Fault migration diagnosis method based on balanced mixed adversarial and smooth rejection labels

Abstract

The application provides a fault migration diagnosis method based on balanced mixed confrontation and smooth suppression label, obtains a vibration signal of a rolling bearing; obtains a fault diagnosis result according to the obtained vibration signal, in combination with an optimized balanced mixed confrontation distribution and a smooth suppression label refinement network; wherein in the optimization of the balanced mixed confrontation distribution and the smooth suppression label refinement network, the fault types of a target domain data set are a subset of a source domain data set; the features of the source domain samples are enhanced to the features of the target domain samples, the data features in the source domain and the target domain are mixed within the domains, and on the basis of the soft pseudo label of the target domain, the supplementary entropy is combined to balance and suppress the uncertainty prediction; and the application alleviates the limitation that the source domain and the target domain data have the same label space in the adaptive network task.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of fault diagnosis, and particularly relates to a fault migration diagnosis method based on balanced hybrid adversarial and smooth label suppression. BACKGROUND

[0002] The statements in this section merely provide background technology related to the present application and do not necessarily constitute prior art.

[0003] Rotating machinery is widely used in manufacturing and related fields. Gears, bearings and other key components of modern equipment are often subjected to different operating loads and harsh external environments, which greatly increases the probability of failure. Sudden failure can cause economic losses and even safety accidents. Therefore, it is very important to develop necessary fault diagnosis technology for equipment health monitoring. Related scholars have developed various advanced signal processing technologies, machine learning and deep learning methods for application in this field.

[0004] In recent years, with the rapid development of intelligent fault diagnosis technology, deep learning-based fault diagnosis methods have also attracted much attention. Data-driven fault diagnosis can better extract complex internal structures from raw input and perform high-dimensional feature representation, and has become an important method for machine health monitoring. Compared with traditional shallow network architecture, deep learning methods realize automatic learning of feature layers, but most of them are still limited to training and testing between the same data distribution. However, due to reasons such as changes in machine operating conditions and working environment, this same distribution assumption will not adapt to real industrial scenarios, resulting in a decline in the generalization performance of the model.

[0005] Transfer learning improves the feature learning of related target domain data by learning the features of source domain data, solving the fault diagnosis task under variable working conditions and cross-device situations. Domain adaptation, as a subclass of transfer learning, usually assumes that the source domain and the target domain have the same label space, but the training and testing data come from different distributions. The purpose is to use the data of the two domains to learn domain-invariant feature representation, thereby maximizing the reduction of domain differences, but the labels of the target domain data are not utilized. However, this domain adaptation still has a necessary premise that the fault types of the source domain and the target domain data are the same, i.e., the label space is consistent.

[0006] The inventors find that in real applications, it is relatively easy to obtain a complete source domain dataset, but it will be relatively difficult to collect a target domain dataset with exactly the same fault category. In general, fault diagnosis scenarios do not have a large number of fault categories contained in the source domain, but only a subset thereof. The existing domain adaptive network will have a significant performance decline in fault diagnosis under such a condition, mainly because of the presence of outlier source classes in the source domain dataset, which will cause negative transfer during inter-domain matching. SUMMARY

[0007] In order to solve the problems of the prior art, the application provides a fault migration diagnosis method based on balanced mixed confrontation and smooth suppression label, and proposes a new partial domain adaptive network for more realistic fault diagnosis tasks, which alleviates the restriction that the source domain and target domain data have the same label space in the adaptive network task.

[0008] In order to achieve the above purpose, the application adopts the following technical scheme:

[0009] The first aspect of the application provides a fault migration diagnosis method based on balanced mixed confrontation and smooth suppression label.

[0010] A fault migration diagnosis method based on balanced mixed confrontation and smooth suppression label comprises the following processes:

[0011] Obtain the vibration signal of the rolling bearing;

[0012] According to the obtained vibration signal, the balanced mixed confrontation distribution and the smooth suppression label refinement network are combined to obtain the fault diagnosis result;

[0013] In the optimization of the balanced mixed confrontation distribution and the smooth suppression label refinement network, the fault categories of the target domain dataset are a subset of the source domain dataset.

[0014] The source domain sample feature is used to enhance the target domain sample feature, the data features in the source domain and the target domain are mixed respectively, and on the basis of the soft pseudo label of the target domain, the supplementary entropy is combined to balance and suppress the uncertainty prediction.

[0015] As a further limitation of the first aspect of the application, when the soft pseudo label corresponding to the index of the real classification is less than or equal to the set threshold, all soft pseudo labels are less than or equal to the set threshold, and the prediction confidence of the sample is low, which will be excluded from the loss function of the smooth suppression label refinement of the target domain data.

