Multi-source domain open set rotating machinery fault diagnosis method based on category transfer

Abstract

The present application relates to the field of fault diagnosis, in particular to a kind of multi-source domain open set rotating machinery fault diagnosis method based on category transfer, the present application combines the multi-source domain deep migration network in unsupervised field and adversarial learning, adopt multi-source fault category matching to let the fault feature alignment between the same fault type in different source domains and the fault feature distant between different fault types;Adversarial learning method is used between each source domain and target domain to reduce the domain gap between source domain and target domain, effectively alleviate the incomplete and suboptimal solution of fault diagnosis result caused by simply fusing multi-source domain fault information in the existing multi-source domain open set fault diagnosis method;In addition, based on the comprehensive decision result of multi-fault classifier, construct global fault alignment between the fault centroid representation of multi-source domain and target domain to further optimize the field invariant fault feature extraction capability of shared feature extractor in fault diagnosis model, further improve fault diagnosis precision.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of fault diagnosis, and particularly relates to a multi-source domain open set rotating machinery fault diagnosis method based on class transfer. BACKGROUND

[0002] In order to ensure the stability and safety of rotating machinery in industrial applications, fault diagnosis is a necessary condition for preventing accidents and ensuring the efficient and reliable operation of rotating machinery. In recent years, data-driven fault diagnosis methods have been deeply developed, and in particular, deep learning-based fault diagnosis methods have increasingly attracted the attention of researchers due to their excellent automatic feature extraction capabilities. However, existing deep learning-based fault diagnosis methods follow the assumption that training data and test data share the same data distribution. Such an assumption is not easy to meet in real industrial scenarios, because the operating conditions of rotating machinery are constantly changing to a large extent, resulting in different fault data distributions collected under different operating conditions. Therefore, the fault diagnosis performance of deep learning-based fault diagnosis methods will be sharply reduced.

[0003] Currently, unsupervised transfer learning can extract domain-invariant features between domains, effectively alleviating the problem of model performance decline caused by distribution differences between domains, providing a new perspective for the above-mentioned problem in fault diagnosis. However, existing unsupervised transfer methods for fault diagnosis mainly focus on utilizing and transferring the fault knowledge of a single source domain to the target domain for fault diagnosis. In fact, we can usually obtain multiple source domains with different data distributions. Researchers have long proven that more comprehensive and universal fault knowledge can be learned from multiple source domains. For example, Figure 1 As shown in (a), existing multi-source domain fault diagnosis methods usually assume that the label space of multiple source domains and the target domain is consistent and the fault types between source domains are the same, that is, multi-source domain closed set fault diagnosis. However, the method based on multi-source domain closed set fault diagnosis does not consider the case that the fault types between different source domains may not be consistent, and the existing model can only identify known fault types contained in multiple source domains, but cannot identify potential unknown fault types in the target domain that are not observed in multiple source domains. Rotating machinery running in harsh environments will inevitably encounter potential unknown faults during use. In addition, it is quite difficult to collect and label source domain fault samples from various operating conditions, often leading to limited fault classes within each source domain. That is, there can be shared fault types and own unique fault types within each source domain, that is, the class transfer problem. The existence of the class transfer problem between multiple source domains will further hinder the transfer of knowledge to the target domain, thereby reducing the fault diagnosis performance. As shown in (b), this new scenario is defined as a multi-source domain open set fault diagnosis with class transfer, where the label spaces of multiple source domains and the target domain are inconsistent. Figure 1 (b), which is defined as a multi-source domain open set fault diagnosis with class transfer, where the label spaces of multiple source domains and the target domain are inconsistent.

[0004] In summary, the current multi-source domain open set rotating machinery fault diagnosis method still has the problems of class transfer between multiple source domains and the inability to separate potential unknown fault types in the target domain. SUMMARY

[0005] In view of the problems of class transfer between multiple source domains and the inability to separate potential unknown fault types in the target domain in the existing multi-source domain open set rotating machinery fault diagnosis method, the present application proposes a multi-source domain open set rotating machinery fault diagnosis method based on class transfer. In the proposed multi-source adversarial deep migration network, the fault features of the fault samples in the source domain training set and the fault samples in the target domain training set are extracted by a shared feature extractor. The centroid representation of the fault classes in each source domain is constructed by the extracted fault features of the multiple source domain training fault samples, and the distances between the same fault types and the distances between different fault types in different source domains are calculated by the Euclidean distance, and a multi-source fault type matching loss is further constructed to make the same fault types close to each other and the different fault types distant from each other. At the same time, the fault features are further input into the domain unique fault classifier between each source domain and the target domain for fault classification, a fault multi-classification loss is constructed, and the network parameters of the shared feature extractor and the domain unique fault classifier are optimized. Then, the fault samples in the target domain training set are input into the domain unique classifier after feature extraction to obtain the target domain fault sample prediction, and an improved binary classification cross-entropy loss with additional constraints is constructed for the fault prediction to reduce the domain gap between each source domain and the target domain and make the domain unique classifier have the ability to distinguish known fault types and potential unknown fault types. Subsequently, a multi-fault classifier comprehensive decision module is constructed by using multiple domain unique fault classifiers to predict the target domain fault samples, the information entropy of the prediction vector of each domain unique fault classifier for the target domain fault samples is calculated, the entropy value is weighted to each domain unique fault classifier as a weight, and a more comprehensive and universal target domain prediction is obtained. Finally, the comprehensive prediction result of the target domain fault samples is used to obtain the pseudo fault type label of the target domain fault samples. The centroid representation of the target domain fault classes is constructed according to the pseudo fault type label, the distances between the centroid of all source domain fault classes after averaging and the centroid of the fault classes in the target domain are calculated by the Euclidean distance, and a global alignment loss is constructed. The ability of the shared feature extractor to extract domain-invariant fault features is optimized by the loss.

