Rotating machinery real damage diagnosis method based on hybrid multi-source domain generalization network model

Abstract

The application provides a rotating machinery real damage diagnosis method based on a hybrid multi-source domain generalization network model, and is used for a new hybrid multi-source domain generalization network model under unknown real working conditions. The core and key of the model is to learn the invariant representation of the feature information of different environment information, and the inner product of the similar feature representation is higher than that of other features through feature discrimination, so that the diagnosis model can learn the feature robust information of a specific working environment and generalize to the unseen working environment. The experimental results of the experiment on the German Paderborn rolling bearing data set prove the effectiveness of the method. Further, through the establishment of a planetary parallel shaft gearbox bearing comprehensive fault simulation test bed, the experimental results of a large number of diagnosis tasks of the collected real data are obtained, and it is verified that the method has improvement in the generalization performance under unknown real working conditions.

Description

Technical Field

[0001] This application belongs to the field of rotating machinery fault diagnosis, and specifically relates to a method for diagnosing actual damage to bearings. Background Technology

[0002] Statistics show that 40% to 70% of electromechanical drive system and motor failures are caused by rolling bearing damage, which can lead to high application costs due to downtime. Therefore, monitoring the operating condition of bearings is crucial in high-risk or high-maintenance applications. Bearing damage detection is typically achieved through vibration analysis using accelerometers. Traditional physics-based fault diagnosis methods using vibration signals have proven effective in detecting faults under defined operating conditions, but they require training specific diagnostic methods for different operating conditions and demand a high level of expert experience and precise mathematical and physical knowledge. This limits their ability to diagnose early, potential faults and currently cannot handle situations in unknown domains, such as... Figure 1A As shown.

[0003] Leveraging the advantages of big data models—the ability to utilize all complete historical data and process it rapidly—data-driven fault diagnosis methods have gradually emerged. Classical data-driven methods include support vector machine models for binary classification and kernel function methods for nonlinear problems, enabling fault state identification and classification. However, due to the complexity of existing devices, the primary requirement is to analyze and process vibration data during operation to diagnose system faults without needing to know the system's precise analytical model. Therefore, methods such as... Figure 1B Domain adaptation (DA), a technique in the field of transfer learning, has been used to address these cross-domain diagnostic tasks. The core of existing DA-based cross-domain fault diagnosis research is the application of knowledge learned from the labeled source domain to the unlabeled target domain. For example, Zhang et al. used an adversarial adaptive learning process to optimize the base CNN for adaptation across different domains. Guo et al. integrated conditional recognition and domain adaptation, achieving machine-intelligent fault diagnosis through maximum mean difference (MMD). Li et al. introduced multi-layer MMD to further align target domain features with source domain features. However, in real-world industrial scenarios, mechanical fault data is often invisible or lacks prior knowledge, degrading the effectiveness of domain adaptation as a transfer diagnostic approach when only minimizing the difference between the source and target domains is considered. Because the unknown domain extends beyond the known domain, and collecting labeled samples under various operating conditions to train the diagnostic model is extremely difficult, if not impossible, these factors can lead to high misdiagnosis rates.

[0004] In recent years, due to the limited generalization ability of domain adaptation, researchers have further explored and improved domain adaptation into domain generalization. Domain generalization methods have been widely applied in fields such as computer vision, natural language processing, and healthcare. Figure 1C As shown, this is a more practical but challenging scenario. The red and blue areas represent existing domains, while the green area represents unseen domains. Since there are no prior distribution estimates available for unseen target tasks, the key is to mine invariant feature representations of the domains contained in the available data from specific source data, enabling these feature representations to generalize to the discrimination of unknown domain faults. However, few scholars have applied this to real and artificial data for fault diagnosis. To this end, it is further considered that even for the same type of rolling bearing fault, the data collected by sensors will differ due to the variability of the working environment (artificial or real) caused by variations in factors such as radial force, rotational speed, and ambient temperature. Especially in industrial applications where specific faults occur, the system must be shut down immediately for maintenance, making it difficult to collect enough fault samples for training deep learning models. Moreover, in industrial applications using inexpensive small electric motors with high power consumption (approximately 1kW) or lower, vibration-based methods would incur the cost of additional sensors, which is economically infeasible.

[0005] Previous work focused on designing generalized features by extracting features from vibration signals collected by sensors. The aim was to find fault features robust to changes in vibration signals under specific operating conditions. Some existing theories suggest that domain-invariant feature representations are general and can be transferred to different domains. Some researchers have used transfer learning to assist in the diagnosis of real-world scenarios with insufficient data; others have proposed a hybrid deep transfer learning method to address the problem of limited labeled bearing fault samples. However, due to the non-smooth sample distribution between the source and target domains, using vibration signal data augmentation methods alone is not accurate enough for fault diagnosis. Researchers have further explored domain-invariant representation learning methods. Liao et al. investigated adversarial learning between labeled and unlabeled source data to learn domain-invariant features. Someone proposed an adversarial generalization network combining feature normalization to learn domain-invariant representations from multiple sources. Another researcher developed a deep DG network using correlation alignment and triplet loss to explore intra-domain features.