[0016] As a further limitation of the first aspect of the application, the supplementary entropy is used to balance and suppress the information of the uncertain category in the source domain to obtain the loss function of the smooth suppression label refinement of the source domain data.

[0017] According to the loss function of the smooth suppression label refinement of the target domain data and the loss function of the smooth suppression label refinement of the source domain data, a loss function of the smooth suppression label refinement is obtained in combination with an adjustable parameter.

[0018] As a further limitation of the first aspect of the application, the loss function of the balanced mixed adversarial distribution comprises: the sum of the balanced adversarial distribution loss function, the mixed adversarial distribution loss function, the mixed source domain loss function and the mixed target domain loss function.

[0019] As a further limitation of the first aspect of the application, in combination with the class-level weight estimated by the target domain data, an optimized source domain classification loss function is obtained, and according to the loss function of the smooth suppression label refinement, the loss function of the balanced mixed adversarial distribution and the optimized source domain classification loss function, a final loss function is obtained.

[0020] The second aspect of the application provides a fault migration diagnosis system based on balanced mixed adversarial and smooth suppression label.

[0021] A fault migration diagnosis system based on balanced mixed adversarial and smooth suppression label, comprising:

[0022] The data acquisition module is configured to acquire the vibration signal of the rolling bearing;

[0023] The fault diagnosis module is configured to obtain a fault diagnosis result according to the acquired vibration signal in combination with the optimized balanced mixed adversarial distribution and smooth suppression label refinement network.

[0024] In the optimization of the balanced mixed adversarial distribution and smooth suppression label refinement network, the fault categories of the target domain data set are a subset of the source domain data set.

[0025] The source domain sample feature is used to enhance the target domain sample feature, the data features in the source domain and the target domain are mixed respectively, and on the basis of the soft pseudo label of the target domain, the supplementary entropy is combined to balance and suppress the uncertainty prediction.

[0026] As a further limitation of the second aspect of the application, in the fault diagnosis module, when the soft pseudo label corresponding to the index of the true classification is less than or equal to the set threshold, all the soft pseudo labels are less than or equal to the set threshold, and the prediction confidence of the sample is low, the sample will be excluded from the loss function of the smooth suppression label refinement of the target domain data.

[0027] As a further limitation of the second aspect of the application, in the fault diagnosis module, the supplementary entropy is used to balance and suppress the information of the uncertain category in the source domain, and the loss function of the smooth suppression label refinement of the source domain data is obtained.

[0028] According to the loss function of the smoothing and inhibition label refinement of the target domain data and the loss function of the smoothing and inhibition label refinement of the source domain data, a loss function of the smoothing and inhibition label refinement is obtained in combination with an adjustable parameter.

[0029] As a further limitation of the first aspect of the application, in the fault diagnosis module, the loss function of balancing the mixed adversarial distribution comprises: the sum of the balancing adversarial distribution loss function, the mixed adversarial distribution loss function, the mixed source domain loss function and the mixed target domain loss function.

[0030] As a further limitation of the first aspect of the application, in the fault diagnosis module, in combination with the class-level weight estimated by the target domain data, an optimized source domain classification loss function is obtained, and according to the loss function of the smoothing and inhibition label refinement, the loss function of balancing the mixed adversarial distribution and the optimized source domain classification loss function, a final loss function is obtained.

[0031] Compared with the prior art, the beneficial effects of the present application are:

[0032] 1. The present application innovatively proposes a fault migration diagnosis method based on balanced mixed adversarial and smoothed inhibition labels, and proposes a new partial domain adaptive network for more realistic fault diagnosis tasks, which alleviates the restriction that the source domain and target domain data have the same label space in the adaptive network task.

[0033] 2. The present application balances the mixed adversarial distribution strategy, not only balances the data features of the source domain and the target domain, but also further expands, so that the label distribution in different domains can be aligned with each other, more internal structures are explored, and an invariant latent space is obtained.

[0034] 3. The present application uses the smoothing-inhibition pseudo-label generated by the target domain to iteratively train the classifier through the smoothing and inhibition label refinement strategy, which not only enhances the distinguishability of the target domain feature space, but also helps to suppress the propagation of uncertain categories. BRIEF DESCRIPTION OF DRAWINGS

[0035] The drawings accompanying the specification of the present application form a part thereof, serve to provide further understanding of the present application, and together with the exemplary embodiments of the present application and their description, serve to explain the present application, and do not constitute an improper limitation of the present application.

