Rotary machinery fault diagnosis method based on partial domain self-adaption
By introducing some fields of adaptive technology in rotary machinery fault diagnosis, using pseudo-label generation and screening strategies, the problem of degradation of diagnostic effects of deep learning methods in the case of data distribution differences and lack of labels is solved, and higher fault classification accuracy and data labeling efficiency are achieved.
Patent Information
- Application Number
- CN202510206284.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-02-25
- Publication Date
- 2025-06-13
AI Technical Summary
The existing rotary mechanical fault diagnosis method based on deep learning has significantly reduced the diagnostic effect in the case of data distribution differences and the lack of effective fault type labels, and it is expensive to obtain high-quality labeled data.
A rotary mechanical fault diagnosis method based on partial domain adaptation is proposed. Through the combination of domain fault feature extractor, data marker, category correlation discriminator, domain attribute discriminator and fault classifier, pseudo-label generation and screening of source domain data is realized, and the classification error of labelless data in the target domain is reduced.
It effectively reduces the irrelevant knowledge error disturbance in the source domain when the target domain fault knowledge is transferred, improves the model's classification accuracy of the target domain fault, and improves the efficiency and quality of data labeling.
Smart Images
Figure CN120145248A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of mechanical fault diagnosis, and particularly relates to a rotating machinery fault diagnosis method based on partial domain adaptation. Background Art
[0002] Rotating machinery plays a key role in the transmission equipment group of industrial applications. However, with the trend of mechanical structures towards complexity, the key components inside rotating machinery face higher fault risks. At the same time, the occurrence of faults will lead to a decline in production quality and pose safety hazards, etc. Therefore, researching effective fault diagnosis and health monitoring methods is of great significance for ensuring and maintaining the safety of rotating machinery.
[0003] With the rapid progress of artificial intelligence algorithms, data-driven fault diagnosis methods have flourished in recent years. Among these methods, deep learning-based fault diagnosis methods have received much attention due to their excellent data processing capabilities. However, deep learning-based fault diagnosis methods require a large number of accurately labeled fault samples as a training basis and assume that the data distributions between the training and test data sets are similar. However, since rotating machinery often operates under different operating conditions, labeling the fault data under each different condition is time-consuming and labor-intensive, which makes the diagnostic effect of deep learning-based fault diagnosis methods significantly decline in scenarios where there are distribution differences between data and a lack of effective fault type label source domain data to guide the fault classification model.
[0004] In actual operation, obtaining such a high-quality labeled data set is a complex and expensive task. For each collected fault sample, detailed fault type labeling needs to be completed by professional engineers based on their rich experience and professional knowledge. This process is not only time-consuming but also costly. This means that even in the case of abundant data, the effective labeled data available for training deep learning models is still limited. This limitation in data directly restricts the learning ability of the models, making them unable to fully exploit the potential fault knowledge in all data, thereby affecting the accuracy and reliability of fault diagnosis.
[0005] Domain adaptation techniques can transfer knowledge from a source domain rich in fault type labels to a target domain lacking fault type labels, thereby weakening the impact of the data distribution difference between domains. Therefore, scholars have explored the application of domain adaptation techniques in cross-domain fault diagnosis under different operating conditions. Although significant progress has been made in fault diagnosis using domain adaptation techniques, existing methods mainly focus on closed-set domain adaptation fault diagnosis, where it is assumed that the label spaces of the source domain and the target domain are the same. However, this closed-set assumption limits the flexibility and applicability of cross-domain fault diagnosis methods. To overcome this limitation, scholars have extended the closed-set assumption by introducing the partial domain adaptation (PDA) fault diagnosis method. Summary of the Invention
[0006] To solve the above technical problems, the present invention proposes a rotating machinery fault diagnosis method based on partial domain adaptation. The structural composition of the rotating machinery fault diagnosis model proposed in the present invention includes a domain fault feature extractor, a data marker, a category correlation discriminator, a domain attribution discriminator, and a fault classifier, which can achieve the fault diagnosis task in the scenario of the partial set label space relationship where only part of the source domain data in the dataset has labels and the other part lacks effective labels. This method can effectively label and screen the unlabeled source domain data in the dataset and accurately classify the known fault types on cross-domain objects (target domain).