[0006] While the above studies have made some progress in fault diagnosis, they still face the following challenges. First, insufficient feature learning. Research suggests that learning domain-invariant representations can be considered as "learning and summarizing" information from a known domain into a shared feature representation, discarding domain-specific features that may cause domain shifts. However, these discarded domain-specific features may carry fault-sensitive information beneficial for diagnosis. Therefore, this leads to a decrease in model performance for each source domain. Second, the above methods mainly focus on fault studies of vibration signals under artificial damage, while fault diagnosis in actual working conditions is usually more variable and challenging. Utilizing only vibration signals, along with discarded features, results in limited feature diversity, thus the model may face the risk of overfitting to the source domain, leading to poor generalization ability. Third, the distinguishability of feature information representations across different domains is low.

[0007] Besides feature extraction, discriminability is another key criterion for achieving domain generalization by further distinguishing feature information from different domains. By considering the compactness between the same class and the separability between different classes, the discriminability of the learned feature representations can be further enhanced.

[0008] In recent years, domain adaptation methods have become increasingly popular in the field of rotating machinery fault diagnosis. However, existing methods require a large amount of labeled prior data distribution. In actual operating condition diagnosis, prior data is often unavailable. Furthermore, significant differences have been found between artificially generated damage datasets (accelerated by artificial damage) and real damage datasets (generated using actual real damage). Therefore, applying models developed based on artificial damage datasets, combined with domain adaptation methods, to industrial data containing real damage is a challenging task. Summary of the Invention

[0009] The purpose of this application is to address the above challenges and propose an improved method for diagnosing faults in rotating machinery.

[0010] To achieve the above objectives, some embodiments of this application provide a method for real damage diagnosis of rotating machinery based on a hybrid multi-source domain generalization network model. The method first designs a hybrid multi-source domain generalization network model for multi-source domain operating condition data, which combines multi-source data augmentation with domain feature information representation. Considering the characteristics of multi-source domains, a contrastive loss of multi-source data augmentation and new feature information is designed to find a domain-invariant class representation, thereby providing the model with a robust and discriminative representation.

[0011] Specifically, the multi-source data augmentation network model proposed in this application generates augmented data by mixing inter-domain data alignment distributions, enabling the augmented data to blend information from multiple source domains and achieve information interaction between these source domains. This augmented data carries and integrates some private information from multiple source domains, which will make the entire data distribution smooth and continuous.

[0012] In some embodiments, the new feature information contrast loss maximizes the common information between features of the same class in different source domains, while minimizing the dissimilar information between features in different source domains.

[0013] In some embodiments, the proposed Hybrid Multi-Source Domain Generalization Network (MDFNet model) consists of two main parts. The first main part is a hybrid enhanced source domain feature extraction network, designed to fully utilize shared information from different types of signal domains for feature extraction and classification. The second main part maximizes the overlap of common information between different classes while minimizing information between different classes. In this way, it can learn domain-independent class representations that can be transferred to new, unseen domains.

[0014] The main beneficial effects of this application include:

[0015] Some embodiments of this application propose a practical fault diagnosis scenario for the model, namely, that the model can be generalized to a real, unseen domain without accessing its data distribution. Compared with traditional data diagnosis methods, the proposed generalization method is more suitable for real-world diagnostic tasks.

[0016] Some embodiments of this application employ a diagnostic framework that can be tailored to different mechanical data. This framework not only fully utilizes vibration and motor current signals to detect bearing damage but also handles situations where only vibration or motor current signals are collected for small rotating machinery. The proposed MDFNet model introduces a novel multi-source domain data augmentation method, enhancing the diversity of invariant representations of mixed feature information from different source domains. Furthermore, a novel feature information contrast loss is designed to maximize the mutual information between the same feature information class in different source domains to consider cross-source domain relationships, while minimizing the learned representation of changes in different feature information within the same source domain. This improves the maximum utilization of data from different source domains and the robustness of common feature information.

[0017] Some embodiments of this application include experiments on the Paderborn rolling bearing dataset in Germany to demonstrate the effectiveness of the proposed model. Furthermore, by constructing a comprehensive fault simulation test bench for planetary parallel shaft gearbox bearings, and through further experimental diagnostic tasks, the results also verify that the model outperforms the comparative model when facing unknown real-world operating conditions. Attached Figure Description

[0018] Figure 1A This is a schematic diagram of existing vibration signal fault diagnosis methods;

[0019] Figure 1B This is a schematic diagram of existing fault diagnosis methods;

[0020] Figure 1C This is a schematic diagram of the rotating machinery fault diagnosis method based on a hybrid multi-source domain generalized network model proposed in this application.