[0036] Figure 1 The flowchart of the fault migration diagnosis method based on balanced mixed adversarial and smoothed inhibition labels provided for the embodiment 1 of the present application is shown in the figure.

[0037] Figure 2 The schematic diagram of the HFZZ rotating machinery fault simulation platform provided for the embodiment 1 of the present application is shown in the figure.

[0038] Figure 3A rolling bearing health condition diagram provided for the embodiment 1 of the present application;

[0039] Figure 4 A migration diagnosis task result schematic diagram under different working conditions provided for the embodiment 1 of the present application. DETAILED DESCRIPTION

[0040] The present application is further illustrated below in conjunction with the accompanying drawings and embodiments.

[0041] It should be noted that the following detailed description is exemplary and is intended to provide further explanation of the present application. Unless otherwise indicated, all technical and scientific terms used herein have the same meaning as would be understood by one of ordinary skill in the art to which the present application pertains.

[0042] It should be noted that the terms used herein are only for the purpose of describing specific embodiments and are not intended to limit the exemplary embodiments according to the present application. As used herein, the singular form is intended to include the plural form unless the context clearly indicates otherwise, and it should also be understood that when the terms "comprise" and / or "include" are used in the specification, there is a presence of a feature, step, operation, device, component, and / or combinations thereof.

[0043] The embodiments in the present application and the features in the embodiments can be combined with each other without conflict.

[0044] Embodiment 1

[0045] As shown in Figure 1 The embodiment 1 of the present application provides a fault migration diagnosis method based on balanced mixed confrontation and smooth suppression label, which includes the following processes:

[0046] S1: Signal acquisition and data set segmentation

[0047] The vibration signals of the mechanical equipment (rolling bearing) under different working conditions and different health conditions are collected from the experimental table, which constitutes the entire data set to be trained and tested. According to a certain proportion, it is divided into a labeled source domain data set and an unlabeled target domain data set, and the data is preprocessed by using the Z-score method. It should be noted that in the partial domain adaptive task, the fault types of the target domain data set should be a subset of the source domain data set.

[0048] S2: Network composition and model construction

[0049] The main architecture of the network is constructed, which contains the necessary feature extractor G, domain discriminator D and label classifier C.

[0050] S3: Objective function and network optimization

[0051] The source domain labeled data xs and target domain unlabeled data x t Input into the network, through the balanced mixed adversarial distribution BBA and the smooth suppression label refinement SSL strategy to design the corresponding loss function, finally unify the objective function, optimize the network parameters , construct the optimal balanced mixed adversarial distribution and smooth suppression label refinement network.

[0052] S4: Network testing and diagnosis results

[0053] In the test phase, the fixed parameter feature extractor G can be directly applied to obtain the high-dimensional mapping of data features, and the label classifier C is used for classification, and then the fault recognition result is obtained.

[0054] The network main body architecture in S2 is mainly composed of domain adversarial neural network DANN, which is introduced as follows:

[0055] Domain adversarial neural network and its corresponding variants are used to solve the problem of domain adaptation. It mainly maximizes the reduction of the difference between the source domain and the target domain by generating field invariant features. DANN is mainly composed of three parts: feature extractor G, domain discriminator D, and label classifier C, whose parameters are θ g , θ d and θ c . G is used to learn field-invariant features to confuse D; at the same time, D tries to distinguish between source domain samples and target samples. G and D are in a minimax game to reduce the difference between the source domain and the target domain. Finally, C is used to classify different labels. Generally speaking, DANN can be defined as:

[0056]

[0057]

[0058]

[0059] Where l ce represents the cross-entropy loss, and ζ is a hyperparameter that balances the source classification risk and domain adversarial. Unlike label flipping in GAN networks, DANN networks use a gradient reversal layer to optimize the objective function, making the two feature distributions as similar as possible in the minmax optimization process:

[0060]

[0061]

[0062] The balanced mixed adversarial distribution BBA strategy in S3 is introduced as follows:

[0063] The prediction score of the target domain sample is used to obtain the class level weight, and then applied to the sample in the source domain, which can filter out the outlier source class to some extent, avoid negative transfer, which can be regarded as a small domain adaptation task, which depends on the accurate prediction of target domain data to a great extent, otherwise, the outlier source class may be regarded as noise interference, participate in inter-domain matching, and cause negative transfer. Unlike this, the application adopts an opposite strategy and designs a large domain adaptation task, instead of filtering out the outlier source class first, but adopts a balanced adversarial distribution strategy to enhance the target domain sample features with the source domain sample features to achieve the purpose of balancing the domains. The formula of balanced adversarial distribution is as follows:

[0064]

[0065] Wherein, e(x) = 1 + e -H(h(x)) represent the entropy perception weight, and It can measure the difficulty of samples on the boundary and samples not on the boundary between different classes in domain adversarial alignment, represent the class level weight calculated by the target domain data, χ represents the variable weighting parameter, which decreases with the increase of the number of iterations.