[0007] In a first aspect, the present invention proposes a rotating machinery fault diagnosis method based on partial domain adaptation, the method comprising: collecting vibration data of key components of a rotating machinery, and inputting the vibration data after data preprocessing into a trained rotating machinery fault detection model to obtain a fault diagnosis result of the rotating machinery;
[0008] S1. Using an acceleration sensor to collect vibration data of key components of a rotating machinery under a plurality of different operating conditions, and performing data preprocessing on the vibration data;
[0009] S2. According to the vibration data after data preprocessing, obtaining a training sample set, which includes source domain data and target domain data, the source domain data includes labeled vibration data and part of unlabeled vibration data, and all the target domain data is unlabeled vibration data;
[0010] S3. Using a domain fault feature extractor to extract features from the source domain data and the target domain data, respectively obtaining source domain data features and target domain data features;
[0011] S4. Using a data marker to generate pseudo-labels for the unlabeled source domain data features, obtaining source domain data features with pseudo-labels, updating the source domain data features, and calculating the marking loss;
[0012] S5. Input the labeled source domain data features, the pseudo-labeled source domain data features, and the target domain features into the class correlation discriminator together to obtain the fault class prediction probabilities respectively, and calculate the class similarity loss L D and the class similarity weight;
[0013] S6. According to the fault class prediction probability results of the source domain and the target domain, combined with the class similarity weight perform common class weighting on the data features of the source domain and the target domain and then input them into the domain attribution discriminator to obtain the corresponding output probabilities of the domain attribution discriminator, and calculate the domain discrimination loss
[0014] S7. Use the fault classifier to perform classification prediction on the updated source domain data features and calculate the fault classification loss
[0015] S8. According to the labeled loss, the class similarity loss L D , the domain discrimination loss and the fault classification loss perform loss backpropagation to optimize the network parameters.
[0016] In a second aspect, the present invention provides an electronic device, including: a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the program, it implements the method for rotating machinery fault diagnosis based on partial domain adaptation described in the first aspect of the present invention.
[0017] In a third aspect, the present invention provides a computer-readable storage medium, on which a computer program is stored. When the computer program is controlled by a processor, it implements the method for rotating machinery fault diagnosis based on partial domain adaptation described in the first aspect of the present invention.
[0018] Advantages of the present invention:
[0019] (1) This method constructs a classification consistency data labeling and screening strategy based on a prototype network and a classifier network through the labeled source domain data features in the source domain dataset, realizes the labeling process of the unlabeled data in the source domain, and eliminates the source domain data features with unclear classification or low quality to ensure the quality of the source domain data feature set.
[0020] (2) This method adopts an adversarial learning strategy to construct a class correlation discriminator, obtains the similarity weight between the source domain class and the target domain class, and constructs a class similarity loss. Similar classes are common classes, and dissimilar classes are source domain specific classes. The similarity weight is used for the source domain data to highlight the common fault classes in the two domains and reduce the influence brought by the source domain specific classes at the same time.
[0021] (3) This method further constructs a domain attribution discriminator by using the weighted high-dimensional features of the source domain fault samples and the target domain sample features through an adversarial learning strategy, and constructs a domain discrimination loss in combination with domain labels. By optimizing this loss function, the feature extractor is updated and optimized to better capture the common invariant features between the source domain and the target domain, thereby promoting domain adaptation. Through this method, when the source domain transfers fault knowledge to the target domain, the interference of irrelevant knowledge of the unique categories in the source domain can be effectively reduced, and the classification accuracy of the model for target domain faults can be improved. Description of the Drawings
[0022] Figure 1 It is a schematic diagram for comparing the fault diagnosis of rotating machinery between traditional domain adaptation and partial domain adaptation;
[0023] Figure 2 It is a flowchart of the training steps of the rotating machinery fault detection model in the embodiment of the present invention;
[0024] Figure 3 It is a schematic diagram of the process in the embodiment of the present invention;
[0025] Figure 4 It is a schematic diagram of the confusion matrix between the predicted labels and the true labels of the diagnosis model for unlabeled source domain data in the experimental verification stage of the present invention. Detailed Embodiments
[0026] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.