[0021] Figure 2 This is a schematic diagram of the overall framework of the MDFNet model of the rotating machinery fault diagnosis method based on the hybrid multi-source domain generalized network model of this application;

[0022] Figure 3 This is a schematic diagram of the loss function and optimization framework of the rotating machinery fault diagnosis method based on the hybrid multi-source domain generalized network model of this application;

[0023] Figure 4A This is a schematic diagram of artificial damage to a bearing;

[0024] Figure 4B This is a schematic diagram of another type of artificial damage to bearings;

[0025] Figure 4C This is a schematic diagram of actual bearing damage;

[0026] Figure 5 The spectrum diagram of bearing data distribution after spectral processing for artificial and natural damage;

[0027] Figure 6 This is a classification diagram showing the results of comparing vibration signals from different axes mixed in different network frameworks with vibration signals from a single axis. Vx axis refers to the classification diagram of vibration along the x-axis, Vz axis refers to the classification diagram of vibration along the z-axis, and V-mix refers to the classification diagram of vibrations from different axes mixed together.

[0028] Figure 7 This is a schematic diagram showing the results of a verification experiment of the rotating machinery fault diagnosis method based on a hybrid multi-source domain generalized network model proposed in this application.

[0029] Figure 8 This is a schematic diagram showing the feature visualization effect obtained by visualizing the results of a representative method trained with hybrid fusion and enhanced data in damage under unknown real working conditions and comparing it with the method of this application;

[0030] Figure 9 This is a diagram showing the effect comparison between including and not including FICL;

[0031] Figure 10 This is a schematic diagram of the distribution of hyperparameters in the rotating machinery fault diagnosis method based on a hybrid multi-source domain generalized network model according to this application;

[0032] Figure 11This is a schematic diagram showing the ablation experimental results of hyperparameter β under different frameworks of the rotating machinery fault diagnosis method based on the hybrid multi-source domain generalized network model according to this application. Detailed Implementation

[0033] The specific embodiments of this application will be described in detail below with reference to the accompanying drawings.

[0034] The rotating machinery real damage diagnosis method based on a hybrid multi-source domain generalization network model proposed in some embodiments of this application proposes a new hybrid multi-source domain network model (MDFNet model) for unknown real working conditions. The core of this model is to capture the feature information of different environmental information invariant representations, and at the same time, through feature discrimination, make the inner product of similar feature representations average higher than other features, so that the diagnostic model can learn the feature robust information of specific working environments (i.e. specific distributions, such as artificial damage datasets) and generalize to unseen working environments (e.g. different distributions, such as real damage datasets).

[0035] Experimental results using the Paderborn rolling bearing dataset in Germany confirm the effectiveness of the proposed method. Further experiments, conducted using a comprehensive fault simulation test bench for planetary parallel shaft gearbox bearings and based on a large amount of real-world data collected, demonstrate that the method does indeed improve its generalization performance under unknown real-world conditions.

[0036] Preliminary and system models:

[0037] In this application, the purpose is to enable diagnostic tasks to be carried out in actual working conditions. However, in actual working conditions, there may be situations where the target domain data is inaccessible, or where there is very little usable data collected because most datasets are based on artificial damage and most fault diagnoses rely on vibration signals for diagnostic generalization.

[0038] This application utilizes a model trained on a dataset of artificially induced damage generated by artificial assistance to perform diagnostic tasks under real-world damage conditions. To clearly illustrate the problem to be solved, multiple source domains Ds are defined, among which... ;

[0039] In this definition, Indicates the number of source domains, the first Each source domain contains 1 labeled sample; and the source sample set , and These represent vibration signal data, current signal data, and mixed data containing both types in the source domain, respectively. The following section uses vibration signal data as an example. Let's take an example to explain.

[0040] in, , and , These represent the raw data in the vibration signal and the output of the machine status, respectively. It is the length of each sample. This indicates the number of machine conditions to be diagnosed.

[0041] The unseen target domain is unlabeled, that is... The source and target domains have the same class label space, meaning that class offset is not considered in this work. , Due to the domain offset, each domain is not exactly the same; there exists... .

[0042] The MDFNet model will be built within a deep learning framework. The domain generalization network consists of a feature extractor. It maps the original input signal to the feature space. and classification networks The goal of the MDFNet model is relative to the source data. go through Training, designing a good loss function to optimize network parameters This allows the model to potentially possess sufficiently strong generalization ability to competently perform diagnostic tasks. The diagnostic task. Although the process can easily capture discriminative features to diagnose source samples, it remains a useful task to make it generalize well to real damage in the domain of artificial damage, that is, to generalize well in the unseen domain; to make the best use of all collected data, such as vibration signals and motor current signals, while preventing the model from overfitting to artificial damage data.

[0043] Multi-source augmentation and hybrid architecture:

[0044] To address the issue of scarce data in real-world scenarios, this paper considers data augmentation in the computer vision (CV) domain, i.e., expanding existing finite datasets. Furthermore, non-smooth data distributions can make invariant representations difficult to learn, especially when there are significant differences between domains. Therefore, this application proposes a multi-source data augmentation method for bearing fault diagnosis to learn rich and general diagnostic knowledge with domain invariance and class discriminability.