[0006] To achieve the above purpose, the present application provides a multi-source domain open set rotating machinery fault diagnosis method based on class transfer, which comprises:

[0007] The acceleration sensor is used to collect the vibration data of the key components of the rotating machinery under multiple different operating conditions, and the vibration data is preprocessed.

[0008] According to whether the vibration data has a fault type label, fault type label data under multiple operating conditions are taken as multiple source domain training sets, and fault type label data under a certain condition are taken as a target domain training set and a target domain test set;

[0009] The multiple source domain training sets and the target domain training set are packaged into batches respectively, and are input into the multi-source adversarial deep migration network in batches;

[0010] The shared feature extractor of the multi-source adversarial deep migration network is used to extract high-dimensional feature representations of fault samples of the multiple source domain training sets and the target domain training set respectively;

[0011] The high-dimensional feature representations of the fault samples are processed by the fault classifier unique to each field of the multi-source adversarial deep migration network, and the fault types of the fault samples are predicted respectively;

[0012] According to the fault type labels of the fault samples in each source domain, fault type centroid representations in each source domain are constructed, and centroid distances of the same fault type labels and centroid distances of different fault type labels between different source domain training sets are calculated, and a multi-source fault type matching loss is constructed;

[0013] According to the fault type prediction of the fault samples output by the fault classifier unique to each source domain of the different source domain training sets and the real fault type label, a fault multi-classification loss is calculated;

[0014] According to the fault type prediction of the fault samples output by the fault classifier unique to each source domain of the target domain training set, a multi-classifier adversarial classification loss is calculated;

[0015] According to the fault type prediction results of the fault samples in the target domain training set by the fault classifier unique to each source domain, the prediction information entropy is calculated, and the confidence weight of each prediction is constructed, the pseudo fault type label of the fault samples in the target domain training set is generated, the fault centroid in the target domain is constructed, and the global fault alignment is constructed between the average centroid of the multiple source domains, and the global fault alignment loss is calculated;

[0016] The multi-source adversarial deep migration network is iteratively trained by optimizing the matching loss, the fault multi-classification loss, the adversarial classification loss and the global fault alignment loss;

[0017] The target domain test set is input into the trained multi-source adversarial deep migration network, and the fault diagnosis result of the target domain test set is output.

[0018] The beneficial effects of the present application are:

[0019] The present application is aimed at the problem that the category transfer exists among multiple source domains, which hinders the fault knowledge transfer and causes the low fault diagnosis precision due to the failure to separate the potential unknown fault types in the target domain, and proposes a multi-source domain open set rotating machinery fault diagnosis method based on category transfer. The constructed multi-source adversarial deep migration network includes a multi-source domain fault type matching module, a multi-fault classifier comprehensive decision module and a global alignment module, which respectively realize the approximation of the same fault class centroid representation in the multi-source domain and the distancing of different fault class centroids, obtain comprehensive and universal target domain fault sample prediction, and further optimize the domain-invariant fault feature extraction capability of the shared feature extractor, thereby improving the low rotating machinery cross multi-source domain open set fault diagnosis precision under the influence of category transfer. The experimental results of the embodiments of the present application show that:

[0020] (1) The method uses multi-source domain fault type matching to make the same fault types in different source domains close to each other and different fault types distant from each other, so as to solve the category transfer problem among multiple source domains.

[0021] (2) The method narrows the domain gap between each source domain and the target domain through an adversarial learning method, and obtains more comprehensive and universal target domain training sample fault prediction through the construction of a multi-fault classifier comprehensive decision module.

[0022] (3) The method obtains the pseudo fault label corresponding to the fault sample by using the target domain training sample fault prediction, and constructs the centroid representation of the fault category in the target domain. The Euclidean distance between the averaged multi-source domain fault category centroid representation and the target domain fault category centroid representation is calculated to construct a global alignment loss for optimizing the domain-invariant fault feature extraction capability of the shared feature extractor.

