Generalized zero sample fault diagnosis method based on prototype fusion reconstruction and domain decoupling enabling

Abstract

The invention relates to a generalized zero sample fault diagnosis method based on prototype fusion reconstruction and domain decoupling enabling, and belongs to the field of rotating machinery fault diagnosis. According to the method, a domain decoupling strategy is constructed to solve the problem of misjudgment of an unknown fault, the unknown fault and a known fault are separated from the source, and confusion between two types of samples is relieved. Based on an M-OpenMax mechanism proposed by open set learning, the reliability of traditional parameter estimation is improved by introducing the Cramer-Rao lower bound, and the sample distinguishability is enhanced by means of more effective parameter estimation. The invention further provides a prototype fusion reconstruction method to relieve the problem of unknown prototype offset. According to the prototype fusion reconstruction method, information is fused from three dimensions of priori knowledge, known domain features and unknown domain features, and a more comprehensive prototype is constructed. Through triple information integration, under the condition that unknown class samples are not needed in the training stage, the constructed unknown prototype is closer to the essential characteristics of the unknown prototype, and therefore the field deviation of the unknown prototype is relieved.

Description

Technical Field

[0001] This invention belongs to the field of rotating machinery fault diagnosis and relates to a generalized zero-sample fault diagnosis method that combines prototype fusion reconstruction and domain decoupling empowerment. Background Technology

[0002] Fault diagnosis plays a crucial role in modern industrial systems, especially in the field of rotating machinery. With the advancement of intelligent upgrading in modern industry, equipment operating conditions are becoming increasingly complex, and the fault modes of rotating machinery are shifting from single faults to complex, coupled faults. Although fault diagnosis technology has made significant progress in recent years, existing research has largely focused on identifying single fault types. Due to methodological limitations, single fault diagnosis struggles to effectively identify overlapping signal features in complex faults, leading to decreased diagnostic accuracy. Compared to single faults, complex faults are characterized by more complex features, stronger coupling, and greater diagnostic difficulty. Therefore, conducting research on complex fault diagnosis in rotating machinery can effectively improve the existing fault diagnosis theoretical system and also has significant engineering value.

[0003] Deep learning methods, with their powerful feature extraction and discrimination capabilities, offer a feasible approach for identifying complex faults. Recent studies have confirmed their effectiveness, such as the unbalanced learning fusion strategy for bearing-gear complex faults proposed by Zhang et al., and the wavelet-enhanced multi-label network used by Liang et al. for gearbox diagnosis. However, the actual diversity of faults makes complete data acquisition difficult to achieve, and the reliance of traditional deep models on large amounts of labeled data limits their industrial applications. To address this issue, zero-shot learning methods have emerged. This method can identify unknown categories using only category attributes. First proposed by Lampert and applied to image recognition, it is now gradually being applied to fault diagnosis. However, most current zero-shot diagnostic methods only focus on diagnosing faults in unknown domains, meaning the test categories only include fault types in unknown domains. Figure 1 This illustrates the zero-shot learning process under different paradigms, such as... Figure 1 As shown in (a), this type of zero-shot learning is called narrow zero-shot learning. In practical engineering, it is even more necessary to diagnose both known and unknown types of faults simultaneously, such as... Figure 1 As shown in (b). This method of simultaneously diagnosing known and unknown faults is usually called generalized zero-shot learning. Currently, research on zero-shot diagnostic methods is still in its early stages, and fault diagnosis methods based on generalized zero-shot learning are particularly scarce. For example, Xu et al. designed a contrastive generation and adaptive smoothing network for generalized zero-shot learning for composite fault diagnosis; Yue et al. constructed a semi-supervised hybrid triplet network for industrial fault diagnosis.

[0004] While the aforementioned methods have achieved good diagnostic results to some extent, none of them systematically consider the domain shift problem in generalized zero-shot learning. In narrow zero-shot learning, since the unknown domain is not involved in training, the constructed prototypes will be shifted, and in this case, it is only necessary to consider correcting the shift between the unknown domain prototypes and features. However, in generalized zero-shot learning, the known class prototypes are trained, causing the constructed unknown class prototypes to be more prone to shifting towards the known class, resulting in a significant decrease in recognition accuracy. Therefore, it is necessary to consider the shift problem of both known and unknown prototypes. Compared to narrow zero-shot learning, the domain shift problem in generalized zero-shot learning is more valuable and challenging to study. Summary of the Invention