[0066] In addition to adopting the strategy of randomly selecting source domain data features to transfer to target domain features to balance the adversarial distribution, the mixed adversarial distribution strategy is also adopted, that is, the data features in the source domain and the target domain are mixed respectively, the mixed adversarial distribution is adopted, the diversity of the intra-domain samples is increased, more internal structures are explored, and then it is beneficial to obtain the invariant latent space features. The formula of mixed adversarial distribution is as follows:

[0067]

[0068]

[0069]

[0070] Wherein, D bs represent the mixed source domain after mixing the source domain, containing sample and label and sample number n bs , D bt represent the mixed target domain after mixing the target domain, containing sample and label and sample number n bt , is the soft pseudo label of the target domain sample , which will be described in detail in the next part, and λ is the weighting mixing ratio, which is randomly selected from a Beta distribution.

[0071] The smoothing inhibition label refinement SSL policy in S3 is introduced as follows:

[0072] Although the domain difference between the source domain and the target domain can be effectively alleviated by balancing the mixed adversarial distribution, and the positive transfer is promoted to the maximum, the distinguishability of the target domain data cannot be guaranteed, because the target domain samples lack class-aware label information. The main purpose of this paper is to classify the target domain data, therefore, the decision boundary of the classifier with high confidence for the target domain data will become necessary. The calculation formulas of the traditional hard pseudo label and soft pseudo label are as follows:

[0073]

[0074]

[0075] wherein, the hard pseudo label is represented by H, the soft pseudo label is represented by S, is the i-th target domain data the Softmax probability output as network input, is a parameter for controlling the label smoothing degree.

[0076] On the basis of the soft label, the present application balances and inhibits the uncertain prediction by introducing supplementary entropy, and the target function of the smoothing inhibition label adopted is as follows:

[0077]

[0078]

[0079] wherein, a represents the index of the real classification, and η is another hyperparameter.

[0080] In the actual application process, the prediction of the model may not have high confidence for part of the target domain data, which indicates that the prediction result is not accurate, therefore, it is necessary to set a threshold value ε to refine the target function of the smoothing inhibition label, when , all will also be less than ε, the prediction confidence of the sample is low, and the sample will be excluded from the target function, then the target function of the smoothing inhibition label refinement of the target domain data is as follows:

[0081]

[0082] The present application not only suppresses the soft label uncertainty in the target domain, but also should suppress the noise in the source domain, which will effectively avoid the uncertainty factor from being propagated to the target prediction and confusing the feature distribution in the target domain. For the information of the uncertain class, the supplementary entropy is still used for balanced suppression, which will balance the lower prediction score for the incorrect class. The target function of the smoothed suppression label refinement of the source domain data is as follows:

[0083]

[0084]

[0085] In formula (15), the refinement strategy is also used, the samples less than the threshold value ε are removed, and the class-level weight w applied in formula (6) is used to weight each sample.

[0086] According to formula (14) and formula (15), the target function of the smoothed suppression label refinement with adjustable parameter v is as follows:

[0087]

[0088] The unified target function in S3 is introduced as follows:

[0089] In the process of training the classifier, the mixed data in the two mixed domains D bs and D bt are also needed to be used to respectively obtain the classification loss of the mixed source domain and the mixed target domain:

[0090]

[0091]

[0092] Of course, the classification loss in the above two formulas should also be included in the mixed adversarial distribution, so that the balanced mixed adversarial distribution loss function is not only composed of formula (6) and formula (7), but also should include the above two formulas:

[0093]

[0094] Wherein, α is an adjustable hyperparameter.

[0095] In actual application process, in order to pay attention to the initial transferable shared class, the source domain classification loss L cls in formula (2) should also apply the class-level weight estimated by the target domain data, so that the task becomes more compact, and the classification loss is as follows:

[0096]

[0097] Finally, all the terms in formula (17), (20), (21) are integrated on the source domain samples, target domain samples and the two constructed hybrid domain samples to facilitate the forward migration of the network, and the unified framework of the proposed network is obtained as follows:

[0098]

[0099]

[0100]

[0101] Wherein, κ1, κ2, κ3 are adjustable parameters, and the purpose of network optimization is to find the optimal parameters For partial domain adaptation task.