[0023] (4) Aiming at the problems of category transfer existing among multiple source domains and the potential unknown faults in the target domain being unable to be identified, the present application proposes a multi-source domain open set rotating machinery fault diagnosis method based on category transfer, which effectively improves the rotating machinery cross multi-source domain open set fault diagnosis precision. BRIEF DESCRIPTION OF DRAWINGS

[0024] In order to more clearly express the purpose, technical scheme and advantages of the present application, the present application will be further described in detail below in combination with the drawings, in which:

[0025] Figure 1 A comparison diagram of multi-source domain closed set fault diagnosis and category transfer multi-source domain open set fault diagnosis;

[0026] Figure 2 A flowchart of a multi-source domain open set rotating machinery fault diagnosis method based on category transfer according to an embodiment of the present application;

[0027] Figure 3 This is a flowchart of a preferred embodiment of the rotating machinery fault diagnosis method based on category transfer using multi-source domain open sets.

[0028] Figure 4 This is a schematic diagram of the structure of a multi-source adversarial deep migration network;

[0029] Figure 5 A bar chart comparing the fault classification accuracy and H-score of different methods on a bearing fault dataset;

[0030] Figure 6 A bar chart comparing the fault classification accuracy and H-score of different methods on a gearbox fault dataset;

[0031] Figure 7 A comparison of the confusion matrices between MS-OSTDN and the Proposed method in bearing fault diagnosis task A2;

[0032] Figure 8 A comparison of the confusion matrix between MS-OSTDN and the Proposed method in gearbox fault diagnosis task B1;

[0033] Figure 9 For different hyperparameters t c A bar chart of fault classification accuracy and H-score for gearbox fault diagnosis task B2;

[0034] Figure 10 Radar plots showing the fault classification accuracy and H-score accuracy of ablation experiments on the bearing fault dataset;

[0035] Figure 11 Radar plot of fault classification accuracy and H-score accuracy in ablation experiments on the gearbox fault dataset. Detailed Implementation

[0036] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0037] Due to class transfers between multiple source domains and the presence of potential unknown faults in the target domain, existing multi-source domain closed-set transfer networks transmit incomplete and suboptimal fault information to the target domain, resulting in low accuracy in fault diagnosis of multi-source domain open-set rotating machinery. Therefore, existing multi-source domain closed-set transfer models are not applicable.

[0038] In the embodiment of the present application, a multi-source domain open set rotating machinery fault diagnosis method based on category transfer is provided, and the structure is as shown in Figure 2 The method mainly comprises:

[0039] S1, collecting vibration data of key parts of rotating machinery under multiple different operating conditions by using an acceleration sensor, and performing data preprocessing on the vibration data;

[0040] S2, according to whether the vibration data has a fault type label, taking the fault type labeled data under multiple operating conditions as multiple source domain training sets, and taking the fault type unlabeled data under a certain operating condition as a target domain training set and a target domain test set;

[0041] S3, packing the multiple source domain training sets and the target domain training set into batches respectively, and inputting them into a multi-source adversarial deep migration network in batches;

[0042] S4, extracting high-dimensional feature representations of fault samples of the multiple source domain training sets and the target domain training set by using a shared feature extractor of the multi-source adversarial deep migration network;

[0043] S5, processing the high-dimensional feature representations of the fault samples by using a fault classifier unique to each domain of the multi-source adversarial deep migration network, and predicting the fault types of the fault samples respectively;

[0044] S6, constructing a fault type centroid representation in each source domain according to the fault type labels of the fault samples in each source domain, calculating the centroid distance of the same fault type labels and the centroid distance of different fault type labels between different source domain training sets, and constructing a multi-source fault type matching loss;

[0045] S7, calculating a fault multi-classification loss according to the fault type prediction of the fault samples output by the fault classifier unique to each source domain of the different source domain training sets and the real fault type label;

[0046] S8, calculating a multi-classifier adversarial classification loss according to the fault type prediction of the fault samples output by the fault classifier unique to each source domain of the target domain training set;

[0047] S9, calculating the prediction entropy and constructing the confidence weight of each prediction according to the fault type prediction results of the fault samples in the target domain training set by the fault classifier unique to each source domain, generating a pseudo fault type label of the fault samples in the target domain training set, constructing a fault class centroid in the target domain, aligning the global fault with the average centroid between the multi-source domains, and calculating a global fault alignment loss;

[0048] S10, iteratively training the multi-source adversarial deep migration network by optimizing the matching loss, the fault multi-classification loss, the adversarial classification loss and the global fault alignment loss.

[0049] S11, inputting the target domain test set into the trained multi-source adversarial deep migration network to output a fault diagnosis result of the target domain test set.

