal, calculate the sample entropy of each category, and identify the category with sample entropy lower than the preset threshold as a scarce fault category;
[0079] Step S32: For each rare fault category, randomly select two samples x from that category. i x j and its corresponding multidimensional physical feature vector p i p j And calculate the Mahalanobis distance d between the two samples. M ;
[0080] Step S33: Based on the Mahalanobis distance described in step S32 and the current number of samples in this category, dynamically determine the interpolation weight α:
[0081] α=1 / (1+(d M / d0) m ×(n i / N) n )
[0082] Where d0 is the baseline distance parameter, m is the distance sensitivity index, and n i N is the current number of samples for this scarce fault category, N is the target number of samples, and n is the scarcity sensitivity index.
[0083] Initial samples are generated through linear interpolation.
[0084] x new = α·x i + (1-α)·x j ;
[0085] Step S34: Dynamically determine the disturbance intensity σ based on the differences in physical characteristics, using the following method:
[0086] σ=σ0×(1+γ×d M )×exp(-β×‖p i -p j ||1)
[0087] Where σ0 is the reference disturbance intensity, γ is the range amplification factor, and d M Let p be the Mahalanobis distance between the two samples, β be the feature difference decay coefficient, and ||p| ... i - p j ||1 represents the L1 norm difference of the physical feature vectors;
[0088] Step S35: Based on the perturbation intensity σ determined in step S34, add anisotropic Gaussian noise to the initial sample generated in step S33 to obtain the enhanced sample;
[0089] Step S36: Perform physical validity verification on the enhanced sample described in step S35. If the preset consistency threshold is met, retain the sample; otherwise, return to step S33 until the verification is passed or the preset maximum number of attempts is reached. If the verification still fails after reaching the maximum number of attempts, return to step S32.
[0090] Step S37: Extract the multidimensional physical feature vector of the sample retained in step S36, and determine whether the Mahalanobis distance with the original sample is less than a preset threshold. If it is less than the threshold, retain the sample; otherwise, return to step S33 until the verification is successful or the preset maximum number of attempts is reached. If the maximum number of attempts is reached and the verification is still unsuccessful, return to step S32.
[0091] Step S38: Repeat steps S32 to S37 until the number of scarce fault category samples reaches the target number of samples N, forming an enhanced source domain dataset.
[0092] Furthermore, in step S4, the pre-trained convolutional neural network includes the following sub-steps:
[0093] Step S41: Construct a one-dimensional convolutional neural network, which includes multiple convolutional layers, batch normalization layers, pooling layers, and a classifier;
[0094] Step S42: Supervised pre-training of the network using the augmented source domain dataset, with the optimization objective being the cross-entropy loss function, to obtain the initial diagnostic model f.θ0 ;
[0095] Step S43: After pre-training is completed, calculate the Fisher information matrix of each parameter in the model to quantify the importance of the parameters to the source domain task.
[0096] Furthermore, in step S5, the lightweight fine-tuning strategy for fixing the elastic weights includes the following sub-steps:
[0097] Step S51: Based on the source domain pre-trained model, calculate the Fisher information matrix of each parameter in the model as a measure of parameter importance;
[0098] Step S52: Based on the Fisher information, freeze all parameters of the feature extraction network and set the classifier parameters to a trainable state;
[0099] Step S53: Dynamically determine the regularization coefficient for each category based on the number of samples in each category of the target domain.
[0100] λ c = λ0×(1+η×(N max / n c ))
[0101] Where λ0 is the baseline regularization coefficient, η is the scarcity sensitivity factor, and N max n is the number of samples from the category with the largest number of samples in the target domain. c Let c be the number of samples in category c;
[0102] Step S54: Dynamically determine the adaptive learning rate lr for each parameter based on the Fisher information of each parameter. i :
[0103] lr i = lr0 / (1+κ×F i / F max )
[0104] Where lr0 is the baseline learning rate, κ is the sensitivity factor, and F... i For the Fisher information of the i-th parameter, F max The largest Fisher information among all parameters;
[0105] Step S55: Construct the joint optimization objective function:
[0106]
[0107] Where L few For small sample classification loss in the target domain, θ i and θ i,oldThese are the current value of the i-th parameter during the fine-tuning process and the initial value of the parameter after pre-training in the source domain, respectively.
[0108] Step S56: Using a small number of labeled samples from the target domain, with the goal of minimizing the joint optimization objective function, the classifier parameters are optimized using gradient descent and an adaptive learning rate, while the feature extraction network parameters are kept frozen, resulting in a diagnostic model f adapted to the target domain. θ1 .