[0045] In particular, by aggregating data from different domains and performing data mixing across these domains to create new instances—cross-domain samples—this approach allows for a deeper investigation of the correlation representations between different domains. First, this method expands the diversity of the data while preserving the uniqueness of each domain, thereby enhancing the generalization effect of feature learning. Second, this method produces a smoothing effect in terms of data distribution, helping the model to better learn domain-invariant feature representations. Furthermore, given the limitations of current diagnostic methods using traditional mixing approaches in multi-source domain data scenarios, a mixing strategy specifically designed for multi-source domain scenarios, such as... Figure 2 As shown.

[0046] If only sensors are used to collect vibration signals from the equipment under actual operating conditions, the data is input into the Vibration Signal Diagnosis (VSD) framework. The augmented samples generated using the MIXUP method for vibration data caused by artificial damage between the two domains are described as follows:

[0047] ,in , This represents the expectation operation. It was Dirac The (Dirac delta) function. Data From the same distribution , yes The enhanced tag is converted to one-hot encoded.

[0048] Similarly, if only the motor current signal of the equipment collected by the frequency converter is available under actual operating conditions, the data is input into the CSD (Current Signal Diagnosis) framework. The augmented sample of the motor current signal caused by artificial damage between the two domains is described as follows:

[0049]

[0050] This application aims to enable models trained using data generated by artificial damage to cope with damage under real working conditions. In order to maximize the acquisition of more comprehensive and accurate system status feedback, data fusion augmentation technology is used to integrate information from multiple sensors. This helps to improve the accuracy of fault diagnosis, reduce false alarms and false negatives, and improve the robustness of the system.

[0051] In MFAD (Mixed Fusion Augmented data Diagnosis), the training data comes from different distributions. To make the mixed distribution suitable, a novel fusion data mixing distribution is proposed, which can integrate different information from different source domains, as described below:

[0052] ,in, and It comes from arbitrary field The characteristics of the vibration signal and the motor current signal—target pair. Mixing from all domains yields a mixed characteristic—target pair. , These are the generated samples. That is the corresponding tag. It is a random variable used to control the degree of mixing. The degree of mixing is generated and normalized as follows: Among them random variables By introducing a hyperparameter To control the vector of random variables The generation of . and The relationship between expected values ​​is as follows With the increase of [the variable], the generated samples become more defined and less random. Specifically, the uncertainty is quantified by calculating the entropy of the sample category prediction, as shown in the following formula:

[0053] ,in It is a category label; yes Predicted labels.

[0054] The obtained data is then applied to fault diagnosis, and its superiority is judged by comparing the category classification loss and feature information loss, which will be described in detail later. Based on the above analysis, multi-source augmentation provides a foundation for mining more diverse domain-invariant representations for the network, thereby improving the model's generalization ability and robustness. The hybrid architecture strategy promotes the maximum utilization and processing of data from various source domains, enhancing the model's discriminability under various conditions.

[0055] MM method and objective optimization:

[0056] Figure 3 This paper provides an overview of the method and network architecture based on the MDFNet model proposed in this application. From an architectural perspective, the MDFNet model mainly consists of a feature extractor. Classifier The system consists of an MFAD module, a MM module (which maximizes the common information among features of the same class in different source domains while minimizing the dissimilar information among features in different source domains), and a weight update module. This application uses an artificially generated damage dataset for diagnostic research, considering that real-world damage data in actual industrial applications is not only scarce but also largely private, making it difficult to obtain large amounts of real-world damage data. Samples randomly selected from each source domain are fed into the MDFNet model. In the extractor... After feature extraction, the extracted features are sent to the classifier. Classifier extractor Mapping to corresponding category labels to learn shared discriminative information across different domains, MFAD then applies this to the extractor. The advanced features are subjected to domain confusion operations, making them invariants of different domains.

[0057] It should be noted that the MFAD and MM modules perform different obfuscation operations on the motor current signal, vibration signal, and mixed signal. This is to ensure that the parameters differ when different signals are augmented and mixed, and also to ensure that fault diagnosis can be performed on data under different real-world conditions. The entire network structure is shown in Table 1. Feature Extractor It consists of four convolutional layers and one LSTM layer. The extracted features are then passed through a final convolutional layer and flattened into a one-dimensional feature embedding in the last flat layer. The feature embedding is then used for forward computation and propagated towards three optimization objectives.

[0058] The feature information contrast loss considers feature alignment in the feature space, where only samples belonging to the same feature are grouped together. This maximizes the common information between features of the same class in different source domains, while minimizing the dissimilar information between features in different source domains. This enables the MFDNet model to learn domain-invariant feature representations across multiple source domains.

[0059] Table 1 Frame Details

[0060]

[0061] The first step is to calculate the intra-domain cross-entropy loss, i.e., the category classification loss, by using the known category labels in the training data.

[0062] Cross-entropy loss can be formalized as the following equation: In this formula, It comes from the domain The i-th sample, yes Category tags, It is a function that maps the original input to class predictions. Cross-entropy loss is expressed as: That is, in a given In the case of class, and All A dimensional vector, where and These represent the predicted and true class labels, respectively. It is particularly important to note that when... It is 1, if the sample If it comes from class c, it is 1; otherwise, it is 0. The predicted score is for class c.