[0005] In view of this, the purpose of this invention is to provide a generalized zero-shot fault diagnosis method with prototype fusion reconstruction and domain decoupling empowerment. First, a prototype fusion reconstruction (PFR) method is proposed to alleviate the bias problem of unknown prototypes. This method fuses theoretical prototypes and feature prototypes as the initial prototype, and then reconstructs the initial prototype through the clustering prototype of the unknown class during the testing phase. The theoretical prototype originates from prior knowledge of known samples, the feature prototype is extracted by a convolutional autoencoder, and the clustering prototype is obtained by clustering the unknown class using a Gaussian mixture model. The PFR method alleviates the domain bias problem of unknown prototypes by introducing multi-view prototypes of samples. Second, a domain decoupling strategy is constructed to solve the problem of unknown faults being misjudged as known faults. This invention introduces an open set learning strategy into generalized zero-shot fault diagnosis, and achieves the separation of known and unknown domains through an improved OpenMax method. Traditional OpenMax algorithms need to fit a Weibull distribution to identify unknown samples and obtain distribution parameters through maximum likelihood estimation. However, the maximum likelihood estimation of the Weibull distribution often has biases, which are particularly obvious when the sample size is limited. Bias in parameter estimation can affect the accuracy of extreme value distribution estimation, reducing the distinguishability of samples between known and unknown domains. Therefore, this invention proposes an improved OpenMax strategy, introducing the Cramer-Rao lower bound to evaluate the effectiveness of parameter estimation. By using more accurate parameter estimation, sample distinguishability is improved, thereby increasing the overall recognition accuracy.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] A generalized zero-shot fault diagnosis method that combines prototype fusion reconstruction and domain decoupling empowerment, comprising: S1. Collect state signal samples of bearings under different working conditions using a rotating machinery fault test bench; S2. Combining prior knowledge, analyze the fault signals to obtain the theoretical prototypes of the corresponding attributes; S3. Set up a convolutional autoencoder and train it using the MSE of the original features and the reconstructed features as the loss function. After training, the convolutional autoencoder extracts the feature prototype from the hidden layer and combines the theoretical prototype and the feature prototype to obtain the initial prototype. S4. Set up the domain discriminator, train the domain discriminator using known samples, use the maximum likelihood estimation method to obtain the initial distribution parameters of the Weibull distribution, and use the Cramer-Robb lower bound combined with a simple network architecture to correct the fitted distribution parameters. S5. Use the trained domain discriminator to extract features from the test samples, use the Openmax method to distinguish between known and unknown fault classes, and obtain the labels of the known classes. S6. Set up a prototype extractor and train it with the goal of aligning the extracted prototype features with the initial prototype. After training, extract features of unknown fault types. S7. Cluster the features of the extracted unknown faults to obtain a feature clustering prototype, and revise the initial prototype based on the feature clustering prototype. S8. In the revised prototype, nearest neighbor estimation is used to obtain the label of the unknown class.

[0008] Furthermore, in step S1, the state signals of bearings under different types and working conditions are acquired by vibration sensors. The acquired bearing fault types include normal state NC, outer ring fault OF, inner ring fault IF, rolling element fault BF, outer and inner ring fault OI, rolling element and outer ring fault BO, and inner ring and rolling element fault IB. The acquired bearing fault working conditions are divided into different load levels; and the sample is expanded using sliding window sampling.

[0009] Furthermore, in step S2, the RMS value of the fault signal and the attribute related to the fault amplitude and fault frequency obtained through spectrum analysis or bearing fault frequency analysis are combined, and the theoretical prototype of the corresponding attribute is specifically one-hot encoding.

[0010] Furthermore, in step S3, the convolutional autoencoder includes two convolutional modules, two fully connected layer modules, and three deconvolutional modules; it takes the original extracted features as input, the output of the autoencoder's hidden layer as the feature prototype, and uses MSE as the loss function.

[0011] in Indicates original features, is Features reconstructed by the autoencoder This indicates the number of training samples for each category; By connecting theoretical prototypes and feature prototype To obtain the initial prototype:

[0012] in This indicates a matrix join operation.

[0013] Furthermore, in step S4, a domain discriminator is set up, which includes four convolutional modules and three fully connected layer modules. The domain discriminator takes known samples as input and outputs type labels for known and unknown samples. The domain discriminator is trained using known class samples, and the test samples include known and unknown fault types. Cross-entropy loss is used as the loss function. Simultaneously, the initial parameters of the Weibull distribution are obtained using the M-Openmax method. First, the average activation vector for each fault type is obtained. j The average activation vector for each category is represented as:

[0014] in, Indicates the first i The activation vector of each sample. This represents the number of samples for each type; Then, the distance between each activation vector and the average activation vector is calculated, resulting in a distance set; i The sample and the first j The distance between different types of faults is:

[0015] Then, the Weibull distribution from extremum theory is used to fit the distance set for each class, where the probability density function and cumulative distribution function of the Weibull distribution are expressed as follows:

[0016]

[0017] In the formula, and These represent the scale parameter and shape parameter of the distribution, respectively. The parameters of the Weibull distribution are estimated using the maximum likelihood estimation method:

[0018] The Cramer-Rao lower bound is introduced to correct and evaluate the parameter estimation process. First, a lower bound is set for the variance of the unbiased estimator:

[0019] in, The Fisher information matrix is ​​calculated using the second-order partial derivatives of the log-likelihood function.