[0109] The method of the present invention will be illustrated below using a specific case of rotating machinery failure as an example.
[0110] The effectiveness of the proposed method is verified using the Case Western Reserve University (CWRU) bearing dataset as an example. The experiment selected drive-end bearing data, with a sampling frequency of 12 kHz and a sample length of 1024 data points. Samples at a rotational speed of 1797 r / min were selected as the source domain, and samples at rotational speeds of 1772 r / min, 1750 r / min, and 1730 r / min were selected as the target domain. Each rotational speed included four fault types: normal, inner race fault, outer race fault, and rolling element fault. 200 samples were selected for each fault type in the source domain, and 1, 3, and 5 samples were selected for each fault type in the target domain as the training set, with the remainder used as the test set. To simulate strong background noise interference in real industrial scenarios, -12 dB Gaussian white noise was introduced into the original vibration signal to verify the robustness of the proposed method under low signal-to-noise ratio conditions.
[0111] The specific implementation process is as follows:
[0112] Step S1: Acquire the rotating machinery vibration signals in the source domain and the target domain, wherein the source domain is the reference working condition and the target domain is the new working condition to be diagnosed;
[0113] In this embodiment, the source domain is the baseline operating condition, and vibration data at a speed of 1797 r / min is selected; the target domain is the new operating condition to be diagnosed, and vibration data at speeds of 1772 r / min, 1750 r / min, and 1730 r / min are selected respectively. Each operating condition includes the four types of fault states listed in Table 1.
[0114] Step S2: Extract multi-domain physical features from the source domain vibration signal and construct a multi-dimensional physical feature vector;
[0115] After unifying and normalizing the above features, they are combined with the fault type code c. Preferably, in this example, a physical feature vector p = [f d / 1000, A d , N s / 50, R s , f c / 6000, B / 1000, λ, f n / 5000,c / 10]. Each source domain sample corresponds to a multidimensional physical feature vector.
[0116] Step S3: Based on the physical feature vector, a difference-aware interpolation enhancement method is used to generate enhanced fault samples that meet physical consistency constraints, forming an enhanced source domain dataset, such as... Figure 2 As shown;
[0117] The baseline distance parameter d0 is set to 1.0, the distance sensitivity index m is set to 2.0, the scarcity sensitivity index n is set to 1.0, and the target sample size N is set to 200. σ0 = 0.05, γ = 1.0, and β = 0.1.
[0118] The enhanced samples undergo dual verification: First, physical validity is verified using a comprehensive evaluation of three indicators: ① kurtosis value within the range of 2.5 to 5.0; ② energy proportion near the fault characteristic frequency is greater than 3%; ③ energy proportion of the envelope spectrum modulation sideband is greater than 20%. When the weighted score of the three indicators (kurtosis weight 0.4, fault frequency energy weight 0.4, modulation sideband energy weight 0.2) is greater than a preset threshold (0.5 in this embodiment), the sample that passes the dual verification is retained; otherwise, it is regenerated. This process is repeated until the number of scarce fault category samples reaches a balanced distribution, forming the enhanced source domain dataset.
[0119] Step S4: Pre-train a convolutional neural network as a feature extractor on the augmented source domain dataset to obtain the initial diagnostic model f. θ0 ;
[0120] Preferably, a one-dimensional convolutional neural network is constructed, with the following structure: the input layer receives a 1×1024 vibration signal; convolutional layer 1 contains 32 convolutional kernels of size 32, with a stride of 4 and padding of 14, followed by batch normalization, ReLU activation, and max pooling of size 2; convolutional layer 2 contains 64 convolutional kernels of size 5, with a stride of 2 and padding of 2, followed by batch normalization, ReLU activation, and max pooling of size 2; convolutional layer 3 contains 128 convolutional kernels of size 3, with a stride of 4 and padding of 14, followed by batch normalization, ReLU activation, and max pooling of size 2; The first convolutional layer has a stride of 2, padding of 1, followed by batch normalization and ReLU activation; the second convolutional layer contains 256 kernels of size 3, with a stride of 2, padding of 1, followed by batch normalization and ReLU activation; the third global average pooling layer compresses the feature map into a 256-dimensional feature vector; the fourth convolutional layer consists of Dropout (0.3), a fully connected layer (256-dimensional input, 128-dimensional output), ReLU, Dropout (0.3), and a fully connected layer (128-dimensional input, 4-dimensional output).
[0121] Supervised pre-training of the network was performed using an enhanced source domain dataset. The cross-entropy loss function was employed, with Adam as the optimizer, an initial learning rate of 0.001, a batch size of 64, and 50 training epochs, resulting in the initial diagnostic model f. θ0 After pre-training is completed, the Fisher information matrix of each parameter in the model is calculated to quantify the importance of the parameters to the source domain task.