[0063] The different working environments across different working domains may cause the same fault to exhibit different characteristics in each source domain. Therefore, the representation of the same feature information class may vary from domain to domain. This inconsistency in the representation of feature information classes across multiple source domains can impair generalization performance on new, unseen target domains in reality. Therefore, optimizing only the cross-entropy loss... Ignoring cross-domain relationships can negatively impact generalization performance. To address this issue and improve the model's robustness and discriminativeness, a novel feature information contrastive loss function is introduced to find domain-invariant class representations. Specifically, cross-domain relationships are considered by maximizing the mutual information between classes of the same feature information in different source domains, while minimizing the learned representations of changes in different feature information across the same domain. This is further applied to the generated augmented data. It can capture class information shared by new generated data and previous source data, and obtain environment-independent class representations, which can improve generalization performance.

[0064] like Figure 3 As shown, the MM module defines three samples, namely anchor samples ( Positive samples ( ) and negative samples ( ) And anchor samples and positive samples They are the same category, negative samples They belong to different categories. Anchor samples Feature information contrast loss is expressed as Its formal representation is as follows:

[0065] ;

[0066] here, It is the cardinality of the positive sample set, given a vector , The similarity scoring function is set as follows: , It is an information comparison hyperparameter. It is the last layer output. This represents the composition of the feature extractor and classifier. Specifically, through optimization... maximizing the anchor sample of The classifier predicts and the mutual information between all anchor samples and corresponding positive samples across different domains, where It is sampled from the positive sample set. Simultaneously, it minimizes the mutual information between the anchor sample and all negative samples. The overall feature information of all samples is compared with the loss. The calculation is as follows: By minimizing This can maximize the lower bound of mutual information between positive samples.

[0067] The overall optimization objective is represented by a convex combination of the following equations: ,in, These are the weights between the two losses. Adam is used as the optimizer to minimize the overall objective, and a feature extractor is learned. and prediction networks Within the framework, and Joint optimization is employed to improve generalization performance on new, unseen domains. Specifically, the cross-entropy loss aims to improve diagnostic performance within each source domain, while the feature-information contrastive loss... The goal is to improve shared feature information by using augmented feature information class representations across domains, which can improve generalization performance on unknown reality domains.

[0068] experiment

[0069] The following section introduces the Paderborn dataset, a publicly available dataset for rolling bearings, and a bearing dataset obtained through experiments using a bearing experimental platform built in the laboratory. Then, the results obtained using several state-of-the-art networks (such as V-Rex, CORAL networks, etc.) for fault diagnosis are presented.

[0070] Description of the Planetary Parallel Shaft Gearbox Bearing Dataset and the Paderborn Dataset:

[0071] The vast majority of damage in bearing datasets is man-made, such as... Figure 4A Grooves and such generated by electrical discharge machining Figure 4B The drilling of holes in the ring, as shown, demonstrates how these preparation methods produce the most commonly used artificial damage, characterized by high precision and ease of replication. However, such artificial damage still lacks the same identity as actual bearing damage. Figure 4A and Figure 4B It can be seen that the damage occurred very suddenly and sharply, which is very obvious in both the damaged and undamaged annular raceway areas. And as... Figure 4C The actual working conditions shown have an irregular structure and a less pronounced sharpness at the damaged and undamaged junctions.

[0072] To examine whether this difference affects actual fault diagnosis, data was processed using the same operating settings named N15_M07_10. This involved processing the spectral results of different damage conditions on the bearing outer ring, obtained from artificial and actual damage. Figure 5 As shown in the figure. These data were provided by the University of Paderborn and normalized to the range of 0 to 1 for intuitive comparison. It is evident that the distributions of artificial damage and real damage differ significantly, indicating potential differences in mean, variance, and failure frequency. Therefore, using artificial damage to explore real damage and improving model transfer efficiency is of great significance for transferring models to real-world operating conditions.

[0073] The Paderborn dataset was generated by the Kat Data Center at the University of Paderborn, with both vibration and motor current signals sampled at 64 kHz. Bearing damage was caused by both artificial and actual damage. The artificial and actual damage data in the Paderborn dataset were categorized into six domains (A to F) based on different operating conditions. Domains A, B, and C represent artificial damage, while domains L, M, and N represent actual damage. Table 2 shows the different damage specifications included in each operating condition. Each operating condition contains three classes: healthy, OR (Obstructive) and IR (Inverse) classes. To standardize the data format, a sliding window with a fixed length of 5120 and a shift size of 409 was used.

[0074] The planetary parallel shaft gearbox bearing dataset, sourced from the HD-CL-012X planetary parallel shaft gearbox bearing comprehensive fault simulation test bench, employs a dataset structure similar to the Paderborn dataset. Vibration signals are collected using accelerometers mounted on the planetary parallel shaft gearbox housing, with accelerometers covering the x, y, and z directions. The sampling frequency is 32 kHz. The experiment is also divided into healthy, outer ring (OR), and inner ring (IR) classes. Since most industrial faults primarily involve vibration signals, this experiment mainly explores the impact of vibration signals from different axes and the hybrid augmented vibration signals on fault diagnosis. The experiment demonstrates that vibration data obtained from different axes can distinguish between fault data of the same or different classes within the same network framework.