[0020] The Fisher information matrix is ​​simplified using the Monte Carlo algorithm. :

[0021] in This represents the number of Monte Carlo simulations. The minimum value is taken. Corresponding parameters and As the final estimated parameter.

[0022] Furthermore, in step S5, the trained domain discriminator is used to extract features from the samples, and a parameterized SoftMax function is introduced to adjust the output probability of the domain discriminator, reducing the activation vector scores of each known category. The adjusted activation vector is then expressed as:

[0023] in, J This represents the total number of known categories. j Indexed by category; The activation vector value for an unknown category is defined as the sum of the differences between the original activation vector and the adjusted activation vector:

[0024] The final output probability is obtained through the SoftMax function:

[0025] Finally, a domain decoupling threshold DS is set to assist in the classification decision. The final output category determination rule is as follows:

[0026] in If it exists However, the probability of the unknown category is also unknown. Even if the value is the maximum, the sample will still be classified as an unseen category.

[0027] Furthermore, in step S6, a prototype extractor is set up. The structure of the feature extractor is similar to that of the domain discriminator, both being convolutional neural networks. Known class samples are used as input, and the output features are used as prototype features. During training, the goal is to align the extracted prototype features with the initial prototype. The alignment loss between the prototype features and the initial prototype is:

[0028] In the formula For the extracted features, This is the initial prototype corresponding to this type of sample; the prototype extractor after training is used to predict unknown fault features in the domain decoupling results.

[0029] Furthermore, in step S7, the domain offset problem is mitigated by fusing the structural features of unknown domain samples. The cluster centers of the Gaussian mixture model are used as the cluster prototype, and the distribution of the Gaussian mixture cluster is represented as follows:

[0030] in The mixing coefficient is, and ,parameter and Let the mean vector and covariance matrix of the i-th Gaussian distribution be represented by the mean vector. Constructing a feature clustering prototype ; The corrected prototype is then represented as:

[0031] In the formula This is a correction factor.

[0032] Furthermore, in step S8, nearest neighbor estimation is used to obtain the label of the unknown class:

[0033] Finally, the label of the unknown class is obtained. By combining the labels of known classes, fault identification can be completed.

[0034] The beneficial effects of this invention are as follows: First, this invention proposes a Prototype Fusion Reconstruction (PFR) method to alleviate the bias problem of unknown prototypes. This method fuses theoretical and feature prototypes as the initial prototype, and then reconstructs the initial prototype during the testing phase using clustered prototypes of unknown classes. The theoretical prototype originates from prior knowledge of known samples, the feature prototype is extracted by a convolutional autoencoder, and the clustered prototype is obtained through Gaussian mixture model clustering of unknown classes. The PFR method alleviates the neighborhood bias problem of unknown prototypes by introducing multi-view prototypes of samples.

[0035] Secondly, this invention constructs a domain decoupling strategy to address the problem of unknown faults being misclassified as known faults. This invention introduces an open-set learning strategy into generalized zero-shot fault diagnosis, achieving separation between known and unknown domains through an improved OpenMax method. Traditional OpenMax algorithms require fitting a Weibull distribution to identify unknown samples and obtaining distribution parameters through maximum likelihood estimation. However, maximum likelihood estimation of the Weibull distribution often suffers from bias, particularly noticeable when the sample size is limited. Bias in parameter estimation affects the accuracy of extreme value distribution estimation, reducing the distinguishability of samples between known and unknown domains. Therefore, this invention proposes an improved OpenMax strategy, introducing a Cramer-Rao lower bound to evaluate the effectiveness of parameter estimation. More accurate parameter estimation enhances sample distinguishability, thereby improving overall recognition accuracy.