[0027] The rotating machinery fault diagnosis method based on deep learning requires a large number of accurately labeled fault samples as the basis for model training. The fault samples with annotations are the source domain, and the source domain may contain some specific fault types that do not appear in the target domain. The unique faults existing in the source domain may cause the diagnosis model to learn unnecessary features, which are not applicable in the target domain and may even cause the model to overfit to the data characteristics of the source domain, thereby reducing its generalization ability in the target domain. Aiming at the limitation that some data in the mechanical fault dataset have labels while some data do not have effective labels, and the problem of label space difference in the traditional domain adaptation-based rotating machinery fault diagnosis model where the target domain does not contain all the fault categories in the source domain, the accuracy of rotating machinery fault detection is low.
[0028] Referring to Figure 1 as shown, the present invention adopts a partial domain adaptation (PDA) fault diagnosis method and improves the existing partial domain adaptation.
[0029] In an embodiment of the present invention, a rotating machinery fault diagnosis method based on partial domain adaptation is proposed. The method includes: collecting vibration data of key components of the rotating machinery, and after data preprocessing, inputting it into a trained rotating machinery fault detection model to obtain a fault diagnosis result of the rotating machinery.
[0030] Referring to Figure 2 as shown, the training process of the rotating machinery fault detection model includes:
[0031] S1. Using an acceleration sensor to collect vibration data of key components of the rotating machinery under multiple different operating conditions, and performing data preprocessing on the vibration data.
[0032] Specifically, collect vibration signals (i.e., vibration data) of key components of the rotating machinery such as bearings or gears under multiple operating conditions (changes in speed or load, etc.). This vibration signal is usually a time-domain signal. Use the Fast Fourier Transform (FFT) to convert the time-domain signal into a frequency-domain signal to obtain its Fourier phase spectrum. According to expert knowledge and experience, manually label the collected vibration data to obtain the fault type and its label. Usually, the fault type is also called the fault category.
[0033] Exemplarily, the gearbox fault data set is that the gearbox fault vibration signals are collected from an experimental test bench at two different speeds (i.e., 30 Hz and 40 Hz). The test bench consists of a motor, a planetary gearbox, a parallel shaft gearbox, and a magnetic drive. This experiment simulates seven gearbox faults, namely 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 has 200 original fault samples with a length of 1024 points. The construction of the gearbox fault data set is for a partial domain adaptation fault diagnosis task with only partial labeled data. After this data set is labeled, the fault type and its label of the gearbox fault data set are obtained, as shown in Table 1.
[0034] Table 1 Fault types and their labels of the gearbox fault data set
[0035]
[0036] S2. Obtain a training sample set according to the vibration data after data preprocessing, which includes source domain data and target domain data. The source domain data includes labeled vibration data and some unlabeled vibration data, and the target domain data are all unlabeled vibration data.
[0037] Specifically, use the data under a certain working condition where the fault type label is accessible during the model training process as the source domain. However, only some of the data in the source domain have valid labels, and the other source domain data do not have fault category labels. At the same time, use the data under another working condition where the fault type label is inaccessible during the model training process as the target domain. Some of the data in the source domain have been marked with fault types using expert knowledge and experience, but there are still some data that have not obtained valid labels. The training set of the rotating machinery fault detection model is all the source domain data and half of the target domain data, while the test set is the remaining target domain data.
[0038] Exemplarily, after empirical experiments and annotation, the source domain data and target domain data of the gearbox fault data set are shown in Table 2.
[0039] Table 2 Source domain data and target domain data of the gearbox fault data set
[0040]
[0041]
[0042] In the embodiment of the present invention, the rotating machinery fault detection model is based on a deep neural network, and includes a domain fault feature extractor, a data marker, a domain attribution discriminator, a category correlation discriminator, and a fault classifier. The data marker includes a Prototypical Networks and a classifier network, and the network structure of the classifier network is the same as that of the fault classifier.
[0043] Exemplarily, the network parameters of the rotating machinery fault detection model are shown in Table 3.
[0044] Table 3 Network parameters of the rotating machinery fault detection model
[0045]
[0046]
[0047] S3. Use the domain fault feature extractor to extract features from the source domain data and the target domain data, and obtain source domain data features and target domain data features respectively.