[0122] Step S5: Transfer the initial diagnostic model f θ0 Deployed to the target domain, and utilizing labeled samples collected from the target domain, the model is fine-tuned using a lightweight fine-tuning strategy with fixed elastic weights to obtain a diagnostic model f adapted to the target domain. θ1 ;
[0123] Preferably, λ0=1000, η=0.5, lr0=0.005, κ=0.1.
[0124] Step S6: Utilize the diagnostic model f θ1 Fault classification and diagnosis are performed on real-time vibration signals in the target domain;
[0125] The adapted diagnostic model is deployed into the actual system to classify the real-time vibration signals collected in the target domain and output fault diagnosis results.
[0126] To verify the effectiveness and advantages of the method of this invention, direct transfer, traditional fine-tuning, and several mainstream domain adaptation methods were used as comparison benchmarks, including Dynamic Subdomain Adaptation Network (DSAN), Structure Regularized Dynamic Clustering (SRDC), and Fixed Ratio Bidirectional Adaptation (FixBi). The comparison results of the cross-load transfer diagnostic accuracy of different methods on the CWRU dataset are shown in Table 1.
[0127] Table 1. Comparison of cross-load migration diagnostic accuracy of different methods on the CWRU dataset (%)
[0128]
[0129] The experimental results show that the method of the present invention achieved an average diagnostic accuracy of 95.62% in 9 cross-load migration experimental scenarios, which is significantly better than direct migration (88.26%), traditional fine-tuning (88.25%) and a number of domain adaptation methods proposed in recent years.
[0130] Compared to the best-performing comparative method DSAN (90.52%), the method of this invention improves by 5.10 percentage points; compared to FixBi (90.42%), it improves by 5.20 percentage points; and compared to SRDC (89.69%), it improves by 5.93 percentage points. Especially in extremely small sample sizes such as 1-shot and 5-shot tests, the method of this invention demonstrates stronger robustness and generalization ability, achieving 97.01% accuracy in the 5-shot task from 1797 to 1730, as high as 99.25% in the 5-shot task from 1797 to 1750, and an even higher accuracy of 99.40% in the 5-shot task from 1797 to 1772.
[0131] It is worth noting that the method of this invention achieved the highest accuracy in 8 out of 9 tasks, which fully verifies the effectiveness and superiority of the proposed method in fault diagnosis of small samples across operating conditions.
[0132] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the invention by those skilled in the art. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the invention should be included within the scope of protection of the invention.
Claims
1. A method for fault diagnosis of cross-domain rotating machinery based on physical feature enhancement, characterized in that, Includes the following steps: Step S1: Acquire the rotating machinery vibration signals in the source domain and the target domain, wherein the source domain is the reference working condition and the target domain is the new working condition to be diagnosed; Step S2: Extract multi-domain physical features from the source domain vibration signal and construct a multi-dimensional physical feature vector; Step S3: Based on the physical feature vector, a difference-aware interpolation enhancement method is used to generate enhanced fault samples that meet physical consistency constraints, forming an enhanced source domain dataset; Step S4: Pre-train a convolutional neural network as a feature extractor on the augmented source domain dataset to obtain the initial diagnostic model f. θ0 ; Step S5: Transfer the initial diagnostic model f θ0 Deployed to the target domain, and utilizing labeled samples collected from the target domain, the model is fine-tuned using a lightweight fine-tuning strategy with fixed elastic weights to obtain a diagnostic model f adapted to the target domain. θ1 ; Step S6: Utilize the diagnostic model f θ1 Fault classification and diagnosis are performed on real-time vibration signals in the target domain.
2. The method for fault diagnosis of cross-domain rotating machinery based on physical feature enhancement as described in claim 1, characterized in that, In step S2, the multi-domain physical feature extraction includes the following sub-steps: Step S21: Extract the fault characteristic frequency f of the vibration signal d and its amplitude A d ; Step S22: Extract sideband information from the envelope spectrum and calculate sideband modulation features: Where N s For the number of sidebands, A i Let be the amplitude of the i-th sideband; Step S23: Extract the center frequency f of the resonant demodulation band of the vibration signal. c and bandwidth B; Step S24: Extract the impact response attenuation coefficient λ and natural frequency f of the vibration signal. n ; Step S25: Extract f d A d , N s , R s , f c , B, λ, f n After dimensional unification and normalization, it is combined with the fault type code c to construct a multidimensional physical feature vector p, with each sample corresponding to a multidimensional physical feature vector.