[0036] 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

[0037] 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 This refers to the zero-shot learning process under different existing paradigms, among which, Figure 1 (a) represents narrow zero-shot learning. Figure 1 (b) represents generalized zero-shot learning; Figure 2 This is a schematic diagram of the overall architecture of the generalized zero-sample fault diagnosis method with prototype fusion reconstruction and domain decoupling empowerment according to an embodiment of the present invention. Figure 3 This is a detailed schematic diagram illustrating the domain decoupling process according to an embodiment of the present invention; Figure 4 This is a schematic diagram of the confusion matrix of the PFR-DD model under four load conditions on the NEEPU test platform according to an embodiment of the present invention, wherein, Figure 4 (a) has a load of 0. Figure 4 (b) has a load of 1. Figure 4 (c) has a load of 2. Figure 4 (d) has a load of 3; Figure 5 This is a schematic diagram of the confusion matrix of the PFR-DD model under four load conditions on the BCF test platform according to an embodiment of the present invention, wherein, Figure 5 (a) has a load of 0. Figure 5 (b) has a load of 1. Figure 5(c) has a load of 2. Figure 5 (d) has a load of 3; Figure 6 This is a schematic diagram of the confusion matrix of the PFR-DD model under four different operating conditions on the BJTU-RAO experimental platform with multimodal input, according to an embodiment of the present invention. Figure 6 (a) represents operating condition 0; Figure 6 (b) is operating condition 1; Figure 6 (c) represents operating condition 2; Figure 6 (d) represents operating condition 3; Figure 7 This is a schematic diagram showing the ablation experimental results of the PFR-DD model under different load conditions in the NEEPU and BCF experimental platforms according to embodiments of the present invention. Figure 7 (a) Results from the NEEPU test platform Figure 7 (b) Results from the BCF test platform; Figure 8 This diagram illustrates the experimental results of the PFR-DD model under load 0 and load 1 conditions in the NEEPU and BCF experiments, respectively, according to embodiments of the present invention. Figure 8 (a) NEEPU test platform with 0 load; Figure 8 (b) Load condition 1 for the NEEPU test platform; Figure 8 (c) BCF test platform with 0 load; Figure 8 (d) shows the load condition of the BCF test platform 1. Detailed Implementation

[0038] 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.

[0039] 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.

[0040] 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.

[0041] Please see Figures 2-8 This is a generalized zero-sample fault diagnosis method that enables prototype fusion reconstruction and domain decoupling.

[0042] Example 1 This embodiment provides a detailed implementation process for a generalized zero-shot fault diagnosis method based on prototype fusion reconstruction and domain decoupling empowerment, such as... Figure 2 As shown, it includes the following steps: S1: Collect state signal samples of bearings under different conditions and of different types using a rotating machinery fault test bench; S2: Combining prior knowledge, analyze fault signals to obtain the theoretical prototype of the corresponding attributes; S3: By using a convolutional autoencoder, feature prototypes are extracted and combined with theoretical prototypes to obtain initial fault prototypes; S4: Train the domain discriminator using known class samples, and obtain the initial distribution parameters using the M-Openmax method; then use Cramer-Rao lower bound (CRLB) combined with a simple network architecture to correct the fitted distribution parameters; S5: Use the trained domain discriminator to extract features from the test samples, use the Openmax method to distinguish between known and unknown fault classes, and obtain the labels of the known classes. S6: Train the prototype extractor, extract the corresponding features, align the extracted prototype features with the initial prototype, and use the prototype extractor to extract the features predicted as unknown type faults in S5. S7: Use unsupervised clustering and matching algorithms to correct the prototypes of unknown classes; S8: Combine nearest neighbor estimation to infer the unknown class label.

[0043] In step S1 of this embodiment, the state signals of bearings under different types and working conditions are acquired by vibration sensors. The acquired bearing fault types include at least normal state (NC), outer ring fault (OF), inner ring fault (IF), rolling element fault (BF), outer ring and inner ring fault (OI), rolling element and outer ring fault (BO), and inner ring and rolling element fault (IB). The acquired bearing fault working conditions are divided into at least 4 load levels (0.0, 0.1, 0.2, 0.3 N·m), and the sample is expanded using sliding window sampling.

[0044] In step S2 of this embodiment, the RMS value of the fault signal and the attribute related to the fault amplitude and fault frequency are obtained by combining the spectrum analysis or the fault frequency of the bearing. The theoretical prototype of the obtained attribute is specifically one-hot encoding.

[0045] In step S3 of this embodiment, the convolutional autoencoder takes the original extracted features as input, the output of the autoencoder hidden layer as the feature prototype, and uses MSE as the loss function.

[0046] in Indicates original features, is Features reconstructed by the autoencoder This indicates the number of training samples for each category.

[0047] By connecting theoretical prototypes and feature prototype You can obtain the initial prototype:

[0048] in This indicates a matrix join operation.

[0049] In step S4 of this embodiment, a domain discriminator is set up. The domain discriminator takes known samples as input and outputs type labels for known and unknown samples. Specifically, the domain discriminator is trained using known class samples, and the test samples include both known and unknown fault types. Cross-entropy loss is used as the loss function.

[0050] Simultaneously, the initial parameters of Weibull are obtained using the M-Openmax method. The first step of the M-OpenMax algorithm is to obtain the average activation vector for each fault type. In this step, training is performed using only data from the known domain. j The average activation vector for each category can be calculated using the following formula:

[0051] in, Indicates the firsti The activation vector of each sample. This represents the number of samples for each type.

[0052] By calculating the distance between each activation vector and the average activation vector, a distance set can be obtained. i The sample and the first j The distance between different types of faults can be calculated using the following formula:

[0053] We use the Weibull distribution from extremum theory to fit the distance set for each class. The probability density function of the Weibull distribution is as follows:

[0054] In the formula, and These represent the scale parameter and shape parameter of the distribution, respectively.