[0048] The domain fault feature extractor is denoted as: f(·), and its input data is denoted as x, where x is sourced from labeled source domain data, unlabeled source domain data, and target domain data. The source domain data features are high-dimensional feature representations, which include labeled source domain data features and unlabeled source domain data features. The target domain data features are also high-dimensional feature representations.
[0049] S4. Use the data marker to generate pseudo-labels for the unlabeled source domain data features, obtain the source domain data features with pseudo-labels, update the source domain data features, and calculate the marking loss.
[0050] In the embodiment of the present invention, a data marker is constructed using the labeled source domain data features, prototype network classification prediction and classifier prediction are performed on the unlabeled source domain data features, the consistency of the two prediction results is judged, and pseudo-labels are generated for the unlabeled source domain data features according to the consistency, obtaining the source domain data features with credible pseudo-labels, and updating and screening the source domain data features.
[0051] In the embodiment of the present invention, step S4 specifically includes:
[0052] S401: Use the prototype network to process the labeled source domain data features to obtain the prototype p of each category in the labeled source domain data features c , which is specifically expressed as:
[0053]
[0054] where p c represents the prototype of each fault category in the labeled source domain data features, c represents the index of the source domain fault category, c ∈ {1, 2,..., C}, C represents the total number of fault categories, N c represents the number of samples in category c, x i represents the i-th labeled sample belonging to category c, and f(·) represents the fault feature extractor.
[0055] S402: Calculate the similarity scores between the unlabeled source domain data features and each fault category in the labeled source domain data features to obtain the classification prediction result P prob (c|j) of the prototype network.
[0056] The calculation formula for the similarity score is:
[0057]
[0058] In the formula, D j,c represents the similarity score between the j-th unlabeled source domain data feature x j and each category c in the labeled source domain data features, Tp c represents the operation on p cFind the transpose, f(x j ) k represents the feature x of the source domain data without labels j in the k-th dimension, p c,k represents the component of the prototype of each category in the feature of the source domain data with labels in the k-th dimension, k ∈ {1, 2, …, D}, and D represents the number of dimensions.
[0059] The classification output result P prob (c|j) of the prototype network is calculated as follows:
[0060]
[0061] In the formula, P prob (c|j) represents the classification probability of the prototype network that the j-th feature x of the source domain data without labels j belongs to the category c, and D j,c′ represents the similarity score between the j-th feature x of the source domain data without labels j and each category c' of the feature of the source domain data with labels.
[0062] S403: Use the classifier network to classify and predict the feature of the source domain data without labels, and obtain the classification prediction result P CLAprob (c|j) of the classifier network, and its specific calculation formula is:
[0063]
[0064] Among them, P CLAprob (c|j) represents the probability of the classifier network that the j-th feature x of the source domain data without labels j belongs to the fault category c, and o j,C represents the logarithmic probability that the classifier network that the j-th feature x of the source domain data without labels j belongs to the specific fault category c, and o j,c ' represents the logarithmic probability that the classifier network that the j-th feature x of the source domain data without labels j belongs to the fault category c'.
[0065] S404: According to the classification prediction result P prob (c|j) of the prototype network and the classification prediction result P CLAprob (c|j) of the classifier network, judge the consistency of the prediction results, and generate pseudo-labels for the feature of the source domain data without labels according to the consistency, and obtain the feature of the source domain data with pseudo-labels.
[0066] S405: Screen and delete the poorly performing data features, and update the feature of the source domain data with pseudo-labels to the set of features of the source domain data with labels.
[0067] During the next round of model training, the source domain data with pseudo-labels is used as labeled data. During the iterative training process, the prototype p of each category in the labeled source domain data features is updated using the momentum centroid update method. c , and its update formula is specifically:
[0068] p c ←(1 - α)p c + αc c ,
[0069]
[0070] In the formula, p c represents the prototype of each fault category in the labeled source domain data features, α represents the momentum coefficient, and c c represents the feature center of fault category c in the current training batch, Y j,c represents the indicator variable used to determine whether the current sample belongs to category c, M represents the number of samples in the batch, and f(x j ) represents the feature representation obtained by mapping the input unlabeled sample x j through the fault feature extractor.
[0071] By generating pseudo-labels for the unlabeled source domain data, the fault category feature information of the source domain dataset is expanded, and in subsequent training, the prototype of each category in the labeled source domain data features is updated using the momentum centroid update method according to the expanded source domain dataset, enabling the updated data tagger to learn the fault feature knowledge contained in the unlabeled source domain data and more accurately predict and classify the remaining unlabeled source domain data.