[0050] In the preferred embodiment of the present application, a multi-source domain open set rotating machinery fault diagnosis method based on class transfer is provided, as shown in the figure. Figure 3 The method mainly includes two processes of multi-source adversarial deep migration network training and fault diagnosis.

[0051] In the training of the multi-source adversarial deep migration network, the fault vibration data under multiple different operating conditions are preprocessed and divided into multiple source domain training sets with fault labels, target domain training sets without fault labels and target domain test sets; the source domain training sets and the target domain training sets are input into the network for network training. The multi-source adversarial deep migration network realizes that the same fault types in different source domains are close to each other and different fault types are far away from each other through the construction of a multi-source fault type matching module, solves the problem of class transfer, reduces the domain gap between each source domain and the target domain in an adversarial learning manner through the construction of an improved binary classification loss with additional constraints, and enables the domain-specific classifier to have the ability to distinguish between known fault types and unknown fault types; the multi-source adversarial deep migration network outputs a more comprehensive and universal prediction of the target domain fault samples through the construction of a multi-fault classifier decision module; the global alignment loss between the centroid representation of the fault classes in the multi-source domain after averaging and the centroid representation of the fault classes in the target domain is constructed through the comprehensive prediction of the target domain fault samples to obtain the pseudo-fault labels corresponding to the fault samples, further optimize the ability of the shared feature extractor to extract domain-invariant fault features, realize the transmission of fault information, and finally enable the network to identify known fault types and separate potential unknown fault types on the target domain.

[0052] In the rotating machinery fault diagnosis, the fault samples in the target domain test set to be tested need to be preprocessed in the same way; the preprocessed data are input into the multi-source adversarial deep migration network with saved model parameters; and the fault diagnosis prediction result of the fault vibration signal to be tested is output.

[0053] In the embodiment of the present application, the fault vibration data of rotating machinery key components such as bearings and gearboxes under multiple different operating conditions (different rotating speeds, different loads, etc.) can be collected; the preprocessing of the fault vibration data includes converting the acquired time-domain fault vibration signal into a frequency-domain signal through fast Fourier transform.

[0054] In the embodiment of the present application, the fault diagnosis can be realized by constructing a multi-source adversarial deep migration network; as shown in the figure. Figure 4As shown, the multi-source adversarial deep migration network at least includes a shared fault feature extractor and a plurality of domain unique fault classifiers; the source domain and the target domain use the same shared fault feature extractor to extract high-dimensional features; the domain unique fault classifier is used to realize the fault classification of the fault samples in each domain.

[0055] In the embodiment of the application, for the fault samples in the plurality of source domain training sets, the fault class centroid representation in each source domain is constructed, the centroid distance of the same fault type and the centroid distance between different fault types between different source domains are calculated by the Euclidean distance, and a multi-source fault class matching loss function is further constructed to make the centroid distance of the same fault type close to each other and the centroid distance between different fault types far from each other; therefore, the multi-source fault type matching loss can be expressed as:

[0056]

[0057] wherein, represents the multi-source fault type matching loss; D shared represents the centroid distance of the same fault type label between different source domain training sets; D private represents the centroid distance between different fault type labels between different source domain training sets.

[0058] The centroid representation of the fault type of the source domain is:

[0059]

[0060] wherein, represents the centroid of the nth fault type in the source domain S m , m∈{1, 2,..., M}, represents the fault sample of the source domain corresponding fault sample label; is the corresponding source domain S m batch, I (·) is an exponential function, is the corresponding fault sample label of the bth source domain fault sample , is extracted by the shared feature extractor G fe (·).

[0061] The smaller D shared indicates that the feature representations of the shared fault classes between the source domains are close to each other, which can make the same fault knowledge more rich and concentrated. On the contrary, the larger D private indicates that the private fault classes are far away from each other, which can make the different fault classes more easily distinguished. Through the multi-source fault type matching loss function​ It can solve the problem of suboptimal fault information transmission to the target domain caused by class transfer issues between multiple source domains, which reduces the fault diagnosis performance of the target domain.

[0062] In this embodiment of the invention, considering that each source domain-specific fault classifier can output fault predictions for fault samples in the training set of that source domain, a fault multi-classification loss can be constructed to optimize shared feature extraction and domain-specific fault classifiers; the fault multi-classification loss is expressed as:

[0063]

[0064]

[0065] in, It is the source domain S m Multi-classification of fault losses It is the source domain S m The number of fault samples in the data. It is a multi-class cross-entropy loss function. It is the source domain S m The i-th fault sample Predicted fault labels, It is the source domain S m The i-th fault sample The corresponding fault sample label, It is a fault classifier unique to the m-th domain.

[0066] In this embodiment of the invention, based on the adversarial learning method, an improved binary classification loss with additional constraints is constructed to reduce the domain gap between each source domain and target domain, and to enable the domain-specific classifier to distinguish between known fault types and unknown fault types.