[0075] Table 2

[0076]

[0077] Comparison methods and experimental setup:

[0078] To verify the effectiveness of the proposed domain generalization method with data augmentation, we first compared and analyzed the experimental results before and after different types of data augmentation for the proposed network structure and several commonly used and state-of-the-art generalization methods. 1) Generalization via Risk Extrapolation (V-Rex): This reduces the model's sensitivity to extreme distribution changes by decreasing the risk differences between training domains. It is a more optimized algorithm than the most commonly used Empirical Risk Minimization (ERM) method, which minimizes the sum of empirical risks among samples from all domains. 2) Deep Correlation Alignment (Deep CORAL): This uses a linear transformation method to align the second-order statistical features of the source and target domain distributions. 3) Conditional Domain Adversarial Network (CDANN): This applies a separate domain classifier to each class to further improve alignment performance. 4) Adaptive Weighted Swing (SASW): This further enhances the discriminative power of feature representations by weakening the weights of unreliable samples. 5) Self-supervised contrastive regularization (SelfReg) for domain generalization (DG), which is cross-domain learning with contrastive regularization through random weighted averaging. 6) Deep domain confusion (DDC), a popular method to reduce the assignment gap, with the overall network optimization objective being the cumulative MMD loss between any two source domains.

[0079] Implementation Details: During the training phase, artificial damage data from multiple source domains in the dataset are used for training, while real damage data is used for validation. Prioritizing detection ensures good convergence. To test the effectiveness of data augmentation on different data types, augmentation is performed on different datasets, and the method's performance on single-type and mixed-type data is verified using different algorithms. Generalization performance is then validated through classification of real damage data. It is worth emphasizing that the standard DG protocol is followed throughout the generalization to the real damage domain; diagnostic performance is measured by an accuracy score, which is the percentage of correctly classified samples divided by the total number of samples.

[0080] The signal is processed by a wide kernel (64x1) in the first convolutional layer to capture fault features in the low-frequency band and suppress high-frequency noise. As the depth feature map length decreases, subsequent convolutional layers use smaller kernels to improve feature representation. The number of kernels learning multidimensional abstract feature representations increases with layer depth. LSTM layers are then used to further process the correlations between signal features. Dropout layers and batch normalization layers are appended to each convolutional and fully connected layer to avoid overfitting and accelerate training. The learning rate is fixed at 0.001. The remaining parameter settings are as follows: For model training, weight decay is fixed at 5e−5 and batch size is fixed at 64 throughout the experiment. For other comparative methods, their hyperparameters are based on the original reference and gradually tuned to optimal values.

[0081] Results Analysis

[0082] Table 3 shows the results of different methods on the Paderborn dataset. Specifically, column C1-Signal in Table 3 means using the motor current signal alone as the feature signal, while C2-Signal (SE) is a retest after applying the data augmentation method. The last column represents the average performance of various methods in the real damage target domain after data augmentation using a mixture of vibration and motor current signals. It should be noted that this embodiment uses a model trained on an artificial damage dataset in an unknown domain representing real-world damage. In Table 3, the best results obtained for different types of data are highlighted in bold, while the second best value is highlighted underlined. It can be seen that the data augmentation method proposed in this embodiment improves the accuracy for both motor current and vibration signals, with vibration signals generally showing higher diagnostic accuracy than motor current signals. Compared to other methods, SelfReg and DDC show better average accuracy on the dataset. Both the weighted averaging of each domain used in SelfReg and the statistical moment matching method used in DDC aim to eliminate distribution differences between different domains and learn invariant representations of feature information between domains.

[0083] Table 3

[0084]

[0085] Figure 7 The experimental results also demonstrate the effectiveness of the proposed strategy. Overall, the proposed MDFNet-based method significantly outperforms all other methods. On average, the MDFNet-based method outperforms the second-best baseline by 4.54%. This indicates that the proposed method is more stable than other comparable methods in generalization tasks across different types of data. Furthermore, experimental results using vibration data from planetary parallel shaft gearbox bearings show that... Figure 6As shown, mixing vibration signals from different axes in different network frameworks yields better classification results than using vibration signals from a single axis, further confirming the reliability of the proposed method.

[0086] Visual analysis:

[0087] To visually demonstrate the effectiveness of the network proposed in the embodiments of this application on real-world damage, the classification effect of the above-described method can be visualized. Specifically, t-SNE technology can be used to map high-dimensional features to low dimensions. The resulting augmented data is used to train all contrastive network models and tested on unknown real-world damage. The test results are then compared with the method proposed in this application. Figure 8 The visualization results show the characteristics of these methods, with three colors representing healthy, outer circle fault, and inner circle fault, respectively. Green indicates healthy, red indicates outer circle fault, and blue indicates inner circle fault. CDANN shows weak separation between the outer and inner circles, with severe confusion in some cases. This indicates that each individual domain classifier trained on the artificial damage dataset cannot generalize well to real damages and is easily misdiagnosed. SASW's visualization results show multiple small clusters of features of the same category, showing better inter-category separation than CDANN, but weak intra-category clustering. This makes samples at the decision boundary highly susceptible to misclassification. In contrast, other methods show improved accuracy, especially the proposed method, which demonstrates good feature matching through the method of maximizing and minimizing feature information. This proves that the proposed method can successfully extract feature information from the artificial loss dataset and generalize the learned invariant feature representation to real damages.