[0072] Calculate the labeling loss during the process of generating pseudo-labels, and during training, optimize the data tagger according to this loss through backpropagation.
[0073] S5. Input the labeled source domain data features, the source domain data features with pseudo-labels, and the target domain features into the category correlation discriminator respectively to obtain the fault category prediction probabilities, and calculate the category similarity loss L D and the category similarity weight.
[0074] The specific process of calculating the category similarity loss L D includes:
[0075] Using the category correlation discriminator, obtain the fault category prediction probabilities of the source domain data features and the target domain features respectively, and calculate the category similarity coefficient of the input features, which is the output result D(f(x)) of the category correlation discriminator. Its calculation formula is:
[0076] D(f(x)) = p(y = 1|f(x)),
[0077]
[0078] where f(x) represents the current input feature, D(·) represents the class correlation discriminator model, p(y = 1|f(x)) represents the probability that the current feature f(x) comes from the source domain, y = 1 indicates that when the indicator variable is 1, the sample is determined to come from the source domain, p s (f(x)) represents the probability that the current feature f(x) comes from the source domain, p t (f(x)) represents the probability that the current feature f(x) comes from the target domain. When the similarity coefficient is closer to 1, it indicates that the sample is more likely to come from the source domain; conversely, the closer the value is to 0, the more likely it is that the sample comes from the target domain.
[0079] Calculate the class similarity loss L of the source domain data features and the target domain features according to the fault class prediction probabilities of the source domain data features and the target domain features D , and its calculation formula is:
[0080]
[0081] where denotes the expectation, x s represents the source domain fault sample, x t represents the target domain fault sample.
[0082] Calculate the class similarity weight of the input feature according to the fault class prediction probability of the class similarity coefficient of the input feature, and its calculation formula is:
[0083]
[0084]
[0085] where represents the class similarity weight between the source domain and the target domain, and can also be abbreviated as ω.
[0086] Through the class similarity weights of the source domain and the target domain it is possible to highlight or make the model pay more attention to the common fault classes of the source domain and the target domain, reduce the influence of irrelevant knowledge in the source domain-specific fault classes, and improve the prediction accuracy of the model.
[0087] S6. According to the fault class prediction probability results of the source domain and the target domain, combined with the class similarity weight perform common class weighting on the data features of the source domain and the target domain and then input them into the domain attribution discriminator to obtain the corresponding domain attribution discriminator output probability, and calculate the domain discrimination loss
[0088] The similar categories are the common categories, and the dissimilar categories are the source-domain specific categories. The similarity weights are used for the source-domain data to highlight the common fault categories in both domains while reducing the impact brought by the source-domain specific categories.
[0089] The domain discrimination loss has the following calculation formula:
[0090]
[0091] In the formula, denotes taking the expectation, w(f(x s )) represents the importance weight of the source-domain samples, which reduces the weights of the source-domain specific category samples while increasing the weights of the common category samples. x s represents the source-domain fault samples, D 0 (·) represents the model of the domain attribution discriminator, D 0 (f(x s )) represents the probability value that the domain attribution discriminator determines the input features come from the source domain, x t represents the target-domain fault samples, D 0 (f(x t )) represents the probability value that the domain attribution discriminator determines the input features come from the target domain. log D 0 (f(x s )) represents taking the natural logarithm of the probability value that the domain attribution discriminator determines to belong to the source domain for the source-domain sample x s , which is used as the confidence score of the discriminator for the feature representation of the source-domain samples.
[0092] When constructing the domain discrimination loss , by multiplying the importance weight w(f(x s )) of the source-domain samples by the confidence score that the sample belongs to the source domain, different degrees of emphasis on different source-domain samples are achieved. If the feature representation f(x s ) of a certain source-domain sample is more similar to the target domain, that is, the importance weight w(f(x s )) is higher, then the contribution of this sample to the domain discrimination loss will be greater, which prompts the model to learn more of the sample features common to these source domains and the target domain. Conversely, because the importance weights of the source-domain specific category samples are small, the influence of the source-domain specific category samples on the domain discrimination loss will also be correspondingly reduced, thereby effectively reducing the interference of the irrelevant knowledge of the source-domain specific categories when the source domain transfers fault knowledge to the target domain and improving the classification accuracy of the model for the target-domain faults.