3. The method for fault diagnosis of cross-domain rotating machinery based on physical feature enhancement according to claim 1, characterized in that, In step S3, the difference-aware interpolation enhancement method includes the following sub-steps: Step S31: Count the number of samples of each category in the source domain vibration signal, calculate the sample entropy of each category, and identify the category with sample entropy lower than the preset threshold as a scarce fault category; Step S32: For each rare fault category, randomly select two samples x from that category. i x j and its corresponding multidimensional physical feature vector p i p j And calculate the Mahalanobis distance d between the two samples. M ; Step S33: Based on the Mahalanobis distance described in step S32 and the current number of samples in this category, dynamically determine the interpolation weight α: α=1 / (1+(d M / d0) m ×(n i / N) n ) Where d0 is the baseline distance parameter, m is the distance sensitivity index, and n i N is the current number of samples for this scarce fault category, N is the target number of samples, and n is the scarcity sensitivity index. Initial samples are generated through linear interpolation. x new = a·x i + (1-a)·x j ; Step S34: Dynamically determine the disturbance intensity σ based on the differences in physical characteristics, using the following method: σ=σ0×(1+γ×d M )×exp(-β×‖p i -p j ‖1) Where σ0 is the reference disturbance intensity, γ is the range amplification factor, and d M Let p be the Mahalanobis distance between the two samples, β be the feature difference decay coefficient, and ||p| ... i - p j ||1 represents the L1 norm difference of the physical feature vectors; Step S35: Based on the perturbation intensity σ determined in step S34, add anisotropic Gaussian noise to the initial sample generated in step S33 to obtain the enhanced sample; Step S36: Perform physical validity verification on the enhanced sample described in step S35. If the preset consistency threshold is met, retain the sample; otherwise, return to step S33 until the verification is passed or the preset maximum number of attempts is reached. If the verification still fails after reaching the maximum number of attempts, return to step S32. Step S37: Extract the multidimensional physical feature vector of the sample retained in step S36, and determine whether the Mahalanobis distance with the original sample is less than a preset threshold. If it is less than the threshold, retain the sample; otherwise, return to step S33 until the verification is successful or the preset maximum number of attempts is reached. If the maximum number of attempts is reached and the verification is still unsuccessful, return to step S32. Step S38: Repeat steps S32 to S37 until the number of scarce fault category samples reaches the target number of samples N, forming an enhanced source domain dataset.
4. The method for fault diagnosis of cross-domain rotating machinery based on physical feature enhancement according to claim 1, characterized in that, In step S4, the pre-trained convolutional neural network includes the following sub-steps: Step S41: Construct a one-dimensional convolutional neural network, which includes multiple convolutional layers, batch normalization layers, pooling layers, and a classifier; Step S42: Supervised pre-training of the network using the augmented source domain dataset, with the optimization objective being the cross-entropy loss function, to obtain the initial diagnostic model f. θ0 ; Step S43: After pre-training is completed, calculate the Fisher information matrix of each parameter in the model to quantify the importance of the parameters to the source domain task.
5. The method for fault diagnosis of cross-domain rotating machinery based on physical feature enhancement according to claim 1, characterized in that, In step S5, the lightweight fine-tuning strategy for fixing the elastic weights includes the following sub-steps: Step S51: Based on the source domain pre-trained model, calculate the Fisher information matrix of each parameter in the model as a measure of parameter importance; Step S52: Based on the Fisher information, freeze all parameters of the feature extraction network and set the classifier parameters to a trainable state; Step S53: Dynamically determine the regularization coefficient for each category based on the number of samples in each category of the target domain. l c = λ0×(1+η×(N max / n c )) Where λ0 is the baseline regularization coefficient, η is the scarcity sensitivity factor, and N max n is the number of samples from the category with the largest number of samples in the target domain. c Let c be the number of samples in category c; Step S54: Dynamically determine the adaptive learning rate lr for each parameter based on the Fisher information of each parameter. i : lr i = lr0 / (1+κ×F i / F max ) Where lr0 is the baseline learning rate, κ is the sensitivity factor, and F... i For the Fisher information of the i-th parameter, F max The largest Fisher information among all parameters; Step S55: Construct the joint optimization objective function: Where L few For small sample classification loss in the target domain, θ i and θ i,old These are the current value of the i-th parameter during the fine-tuning process and the initial value of the parameter after pre-training in the source domain, respectively. Step S56: Using a small number of labeled samples from the target domain, with the goal of minimizing the joint optimization objective function, the classifier parameters are optimized using gradient descent and an adaptive learning rate, while the feature extraction network parameters are kept frozen, resulting in a diagnostic model f adapted to the target domain. θ1 .