[0088] FICL ablation:

[0089] This application primarily benefits from the designed data augmentation and feature-contrast loss, which improve the invariance and class discriminability of artificial damage feature information, maximizing generalization performance in the unseen real loss domain. To verify their effectiveness on the MDFNet model, the advantages of multi-source domain data augmentation over single-domain data augmentation were verified while maintaining network structure and parameters invariance. Furthermore, the impact of feature-contrast loss on classification accuracy was explored through a series of experiments, and the results are compared as follows: Figure 7 As shown in the figure. The results comparison shows that multi-source domain data hybrid augmentation is significantly better than single-domain data augmentation, and the method also has an improvement effect on other network structures, with a maximum improvement of 7.32%.

[0090] exist Figure 9As can be clearly seen, the test accuracy of the MDFNet model drops sharply by 10.61% when the Feature Information Contrast Loss (FICL) is omitted. In contrast, the FICL method, in the unknown domain of the true loss, learns fault feature information better and generalizes the results more significantly than the multi-source data augmentation strategy. Furthermore, Figure 9 In the diagram, black and red represent losses with and without FICL, respectively. The comparison reveals that the loss with FICL is smoother, indicating better generalization and consistent results under different conditions and environments. The comparison demonstrates the effectiveness of this feature-based contrastive learning loss function for generalization. Therefore, the superior generalization performance of the MDFNet model can be attributed to the implementation of FICL and multi-source data augmentation.

[0091] Parameter analysis:

[0092] The hyperparameters involved in the data augmentation section include the symmetric distribution hyperparameter α (with a Beta symmetric distribution) and the mixing intensity hyperparameter β (affecting the mixing intensity); during classification, there is the information contrast hyperparameter τ (affecting information contrast). Considering the impact of these hyperparameters on model performance, the experiment analyzed the sensitivity of each hyperparameter. For example... Figure 10 As shown, when the symmetric distribution hyperparameter α approaches 0, and when λ is centered at 0 or 1, the trend for model accuracy is initially an increase followed by a decrease. For the mixture intensity hyperparameter β, a larger value of β indicates more defined generated samples and less randomness. The trend for model accuracy is also an initial increase followed by a decrease. It is speculated that the initial increase may be due to enhanced robustness, while the subsequent decrease in performance may be due to insufficient data.

[0093] Since there is more than one hyperparameter, when analyzing one of the hyperparameters, the others are fixed at their optimal values. The results are as follows: Figure 10As shown, when the symmetric distribution hyperparameter α approaches 0.01, the model's diagnostic accuracy decreases, indicating that α should be selected within the range of [0.2, 0.6]. Conversely, when the information contrast hyperparameter τ approaches 0 or exceeds 0.2, the model's diagnostic accuracy decreases, indicating that τ should be selected within the range of [0.01, 0.2]. When the symmetric distribution hyperparameter α is greater than 0.2, its impact on model performance fluctuates less, indicating that it is not a major influencing factor. As the mixing intensity hyperparameter β increases from 0 to 10, the test accuracy shows a trend of first increasing and then decreasing. This may be because smaller β values ​​can cause some unreliable samples to damage model performance, while larger β values ​​lead to more explicit and less random generated augmented samples, reducing generalization ability and thus limiting the model's ability to learn general diagnostic knowledge. Therefore, under the current experimental settings, selecting a value of β within the interval [7, 9] is appropriate. When β is changed, the model's accuracy fluctuates accordingly.

[0094] Further experimental studies investigated the varying impacts of different mixing intensity hyperparameters β on the results for different data frame choices. For mixed data, a β value of 8 resulted in higher accuracy, while a β value of 6 was more suitable for simple vibration signal data. Relatively speaking, the mixing intensity hyperparameter β had a smaller impact on motor current signals. Figure 11 As shown.

[0095] This application proposes a novel domain generalization method based on the MDFNe model to address real-world fault diagnosis tasks, particularly in scenarios with limited data and unknown operating conditions. More importantly, it aims to demonstrate fault diagnosis in such situations. This method learns generalized representations of feature information from multiple source domains and uses these representations to diagnose the target domain in practical contexts.

[0096] Specifically, the main features of the novel domain generalization method proposed in this application are the use of a multi-source data hybrid enhancement generalization module and a feature information contrastive learning module. These not only enhance the robustness and generalization ability of the general representation of feature information, but also utilize maximization and minimization to make the diagnostic boundary more compact and generalized, thereby reducing the distribution shift caused by fluctuating operating conditions. Thanks to these two features, the MDFNet model proposed in this application fully utilizes existing data mining techniques to extract common feature information from different types of data and promptly detect faults.