[0055] The cumulative distribution function of the Weibull distribution can be expressed as:

[0056] In the traditional OpenMax algorithm, maximum likelihood estimation is typically used to estimate the parameters of the Weibull distribution. Its likelihood function expression is:

[0057] This invention then proposes introducing the Cramer-Rao lower bound to correct and evaluate the parameter estimation process. The CRLB sets a lower bound on the variance of the unbiased estimator, expressed as follows:

[0058] in, The Fisher information matrix is ​​obtained by calculating the second-order partial derivative of the log-likelihood function:

[0059] The formula for calculating the second-order partial derivative is:

[0060]

[0061]

[0062] However, directly calculating the Fisher information matrix The expectation is extremely challenging. Therefore, the Monte Carlo algorithm is used to simplify the calculation, which can be simplified to the following form:

[0063] in This represents the number of Monte Carlo simulations. The minimum value is taken. Corresponding parameters and As the final estimated parameter.

[0064] By introducing CRLB to measure and constrain the effectiveness of parameter estimation, the M-OpenMax method proposed in this invention can obtain more accurate Weibull distribution parameter estimates, thereby improving the ability to identify unknown samples and ultimately improving the overall performance of generalized zero-sample fault diagnosis.

[0065] In step S5 of this embodiment, a parameterized SoftMax function is introduced to further adjust the model's output probabilities. This function aims to reduce the activation vector scores of each known class and assign the reduced score difference to the unknown class. The adjusted activation vectors are calculated as follows:

[0066] in, J This represents the total number of known categories. j For category indexing.

[0067] The activation vector value for an unknown category is defined as the sum of the differences between the original activation vector and the adjusted activation vector, as shown in the following formula:

[0068] The final output probability can be obtained using the SoftMax function:

[0069] Finally, a domain decoupling threshold DS is set to assist in the classification decision. The final output category determination rule is as follows:

[0070] in It should be noted that the classification of a sample into a known or unknown category is not solely determined by the aforementioned threshold criteria. If there are... However, the probability of the unknown category is also unknown. In the case of the maximum value, the sample will still be classified as an unseen category. The detailed process of domain decoupling in S4-S5 is as follows: Figure 3 As shown.

[0071] In step S6 of this embodiment, a prototype extractor is set up. The structure of the feature extractor is similar to that of the domain discriminator, both being convolutional neural networks. Known class samples are used as input, and the output features are used as prototype features. During training, the goal is to align the extracted prototype features with the initial prototype. The alignment loss between the prototype features and the initial prototype is:

[0072] In the formula For the extracted features, This serves as the initial prototype for this type of sample. The prototype extractor, after training, is used to predict unknown fault features in the domain decoupling results.

[0073] In step S7 of this embodiment, since there is no data from the unknown class during training, the corresponding prototype will inevitably be transferred to the unknown class. This invention introduces a feature clustering prototype to correct this transfer. This method alleviates the domain shift problem by fusing structural features of samples from the unknown domain. During the testing phase, the extracted features are clustered to obtain a clustering prototype. The proposed model uses the cluster centers of a Gaussian mixture model as the prototype. The distribution of Gaussian mixture clustering can be defined as follows:

[0074] in The mixing coefficient is, and .parameter and Let represent the mean vector and covariance matrix of the i-th Gaussian distribution. In this module, the mean vector is used. Constructing a feature clustering prototype .

[0075] The revised prototype can be represented as:

[0076] In the formula This is a correction factor.

[0077] In step S8 of this embodiment, nearest neighbor estimation is finally used to obtain the label of the unknown class:

[0078] This completes the fault identification process.

[0079] Example 2 To evaluate the effectiveness of the method proposed in Example 1, this embodiment conducts experiments on three different test benches to assess the performance of the proposed PFR-DD model. The first experiment is based on the bearing dataset (NEEPU) collected by Northeast Electric Power University. The second experiment uses data from a self-made bearing composite fault test platform (BCF) for validation. The third experiment is based on the bogie dataset (BJTU-RAO) from Beijing Jiaotong University. Furthermore, ablation experiments are used to further validate different modules.

[0080] The experiment compared six classic and state-of-the-art algorithms, including: Direct Attribute Prediction (DAP), Fault Description-Based Attribute Transfer (FDAT), Semantic Consistent Embedding (SCE), Zero-Shot Attribute Embedding Composite Fault Diagnosis Model (ZSAECFD), and two GZSL methods: Adaptive Weighted Semantic Autoencoder (AWSAE) and Contrastive Generative and Adaptive Smoothing Network (CGASNet).