[0093] S7. Use the fault classifier to classify and predict the updated source domain data features, and calculate the fault classification loss.
[0094] The fault classification loss is the cross-entropy loss, and its calculation formula is:
[0095]
[0096] Among them, represents the cross-entropy loss, x s represents the input source domain data features, y represents the true label of the source domain data features, and C(·) represents the model of the fault classifier.
[0097] The updated source domain data features are obtained after being processed in step S4, and they include labeled source domain data features, unlabeled source domain data features, and source domain data features with pseudo-labels.
[0098] S8. According to the marker loss, the class similarity loss L D , the domain discrimination loss and the fault classification loss perform loss backpropagation to optimize the network parameters.
[0099] Among them, the class similarity loss L D and the domain discrimination loss adopt the adversarial training method when training the corresponding discriminator.
[0100] After the training of the rotating machinery fault detection model is completed, the network parameters of the domain fault feature extractor and the fault classifier have been optimized. When applying the model to the actual detection or prediction stage, only the fault feature extractor and the fault classifier are used, and its prediction process includes:
[0101] Use an acceleration sensor to collect vibration data of key components of the rotating machinery under multiple different operating conditions, and perform data preprocessing on the vibration data;
[0102] Extract features through the optimized fault feature extractor;
[0103] Input the extracted features into the optimized fault classifier to obtain the fault diagnosis result of the rotating machinery.
[0104] An embodiment of the present invention proposes an electronic device, including: a memory, a processor, and a computer program stored on 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 the present invention.
[0105] An embodiment of the present invention provides a computer-readable storage medium, on which a computer program is stored. When the computer program is controlled by a processor, it implements the method for diagnosing faults of rotating machinery based on partial domain adaptation according to the first aspect of the present invention.
[0106] Experimental verification:
[0107] To verify the effectiveness of the present invention, an experimental verification was carried out using the gearbox fault vibration data set (DDS) collected from a Drivetrain Diagnostics Simulator (DDS).
[0108] In the gearbox fault data set (DDS), the gearbox fault vibration signals were collected from an experimental test bench at two different rotational speeds, namely 30 Hz and 40 Hz. The test bench consists of an electric motor, a planetary gearbox, a parallel shaft gearbox, and a magnetic drive. Seven types of gearbox faults were simulated, namely 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 has 200 original fault samples with a length of 1024 points.
[0109] Based on the above available gearbox fault data, the deep learning-based rotating machinery fault diagnosis model (Proposed) proposed by the present invention was compared with the DANN model proposed in the literature "Unsupervised domain adaptation by backpropagation" in the International conference on machine learning, Jun. 2015, and the IWAN model proposed in the literature "Importance weighted adversarial nets for partial domain adaptation" in Proc. IEEE / CVF Conf. Comput. Vis. Pattern Recognit. (CVPR)., 2018. To ensure the fairness of the experiment, the above methods all adopt the same feature extractor and classifier structure.
[0110] To obtain stable experimental results and minimize the influence of randomness on the experimental results, each comparison method was conducted five times, and each experiment was trained 500 times. The fault classification accuracy was compared with the proposed method under two data conditions: training only with labeled data in the source domain and when all data in the source domain have valid true labels. The specific fault diagnosis accuracy results are shown in Table 4. The accuracy of the pseudo-labeled data by the data marker is asFigure 4 As shown in the confusion matrix, Figure 4 in which, the horizontal axis true Lalel represents the true label, and Predicted Lalel represents the label predicted by the model.
[0111] Table 4 Cross-domain fault diagnosis results of the gearbox fault dataset
[0112]
[0113]
[0114] It can be seen from the experimental results that when using only the labeled source domain data for training with the target domain, the accuracy of the traditional adversarial domain adaptation method is only 84.65%, while the accuracy is improved when using the partial domain adaptation method. When all the data in the source domain have true labels and the training data volume is sufficient, the fault knowledge can be better applied to the classification and recognition of target domain faults. At this time, the accuracy of the traditional adversarial domain adaptation method is increased to 87.75%, and the accuracy of the partial domain adaptation method is increased to 90.03%. This shows that constructing an accurate model requires a large amount of high-quality labeled data, and the traditional fault diagnosis model cannot utilize the fault knowledge in the unlabeled data samples in the source domain for the recognition and classification of target domain faults. The proposed method can screen and label the unlabeled samples in the source domain, realize the expansion of the number of labeled samples in the source domain for training, and thus improve the accuracy of the fault diagnosis model.