[0067] Fault samples in the target domain training set are used to construct an improved binary classification cross-entropy loss with additional constraints. This reduces the domain gap between the source and target domains, while enabling the fault classifier to distinguish between known and unknown fault types. Specifically, a binary classification cross-entropy loss function driven by a fixed threshold T is used. Adversarial learning is used to reduce the domain gap between each source and target domain and enable the domain-specific fault classifier to identify unknown faults. Meanwhile, a constraint loss is constructed to make the prediction results more reliable. The method is used for restraining the field unique fault classifier, so that the classification boundary between each fault category is more clear. In addition, the fault samples in the target domain training set are output by multiple field unique fault classifiers, and multiple prediction results are obtained. A multi-fault classifier comprehensive decision module is constructed to comprehensively predict the target domain fault samples. Specifically, the information entropy of the prediction vector of the target domain fault sample is calculated as a weight to weight the prediction of the fault sample by each field unique fault classifier, and a comprehensive prediction is obtained.

[0068] It should be noted that the information entropy can reflect the stability and reliability of fault prediction. The greater the information entropy value, the more unstable and unreliable the fault prediction is. Conversely, the smaller the information entropy value, the more stable and reliable the fault prediction is. However, the prediction dimension output by each field unique fault classifier is not necessarily consistent. The prediction output dimension of the field unique fault classifier is dimension. Therefore, in order to ensure the consistency of the dimension and give the final weighted decision result, it is necessary to expand the prediction output dimension of each field unique classifier to (|K|+1) dimension. Specifically, the corresponding fault category in the field unique fault classifier does not exist, and the dimension expansion is filled with zero value. It should be noted that, is the number of fault categories in each source domain, and |K| is the number of fault categories in all source domains. The related calculation formula of the multi-classifier adversarial classification loss is as follows:

[0069]

[0070]

[0071]

[0072] Among them, represents the multi-classifier adversarial classification loss of the mth field; represents the binary cross-entropy loss of the m fields, represents the constraint loss of the mth source domain; is the target domain fault sample training set, T is a fixed threshold used as a boundary threshold for distinguishing known fault types and potential unknown fault types in the target domain, is the target domain fault sample the first dimension probability value of the fault prediction output by the shared feature extractor and the field unique fault classifier, is the target domain fault sample the fault sample feature extracted by the shared feature extractor G fe (·), represents the mth field unique fault classifier; t cis a predefined threshold and is used to constrain the classification boundary between failure classes, is the number of failure classes in each source domain.

[0073] In the embodiments of the present application, the failure samples in the target domain training set are input into the unique failure classifier of each domain, a multi-failure classifier comprehensive decision module is constructed to make comprehensive and universal prediction for the failure samples; finally, the comprehensive prediction results of the target domain failure samples are used to generate the pseudo-failure labels of the failure samples, the failure type centroid representation in the target domain is constructed, and the distance between the centroid representation and the average centroid representation in multiple source domains is calculated by using the Euclidean distance, a global failure alignment loss is constructed to further optimize the neighborhood invariant failure feature extraction capability of the shared feature extractor and improve the failure diagnosis precision; wherein the global alignment loss is represented as:

[0074]

[0075] wherein, is the global alignment loss; |K| represents the number of failure classes in all source domains; is the centroid representation of the lth failure class in the target domain training set, is the failure class centroid of all source domain failure classes.

[0076] The centroid representation of the failure type of the target domain is:

[0077]

[0078]

[0079]

[0080]

[0081]

[0082]

[0083] wherein, is the pseudo-label of the target domain failure sample, represents the comprehensive prediction of the target domain failure sample; is the failure prediction weight of the corresponding domain unique failure classifier, is the prediction vector of the target domain failure sample and its length is H(·) is an information entropy function, is the failure sample feature of the target domain failure sample extracted by the shared feature extractor G fe (·), represents the mth domain-specific fault classifier.

[0084] To verify the effectiveness of the category transfer multi-source domain open set fault diagnosis method disclosed in the present application, bearing fault vibration data collected on a machine fault simulator (MFS) and gearbox fault vibration data collected on a drivetrain diagnostics simulator (DDS) were used for experimental verification. The MFS mainly consists of a 1-horsepower 3-phase motor, a 1-horsepower variable frequency AC driver with a multifunction programmable control panel, an internal tachometer (Model: PCB 352C03) with LCD display, a pulse analog TTL output for data acquisition, a shaft, two rotors, two bearing supports and a gearbox. Four types of bearing faults were simulated at three different constant speeds (22 Hz, 35 Hz and 45 Hz), as shown in Table 1.

[0085] Table 1 Bearing fault type description

[0086]

[0087]

[0088] The DDS consists of a speed controller, a motor, a planetary gearbox, a parallel gearbox and an electromagnetic brake. An accelerometer is installed in the vertical direction of the planetary gearbox to collect the original vibration acceleration signal. Several gearbox faults under different speed operating conditions are simulated by changing the speed of the motor. Seven different fault types of bearings and gears in the parallel shaft gearbox under three different speed operating conditions (20 Hz, 30 Hz and 40 Hz) are simulated, as shown in Table 2.