[0081] The NEEPU experimental platform includes a motor, shaft, data acquisition unit, magnetic brake, and bearing testing module. The experiment includes seven bearing health states: normal (NC), outer ring fault (OF), inner ring fault (IF), rolling element fault (BF), outer and inner ring fault (OI), rolling element and outer ring fault (BO), and inner ring and rolling element fault (IB). Four load levels (0.0, 0.1, 0.2, and 0.3 N·m) were tested, corresponding to loads 0-3 respectively.

[0082] Table 1 shows three average diagnostic accuracy metrics.

[0083] Table 1

[0084] Where K represents the accuracy of known (single) faults, U represents the accuracy of unknown (compound) faults, and A represents the overall accuracy. The values ​​presented in Table 1 are the average results of five experiments. Taking ".980±.008" as an example, it represents a diagnostic accuracy of 0.980 and a standard deviation of 0.008. Due to the use of traditional machine learning techniques, the overall diagnostic accuracy of the FDAT and SCE methods is relatively low. The DAP and ZSAECFD models only focus on faults in the unknown domain and fail to effectively distinguish between known and unknown faults, resulting in diagnostic results that are significantly biased towards known faults. The two generalized zero-shot diagnostic baseline methods, AWSAE and CGANet, are characterized by their relatively low ability to distinguish between known and unknown faults and their generally low diagnostic accuracy for unknown faults, resulting in a low overall diagnostic accuracy. In contrast, although the PFR-DD model proposed in this invention sacrifices slightly in the diagnostic accuracy of single faults, its overall diagnostic accuracy far exceeds that of other methods. Figure 4 The diagram shows the confusion matrix of the PFR-DD model under four loading conditions. Figure 4 (a) has a load of 0. Figure 4 (b) has a load of 1. Figure 4 (c) has a load of 2. Figure 4(d) has a load of 3; the confusion matrices under four different load conditions demonstrate that the PFR-DD model maintains high diagnostic accuracy for single faults while also achieving satisfactory accuracy in detecting complex faults. Particularly noteworthy is that these complex faults are completely excluded during training. This performance, capable of crossing the boundary between known and unknown faults, fully demonstrates the enormous potential of the proposed PFR-DD model.

[0085] In addition, further experiments were conducted on a self-made BCF testing platform to further verify the generalization ability of the proposed PFR-DD model. The experimental setup consisted of five key components: a load motor, a torque and speed measuring instrument, a planetary gear reducer, a bearing testing module, and a drive motor. This study evaluated seven bearing health conditions (NC, OF, IF, BF, OI, BO, and IB) using SKF 61908 bearings. Four load conditions were applied via the load motor: load 0: no load, load 1: 2 N·m, load 2: 4 N·m, and load 3: 6 N·m. All tests were conducted at a constant speed of 500 rpm and a sampling frequency of 12 kHz. The same data partitioning scheme and hyperparameter configuration were used for both the proposed model and the comparative model to ensure methodological consistency.

[0086] Table 2 lists the average diagnostic accuracy under four different load conditions.

[0087] Table 2

[0088] As shown in Table 2, the overall diagnostic accuracy of the PFR-DD model proposed in this invention under four different operating conditions is 82.6%, 88.3%, 95.6%, and 92.5%, respectively. Compared with other methods, the PFR-DD model exhibits superior domain decoupling capability and unknown fault identification capability. Therefore, this model ensures high diagnostic accuracy for single faults while maintaining a fairly high diagnostic accuracy for complex faults. Furthermore, Figure 6 The confusion matrices of the PFR-DD model under four different operating conditions on the BCF experimental platform are shown. Figure 6 (a) represents operating condition 0; Figure 6 (b) is operating condition 1; Figure 6 (c) represents operating condition 2; Figure 6 (d) represents operating condition 3. The confusion matrix for diagnostic accuracy for each fault type is shown. Overall, the model demonstrates satisfactory diagnostic accuracy across different operating conditions, particularly under loads 2 and 3. Even without training on composite faults and subjected to interference from single faults, the combined diagnostic accuracy for these two load conditions exceeds 90%, fully demonstrating the potential of the proposed model.

[0089] To verify the model's generalization ability under multimodal inputs, this embodiment conducted experiments on the BJTU-RAO dataset. Six fault types were tested: normal, inner race fault, outer race fault, ball fault, combined inner and outer race fault, and combined outer race and ball fault. For each fault type, data from three different sensors were used, including X-direction vibration signals, Y-direction vibration signals, and acoustic sensor signals. The experiments were conducted under four different operating conditions: Condition 0: 20 Hz speed, 0 kN load; Condition 1: 40 Hz speed, 0 kN load; Condition 2: 60 Hz speed, 0 kN load; Condition 3: 20 Hz speed, 10 kN load (load direction towards the motor side). The sampling frequency was 64 kHz. Similar to the experiments on the NEEPU and BCF datasets, corresponding fault attributes were defined for different fault types, including fault category, fault amplitude, and fault frequency. Fault amplitude can be calculated using the root mean square value. Since characteristic fault frequencies cannot be calculated from acoustic signals, this embodiment uses the same attribute type as the fault category. Furthermore, to better adapt to multimodal data, corresponding improvements were made to the neighborhood discriminator and prototype extractor. For the neighborhood discriminator, three identical convolutional feature extractors were used to extract features from different sensors, followed by feature-level fusion through a fully connected layer. Similarly, the prototype extractor also used a fully connected layer to perform feature-level fusion on the features extracted from the three types of signals. For each operating condition, the experiment was repeated five times, and the confusion matrix of one trial under each of the four operating conditions is shown below. Figure 8 As shown, the diagnostic accuracy for each category consistently remains above 90%, and the overall accuracy exceeds 85% in all cases, effectively demonstrating the strong generalization ability of the PFR-DD model under multimodal data input.