[0115] Those of ordinary skill in the art can understand that all or part of the steps in the various methods of the above embodiments can be completed by instructing relevant hardware through a program, and this program can be stored in a computer-readable storage medium. The storage medium can include: ROM, RAM, magnetic disk or optical disk, etc.
[0116] Although the embodiments of the present invention have been shown and described, for those of ordinary skill in the art, it can be understood that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principle and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.
Claims
1. A rotating machinery fault diagnosis method based on partial domain adaptation, characterized in that: include: Collect vibration data of key parts of rotating machinery, input the data into the trained rotating machinery fault detection model after data preprocessing, and obtain the fault diagnosis results of rotating machinery; The training process of the rotating machinery fault detection model includes: Use acceleration sensors to collect vibration data of key components of rotating machinery under multiple different operating conditions and perform data preprocessing; The source domain data and the target domain data are obtained according to the vibration data after data preprocessing, wherein the source domain data includes vibration data with labels and some vibration data without labels, and the target domain data are all vibration data without labels; The domain fault feature extractor is used to extract features from the source domain data and the target domain data to obtain source domain data features and target domain data features respectively; Use the data tagger to generate pseudo labels for the unlabeled source domain data features, obtain the source domain data features with pseudo labels, update the source domain data features, and calculate the labeling loss; The labeled source domain data features, pseudo-labeled source domain data features, and target domain features are input into the category relevance discriminator to obtain the fault category prediction probability, and calculate the category similarity loss and category similarity weight of the source domain and the target domain; According to the fault category prediction probability results of the source domain and the target domain, combined with the category similarity weight, the data features of the source domain and the target domain are weighted by the common category and then input into the domain attribution discriminator to obtain the corresponding domain attribution discriminator output probability and calculate the domain discrimination loss; Using the fault classifier, classify and predict the updated source domain data features and calculate the fault classification loss; Loss backpropagation is performed based on labeling loss, category similarity loss, domain discrimination loss, and fault classification loss to optimize network parameters.
2. The rotating machinery fault diagnosis method based on partial domain adaptation according to claim 1 is characterized in that: Using the data marker, we predict the labeled source domain data features and the unlabeled source domain data features respectively, judge the consistency of the prediction results, generate pseudo labels for the unlabeled source domain data features based on the consistency, obtain the source domain data features with pseudo labels, and update and filter the source domain data features.
3. The rotating machinery fault diagnosis method based on partial domain adaptation according to claim 1 is characterized in that: The data tagger includes a prototype network and a classifier network. Through the prototype network, the labeled source domain data features are processed to obtain the prototype p of each category in the labeled source domain data features. c ; Calculate the similarity score of each fault category in the unlabeled source domain data features and the labeled source domain data features, and obtain the classification prediction result P of the prototype network prob ; Use the classifier network to classify and predict the unlabeled source domain data features, and obtain the classification prediction result P of the classifier network CLAprob ; According to the classification prediction result P of the prototype network prob And the classification prediction result P of the classifier network CLAprob , judge the consistency of the prediction results, generate pseudo labels for the unlabeled source domain data features according to the consistency, and obtain the source domain data features with pseudo labels.
4. The rotating machinery fault diagnosis method based on partial domain adaptation according to claim 1 is characterized in that: The source domain data features with pseudo labels are used as labeled data and updated to the labeled source domain data feature set. They are then applied to the iterative training process of the model, and the momentum centroid update method is used to update the prototype p of each category in the labeled source domain data features. c .
5. An electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the rotating machinery fault diagnosis method based on partial domain adaptation as described in any one of claims 1 to 4 is implemented.
6. A computer storable medium, characterized in that: A computer program is stored thereon, and when the computer program is controlled by a processor, the computer program implements any one of the rotating machinery fault diagnosis methods based on partial domain adaptation described in 1-4.