[0089] Table 2 Gearbox fault type description

[0090]

[0091] Based on the available bearing fault data and gearbox fault data above, the proposed multi-source adversarial deep migration network (Proposed) is compared with the multi-source domain convolutional neural network (MS-CNN), the MSANN model proposed in the literature Proc. AAAI Conf. Artif. Intell., Apr. 2018, “Multi-adversarial domain adaptation.”, the OSBP model proposed in the literature Proc. Eur. Conf. Comput. Vis., Sept. 2018, “Open set domain adaptation by backpropagation.”, the CDOAN model proposed in the literature Reliab. Eng. Syst. Saf., “Dual adversarial network for cross-domain open set fault diagnosis.” and the MS-OSDTN model proposed in the literature IEEE Trans. Cybern., Aug. 2022, “A Multi-Source Weighted Deep Transfer Network for Open-Set Fault Diagnosis of Rotary Machinery.”. To ensure the fairness of the experiment, the same feature extractor and classifier structure are used for the above methods except for the final fault prediction output layer. The network structure of the proposed method is shown in Table 3

[0092] Table 3 Network structure of the proposed method

[0093]

[0094]

[0095] In this experiment, a total of 6 bearing fault diagnosis tasks and 9 gearbox fault diagnosis tasks are set, as shown in Tables 4 and 5.

[0096] Table 4 Bearing fault multi-source domain open set fault diagnosis tasks

[0097]

[0098] Table 5 Gearbox fault multi-source domain open set fault diagnosis tasks

[0099]

[0100] In open set fault diagnosis, there are three important fault diagnosis effect evaluation indexes: potential unknown fault type recognition accuracy (Unk), known fault type classification accuracy (OS and overall accuracy (OA). The OA is the average of Unk and OS. The OA of the proposed method is 0.9999, which is 0.0001 higher than the MS-OSDTN model, 0.0002 higher than the MS-CNN model, 0.0003 higher than the MSANN model, 0.0004 higher than the CDOAN model and 0.0005 higher than the OSBP model.* ) and overall failure recognition accuracy (OS), which is calculated as follows:

[0101] Unk = C unk / N unk

[0102]

[0103]

[0104] where C Unk and N are the number of failure samples correctly recognized as unknown failure types and the number of failure samples correctly recognized as known failure types, respectively; N Unk and N are the number of unknown failure samples and the number of known failure samples, respectively. It is worth mentioning that when the proportion of unknown failures is too large, considering Unk or OS * alone as an evaluation index of the method will no longer be reasonable. Therefore, H-score is introduced as an evaluation index of the comprehensive accuracy of Unk and OS * . The formula is as follows:

[0105]

[0106] To obtain stable experimental results and minimize the influence of randomness on experimental results, five experiments were performed for each comparison method, and 500 model training was performed for each experiment. The specific failure diagnosis accuracy results are shown in Tables 6 and 7, and the accuracy bar charts of overall accuracy OS and H-score are shown in Figure 5 and Figure 6 . In order to further show the superiority of the proposed method (Proposed), the confusion matrix display means is adopted. The confusion matrix can reflect the failure recognition accuracy of each failure type. The confusion matrices of failure diagnosis tasks A2 and B1 randomly selected are plotted in Figure 7 and Figure 8 .

[0107] Table 5 Bearing fault multi-source domain open set failure diagnosis results

[0108]

[0109] Table 6 Gearbox fault multi-source domain open set failure diagnosis results

[0110]

[0111]

[0112] It can be seen that the proposed method (Proposed) is superior to the comparative method on all multi-source domain open set fault diagnosis tasks. Proposed achieves 92.06% of the OS average accuracy and 91.55% of the H-Score average accuracy in the bearing fault data set, and achieves 86.10% of the OS average accuracy and 86.81% of the H-Score average accuracy in the gearbox fault data set. The typical multi-source domain closed set method (MSANN) shows better performance than MS-CNN, because the alignment between domains is achieved through adversarial learning. However, MS-CNN and MSANN cannot identify potential unknown fault types in the target domain. Therefore, the H-Score of MS-CNN and MSANN is zero. In addition, CDOAN is superior to OSBP through double adversarial learning. As the latest multi-source domain open set fault diagnosis method, MS-OSDTN can also achieve multi-source open set fault diagnosis by considering the fusion of multiple source domain data as input, achieving 80.93% of the OS average accuracy and 83.59% of the H-score accuracy. But the diagnosis effect of MS-OSDTN is not as good as Proposed, because it simply fuses multiple source domain data into a single source domain, without considering the incomplete and suboptimal fault knowledge transfer caused by the domain difference and class transfer problems between multiple source domains, further degrading the fault diagnosis precision.