[0090] To evaluate the contributions of the three core components—the PFR strategy, cluster correction, and the M-OpenMax module—ablation experiments were conducted under eight load conditions on two datasets. The results are as follows: Figure 7 As shown, Figure 7 The diagram shows the ablation experimental results of the PFR-DD model under different load conditions in the NEEPU and BCF experimental platforms. Figure 7 (a) shows the experimental results from the NEEPU test platform. Figure 7(b) shows the experimental results of the BCF experimental platform; where "WP" represents the model without the PFR strategy, "WM" represents the model using the unmodified standard OpenMax method, and "WC" represents the model without cluster-based prototype correction. In all eight subgraphs, the three modules work synergistically but each has its own function: the PFR strategy primarily improves the ability to identify faults in unknown domains, while cluster correction provides further support; in contrast, M-OpenMax affects the accuracy of identification in both known and unknown domains. These components work together to enable the proposed PFR-DD framework to achieve excellent overall diagnostic performance.

[0091] Domain decoupling threshold The ability to distinguish between known and unknown domains plays an auxiliary role in the diagnostic process. Therefore, this embodiment investigates the impact of different threshold settings on the overall diagnostic accuracy. The threshold was adjusted from 0 to 0.8 in steps of 0.1: when the threshold was set to 0, the classification result completely depended on the output of M-OpenMax; while setting the threshold to 1 would cause all samples to be classified as unknown. Experimental results under load 0 and load 1 conditions in NEEPU and BCF experiments are shown below. Figure 8 As shown, when the threshold is not set too high, the accuracy fluctuations under the four operating conditions are small, indicating that the proposed PFR-DD method has low parameter sensitivity. In practical applications, even without setting a threshold (i.e., setting the threshold to 0), a high diagnostic accuracy can still be obtained. In the NEEPU experiment, the highest accuracy was achieved with a threshold of 0.5; in the BCF experiment, the optimal threshold for load 0 was 0.6, and the optimal threshold for load 1 was 0.5. Based on this, the domain decoupling threshold was uniformly set to 0.5 in both the NEEPU and BCF experiments.

[0092] 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 generalized zero-shot fault diagnosis method that combines prototype fusion reconstruction and domain decoupling empowerment, characterized in that: The method includes: S1. Collect state signal samples of bearings under different working conditions using a rotating machinery fault test bench; S2. Combining prior knowledge, analyze the fault signals to obtain the theoretical prototypes of the corresponding attributes; S3. Set up a convolutional autoencoder and train it using the MSE of the original features and the reconstructed features as the loss function. After training, the convolutional autoencoder extracts the feature prototype from the hidden layer and combines the theoretical prototype and the feature prototype to obtain the initial prototype. S4. Set up the domain discriminator, train the domain discriminator using known samples, use the maximum likelihood estimation method to obtain the initial distribution parameters of the Weibull distribution, and use the Cramer-Robb lower bound combined with a simple network architecture to correct the fitted distribution parameters. S5. Use the trained domain discriminator to extract features from the test samples, use the Openmax method to distinguish between known and unknown fault classes, and obtain the labels of the known classes. S6. Set up a prototype extractor and train it with the goal of aligning the extracted prototype features with the initial prototype. After training, extract features of unknown fault types. S7. Cluster the features of the extracted unknown faults to obtain a feature clustering prototype, and revise the initial prototype based on the feature clustering prototype. S8. In the revised prototype, nearest neighbor estimation is used to obtain the label of the unknown class.

2. The generalized zero-shot fault diagnosis method based on prototype fusion reconstruction and domain decoupling empowerment as described in claim 1, characterized in that: In step S1, the state signals of bearings under different types and working conditions are acquired by vibration sensors. The acquired bearing fault types include normal state NC, outer ring fault OF, inner ring fault IF, rolling element fault BF, outer and inner ring fault OI, rolling element and outer ring fault BO, and inner ring and rolling element fault IB. The acquired bearing fault working conditions are divided into different load levels. The sample is expanded using sliding window sampling.