[0097] Extensive data analysis conducted on the rolling bearing dataset collected by the applicant's self-built platform and the German Paderborn rolling bearing dataset demonstrates that this method significantly improves the generalization performance of real data in unknown domains, outperforming existing methods.

[0098] Other embodiments of the present invention also provide an apparatus including a memory, a processor, and a program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the rotating machinery real damage diagnosis method based on a hybrid multi-source domain generalized network model as described in the above embodiments of the present invention.

[0099] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of hardware embodiments, software embodiments, or embodiments combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage and optical storage) containing computer-usable program code.

[0100] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions specified in one or more blocks of the flowchart illustrations and / or one or more blocks of the block diagrams.

[0101] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means that implement the functions specified in one or more flowcharts and / or one or more block diagrams.

[0102] These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process, such that the instructions, which execute on the computer or other programmable apparatus, provide steps for implementing the functions specified in one or more flowcharts and / or one or more block diagrams.

[0103] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the methods described above. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc.

[0104] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the technical scope disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for diagnosing real damage to rotating machinery based on a hybrid multi-source domain generalized network model: characterized in that, Including the following steps: Provides a hybrid multi-source domain generalized network model; The hybrid multi-source domain generalization network model is configured to aggregate data from different domains based on a preset hybrid strategy suitable for multi-source domain scenarios, and to mix data between different domains to create new instances. The fused data hybrid distribution, which integrates different information from multiple different source domains in the new instance, is described as follows: ； in, and It comes from arbitrary field The vibration signal and the motor current signal are characteristic-target pairs; Ds is the defined source domain, where, , Indicates the number of source domains, the first Each source domain contains 1 labeled sample; and the source sample set , and These represent vibration signal data, current signal data, and mixed data containing both types from the source domain, respectively; the mixture from all domains yields a mixed feature-target pair. , These are the generated samples. These are the corresponding tags; It is a random variable used to control the degree of mixing. The degree of mixing is generated and normalized as follows: Among them random variables Among them, the mixing intensity hyperparameter To control the random variable The generation of the mixing intensity hyperparameter and The relationship between the expected values ​​and the mixing intensity hyperparameter is as follows: With the increase of [something], the randomness of the generated samples is reduced; and The obtained data is applied to fault diagnosis, and the superiority of the loss function is judged by comparing the category classification loss function and the feature information.

2. The method for diagnosing real damage to rotating machinery based on a hybrid multi-source domain generalization network model according to claim 1, characterized in that, The mixing intensity hyperparameter β is selected within the interval [7, 9].

3. The method for diagnosing real damage to rotating machinery based on a hybrid multi-source domain generalized network model according to claim 1, characterized in that, It also includes designing the feature information contrast loss function to maximize the mutual information between the same feature information class in different source domains to consider cross-source domain relationships, while minimizing the learning representation of different feature information changes between the same source domain.

4. The method for diagnosing real damage to rotating machinery based on a hybrid multi-source domain generalization network model according to claim 1, characterized in that, It also includes enabling the learning of artificial damage data feature information from different source domains and then generalizing it to the unknown domain of real damage. That is, by using feature discrimination, the inner product of similar feature representations is averaged higher than other features, thereby learning the same fault feature information from different domains in artificial damage data and then generalizing it to data under real damage.

5. The method for diagnosing real damage to rotating machinery based on a hybrid multi-source domain generalization network model according to claim 1, characterized in that, Define anchor sample Positive samples and negative samples Three samples, and anchor sample and positive samples They are the same category, negative samples Belonging to different categories; anchor samples The feature information comparison loss function is expressed as: The formal representation is as follows: ； in, It is the cardinality of the positive sample set, given a vector , The similarity scoring function is set as follows: , It is an information comparison hyperparameter; It is the output of the last layer; This represents the composition of the feature extractor and classifier; the overall feature information of all samples is compared with the loss function. The calculation is as follows: .

6. The method for diagnosing real damage to rotating machinery based on a hybrid multi-source domain generalization network model according to claim 5, characterized in that, The information comparison hyperparameter τ is selected from [0.01, 0.2].

7. The method for diagnosing real damage to rotating machinery based on a hybrid multi-source domain generalization network model according to claim 5, characterized in that, The category classification loss function is the cross-entropy loss. function.

8. The method for diagnosing real damage to rotating machinery based on a hybrid multi-source domain generalization network model according to claim 7, characterized in that, The overall optimization objective consists of the feature information contrast loss and the cross-entropy loss. Convex combination representation of functions: , where α is a symmetric distribution hyperparameter.

9. The method for diagnosing real damage to rotating machinery based on a hybrid multi-source domain generalization network model according to claim 8, characterized in that, The symmetric distribution hyperparameter α is selected from [0.2, 0.6].

10. The method for diagnosing real damage to rotating machinery based on a hybrid multi-source domain generalization network model according to claim 1, characterized in that, The rotating machinery mentioned is a bearing.

Images