[0113] In addition, the following conclusions can also be drawn. First, the unsupervised transfer learning method can reduce the distribution difference between the data collected under different operating conditions, realize the transfer of fault knowledge and the fault diagnosis task. Second, compared with a single source domain, multiple source domains can contain more useful fault information. Third, merging or fusing multiple source domains into a single source domain is a solution to utilizing multiple source domain fault samples, but when the data distribution and label space of multiple source domains are different, the fault diagnosis performance will be hindered. Finally, for the above problems in multi-source domain open set fault diagnosis, Proposed can promote the transfer of fault knowledge and improve the fault diagnosis precision by separately considering the influence of distribution difference and class transfer between multiple source domains, instead of simply merging or fusing.

[0114] We also selected the gearbox fault diagnosis task B2 to examine the influence of different preset hyperparameters t c The influence on the fault diagnosis precision, and the bar chart results are as Figure 9are shown. In addition, we further carry out ablation experiments to verify the effectiveness of each module in the proposed method. Among them, the proposed method removes the multi-source fault type matching loss (No-MSFCM) and the proposed method removes the global alignment loss (No-GA) as two kinds of ablation methods, and the precision radar chart is drawn on the bearing fault multi-source domain open set fault diagnosis task and the gearbox multi-source domain open set fault diagnosis task as shown in Figure 10 and Figure 11 as shown.

[0115] The present application aims at the problems of class transfer between multiple source domains and the problem that potential unknown fault types in the target domain cannot be identified in multi-source domain open set fault diagnosis, and proposes a multi-source domain open set rotating machinery fault diagnosis method based on class transfer. The experimental results show that:

[0116] (1) The multi-source domain fault type matching module constructed by the present application can solve the class transfer problem by narrowing the same fault class centroid representation between different source domains and distancing the different fault class centroid representation, so as to enable the model to obtain more accurate and rich fault knowledge and promote the knowledge transfer of fault knowledge between source domains and target domains.

[0117] (2) The improved binary cross-entropy loss with additional constraints is constructed to reduce the domain gap between each source domain and the target domain and enable the domain-specific fault classifier to distinguish known fault types and unknown fault types. At the same time, after the target domain fault samples are predicted by multiple domain-specific fault classifiers, a multi-fault classification comprehensive decision module is constructed to obtain more comprehensive and universal target domain fault sample prediction.

[0118] (3) According to the comprehensive prediction of the target domain fault samples, the pseudo-fault label corresponding to the fault samples is obtained, and the centroid representation of the fault classes in the target domain is constructed. The Euclidean distance between the averaged multi-source fault class centroid and the target domain fault type centroid is calculated, and the global alignment loss is constructed to optimize the domain-invariant fault feature extraction ability of the shared feature extractor.

[0119] The proposed multi-source adversarial deep transfer network can effectively solve the problems of class transfer between multiple source domains and the problem that potential unknown fault types in the target domain cannot be identified, and can effectively improve the precision of class transfer multi-source domain open set rotating machinery fault diagnosis.

[0120] Those skilled in the art can understand that all or part of the steps in the above embodiments can be completed by programs instructing related hardware, and the programs can be stored in a computer readable storage medium, which can include ROM, RAM, magnetic disk or optical disk, etc.

[0121] While embodiments of the application have been shown and described, it is to be understood that the embodiments described are merely exemplary of the principles and application of the present application. Numerous modifications and adaptions can be effected without departing from the spirit and scope of the present application, which is not limited to the exact construction and arrangement described. It is intended, therefore, to cover all modifications and adaptions that fall within the scope of the claims and their equivalents.