3. The generalized zero-shot fault diagnosis method based on prototype fusion reconstruction and domain decoupling empowerment as described in claim 2, characterized in that: In step S2, the RMS value of the fault signal and the attributes related to the fault amplitude and fault frequency obtained through spectrum analysis or bearing fault frequency analysis are combined, and the theoretical prototype of the corresponding attribute is specifically one-hot encoding.

4. The generalized zero-shot fault diagnosis method based on prototype fusion reconstruction and domain decoupling empowerment as described in claim 3, characterized in that: In step S3, the convolutional autoencoder includes two convolutional modules, two fully connected layer modules, and three deconvolutional modules; it takes the original extracted features as input, the output of the autoencoder's hidden layer as the feature prototype, and uses MSE as the loss function. in Indicates original features, is Features reconstructed by the autoencoder This indicates the number of training samples for each category; By connecting theoretical prototypes and feature prototype To obtain the initial prototype: in This indicates a matrix join operation.

5. The generalized zero-shot fault diagnosis method based on prototype fusion reconstruction and domain decoupling empowerment according to claim 4, characterized in that: In step S4, a domain discriminator is set up. The domain discriminator includes four convolutional modules and three fully connected layer modules. The domain discriminator takes known samples as input and outputs type labels for known and unknown samples. The domain discriminator is trained using known class samples. The test samples include known and unknown fault types, and cross-entropy loss is used as the loss function. Simultaneously, the initial parameters of the Weibull distribution are obtained using the M-Openmax method. First, the average activation vector for each fault type is obtained. j The average activation vector for each category is represented as: in, Indicates the first i The activation vector of each sample. This represents the number of samples for each type; Then, the distance between each activation vector and the average activation vector is calculated, resulting in a distance set; i The sample and the first j The distance between different types of faults is: Then, the Weibull distribution from extremum theory is used to fit the distance set for each class, where the probability density function and cumulative distribution function of the Weibull distribution are expressed as follows: In the formula, and These represent the scale parameter and shape parameter of the distribution, respectively. The parameters of the Weibull distribution are estimated using the maximum likelihood estimation method: The Cramer-Rao lower bound is introduced to correct and evaluate the parameter estimation process. First, a lower bound is set for the variance of the unbiased estimator: in, The Fisher information matrix is ​​calculated using the second-order partial derivatives of the log-likelihood function. The Fisher information matrix is ​​simplified using the Monte Carlo algorithm. : in Indicates the number of Monte Carlo simulations; take the minimum. Corresponding parameters and As the final estimated parameter.

6. The generalized zero-shot fault diagnosis method based on prototype fusion reconstruction and domain decoupling empowerment according to claim 5, characterized in that: In step S5, the trained domain discriminator is used to extract features from the samples, and a parameterized SoftMax function is introduced to adjust the output probability of the domain discriminator, reducing the activation vector scores of each known class. The adjusted activation vector is then expressed as: in, J This represents the total number of known categories. j Indexed by category; The activation vector value for an unknown category is defined as the sum of the differences between the original activation vector and the adjusted activation vector: The final output probability is obtained through the SoftMax function: Finally, a domain decoupling threshold DS is set to assist in the classification decision. The final output category determination rule is as follows: in If it exists However, the probability of the unknown category is also unknown. Even if the value is the maximum, the sample will still be classified as an unseen category.

7. The generalized zero-shot fault diagnosis method based on prototype fusion reconstruction and domain decoupling empowerment according to claim 6, characterized in that: In step S6, a prototype extractor is set up. The structure of the feature extractor is similar to that of the domain discriminator, both being convolutional neural networks. Known class samples are used as input, and the output features are used as prototype features. During training, the goal is to align the extracted prototype features with the initial prototype. The alignment loss between the prototype features and the initial prototype is: In the formula For the extracted features, This is the initial prototype corresponding to this type of sample; The prototype extractor, after training, is used to predict unknown fault features in the domain decoupling results.

8. The generalized zero-shot fault diagnosis method based on prototype fusion reconstruction and domain decoupling empowerment according to claim 7, characterized in that: In step S7, the domain offset problem is mitigated by fusing the structural features of unknown domain samples. The cluster centers of the Gaussian mixture model are used as the cluster prototype, and the distribution of the Gaussian mixture cluster is represented as follows: in The mixing coefficient is, and ,parameter and Let the mean vector and covariance matrix of the i-th Gaussian distribution be represented by the mean vector. Constructing a feature clustering prototype ; The corrected prototype is then represented as: In the formula This is a correction factor.

9. The generalized zero-shot fault diagnosis method based on prototype fusion reconstruction and domain decoupling empowerment according to claim 8, characterized in that: In step S8, nearest neighbor estimation is used to obtain the label of the unknown class: Finally, the label of the unknown class is obtained. By combining the labels of known classes, fault identification can be completed.

Images