Claims

1. A method for fault diagnosis of rotating machinery based on multi-source domain open sets with category transfer, characterized in that, The method includes: Vibration data of key components of rotating machinery under multiple different operating conditions are collected using accelerometers, and the vibration data is preprocessed. Based on whether the vibration data has fault type labels, the fault type labeled data under multiple operating conditions are used as multiple source domain training sets, and the fault-free type labeled data under a certain operating condition are used as the target domain training set and the target domain test set. Multiple source domain training sets and target domain training sets are packaged into batches and then sequentially input into the multi-source adversarial deep transfer network. High-dimensional feature representations of fault samples from multiple source domain training sets and target domain training sets are extracted using the shared feature extractor of a multi-source adversarial deep transfer network. By utilizing the unique fault classifiers in various fields of multi-source adversarial deep transfer networks, the high-dimensional feature representations of fault samples are processed to predict the fault types of the fault samples. Based on the fault type labels of fault samples in each source domain, construct the centroid representation of fault type in each source domain. At the same time, calculate the centroid distance between the same fault type labels in different source domain training sets and the centroid distance between different fault type labels to construct the matching loss of multi-source fault types. The matching loss for the multi-source fault type is expressed as: in, D represents the matching loss for multi-source fault types; shared D represents the centroid distance between training sets of the same fault type from different source domains; private This represents the centroid distance between different fault type labels in different source domain training sets. The centroid distances between the fault type labels are respectively represented as: Where M represents the number of training sets in the source domain, and Representing the source domain S respectively g and source domain S h Logarithm of the same fault type and logarithm of different fault types; Represents the source domain S g The fault type label space within; Represents the source domain S h The fault type label space within; For source domain S g The centroid of the l-th fault type in the data. For source domain S h The centroid of the l-th fault type; For source domain S g The centroid of the p-th fault type in the data. For source domain S h The centroid of the q-th fault type in the data; Represents the L2 norm; The centroid of the fault type in the source domain is represented as: in, Represents the source domain S m The centroid of the nth fault type in the matrix, m∈{1,2,...,M}, For the corresponding source domain S m Batch,I (·) It is an exponential function. This is the b-th source domain fault sample. The corresponding fault sample label, yes via shared feature extractor G fe (·) Extracted fault sample features; Based on the fault type prediction and the true fault type label of the fault samples output by the fault classifier unique to the source domain in different source domain training sets, the fault multi-classification loss is calculated. Based on the fault type prediction of the fault samples output by the fault classifiers unique to each source domain in the target domain training set, the adversarial classification loss of the multi-classifier is calculated. Based on the fault type prediction results of the fault samples in the target domain training set by the fault classifiers unique to each source domain, the prediction information entropy is calculated and the confidence weight of each prediction is constructed. Pseudo-fault type labels of fault samples in the target domain training set are generated, the fault class centroid in the target domain is constructed, and the global fault alignment is constructed with the average centroid between multiple source domains. The global fault alignment loss is calculated. The multi-source adversarial deep transfer network is iteratively trained by optimizing the matching loss, fault multi-classification loss, adversarial classification loss, and global fault alignment loss. The target domain test set is input into the trained multi-source adversarial deep transfer network, and the fault diagnosis results of the target domain test set are output.

2. The method for fault diagnosis of rotating machinery based on class transfer and multi-source domain open sets according to claim 1, characterized in that, Preprocessing the fault vibration data includes converting the acquired time-domain fault vibration signal into a frequency-domain signal using a fast Fourier transform.

3. The method for fault diagnosis of rotating machinery based on class transfer and multi-source domain open sets according to claim 1, characterized in that, The fault multi-classification loss is represented as follows: in, It is the source domain S m Multi-classification of fault losses It is the source domain S m The number of fault samples in the data. It is a multi-class cross-entropy loss function. It is the source domain S m The i-th fault sample Predicted fault labels, It is the source domain S m The i-th fault sample The corresponding fault sample label, It is a fault classifier unique to the m-th domain.

4. The method for fault diagnosis of rotating machinery based on class transfer and multi-source domain open sets according to claim 1, characterized in that, The adversarial classification loss of the multi-classifier is expressed as: in, This represents the adversarial classification loss of the multi-classifier in the m-th domain. Represents the binary cross-entropy loss across m domains. Let represent the constraint loss of the m-th source domain; The target domain fault sample training set is given, and T is a fixed threshold used as a boundary threshold to distinguish between known fault types and potentially unknown fault types in the target domain. It is a target domain fault sample The fault prediction output by the shared feature extractor and the domain-specific fault classifier The probability value of the dimension. It is a target domain fault sample via shared feature extractor G fe (·) Extracted fault sample features, This represents a fault classifier unique to the m-th domain; t c It uses a predefined threshold to constrain the classification boundaries between fault categories. It represents the number of fault categories in each source domain.

5. The method for fault diagnosis of rotating machinery based on class transfer and multi-source domain open sets according to claim 1, characterized in that, The global alignment loss is expressed as: in, It is the global alignment loss; |K| represents the number of fault categories in all source domains; It is the centroid representation of the l-th fault category in the target domain training set. It is the fault category centroid that averages all source domain fault categories.

6. The method for fault diagnosis of rotating machinery based on class transfer and multi-source domain open sets according to claim 5, characterized in that, The centroid of the fault type in the target domain is represented as: in, It is the global alignment loss; |K| represents the number of fault categories in all source domains; B is the centroid representation of the l-th fault category in the target domain training set. T It is a batch of fault samples participating in the training set of the target domain. It is the fault category centroid that averages all source domain fault categories; These are pseudo-labels for fault samples in the target domain. This represents the comprehensive prediction of fault samples in the target domain. These are the fault prediction weights of a fault classifier unique to the corresponding domain. It is a target domain fault sample The prediction vector and its length is H(·) is the information entropy function. It is a target domain fault sample via shared feature extractor G fe (·) Extracted fault sample features, This represents a fault classifier unique to the m-th domain.

Images