[0102] Specifically, the following examples are provided to build a HFZZ rotating machinery fault simulation platform, as shown in Figure 2 The platform is composed of a motor, a control system, a radial loading device, an acceleration sensor, etc. The original data is collected by the acceleration sensor with a collection frequency of 12.8 kHz. Three different rotating speeds are used to simulate the bearing running state under different working conditions, which are working condition A (1750 rpm), working condition B (2000 rpm) and working condition C (2250 rmp). The bearing fault is generated by processing, and eight kinds of health conditions including normal (NM), inner ring fault (IR), outer ring fault (OR), rolling element fault (BA) and various mixed faults are generated. The detailed information is shown in Figure 3 In the experiment, 200 groups of data are collected from each health condition under each running speed, each data contains 1024 data points, and the training set and test set are divided according to the ratio of 8:2. It is worth noting that this is the division of the initial data set in the domain adaptation task, and since the present application mainly verifies the partial domain adaptation task, the fault types in the test set should be correspondingly reduced.

[0103] In order to verify the performance of the algorithm proposed in the present application in the partial domain adaptation task, the target domain data set with relatively reduced fault number is constructed, which is H1-H6, wherein the former represents the source domain and the latter represents the target domain. The detailed description is shown in Table 1.

[0104] Table 1: Detailed description of partial domain adaptation task

[0105]

[0106] To verify the superiority of the proposed balanced mixed adversarial distribution and smooth suppression label refinement network, several state-of-the-art deep neural network algorithms are used, such as CNN (Baseline), DA-adv, WATN and two ablation strategies, namely no balanced mixed adversarial distribution (No-BBA) and no smooth suppression label (No-SSL), to compare the partial domain fault diagnosis tasks.

[0107] In the fault diagnosis experiment, ten tests were carried out, and the average value was taken as the experimental result as shown in Table 2. From the results, it can be seen that the proposed method obtains a relatively high average fault diagnosis accuracy of 96.57%. Next, it can be found that DA-adv produces a negative transfer effect when solving the partial domain adaptation problem, resulting in poor diagnosis performance. Then, the WATN method can greatly improve the performance of the partial domain adaptation task. Of course, the method proposed in this paper still occupies a higher accuracy among the compared methods. At the same time, the ablation method using two different strategies, although inferior to the WATN method based on partial domain adaptation in some tasks, has similar overall effect to WATN and also achieves good diagnosis effect. In comparison, the No-SSL method equipped with BBA strategy has slightly better effect. Overall, the proposed method achieves superior performance and improves the overall average accuracy.

[0108] Table 2: Diagnosis results of partial domain transfer tasks

[0109]

[0110] It is also necessary to verify the diagnosis performance of the proposed method under different working conditions, so six different partial domain adaptation tasks are set in this section, and the specific details are described in Table 3, Figure 4 The diagnosis results are shown. It can be found that the diagnosis performance of all methods has decreased to a certain extent. Especially in task S3, the proposed algorithm does not achieve the optimal diagnosis performance. However, overall, the proposed algorithm still occupies a relatively high diagnosis accuracy. This is also in line with the expected result.

[0111] Table 3: Table of transfer diagnosis tasks under different working conditions

[0112]

[0113] Example 2

[0114] The embodiment 2 of the present application provides a fault transfer diagnosis system based on balanced mixed adversarial and smooth suppression label, comprising:

[0115] The data acquisition module is configured to acquire the vibration signal of the rolling bearing.

[0116] The fault diagnosis module is configured to obtain a fault diagnosis result according to the acquired vibration signal, in combination with the optimized balanced mixed adversarial distribution and the smoothing and inhibiting label refinement network.

[0117] In the optimization of the balanced mixed adversarial distribution and the smoothing and inhibiting label refinement network, the fault categories of the target domain data set are a subset of the source domain data set.

[0118] The source domain sample feature is used to enhance the target domain sample feature, the data features in the source domain and the target domain are mixed within the domain, and on the basis of the soft pseudo label of the target domain, the uncertainty prediction is balanced and inhibited in combination with the supplementary entropy.

[0119] The working method of the system is the same as the fault migration diagnosis method based on balanced mixed adversarial and smoothing and inhibiting label provided in Embodiment 1, and will not be described here.

[0120] Those skilled in the art should understand that the embodiments of the present application can be provided as a method, a system, or a computer program product. Therefore, the present application can take the form of a hardware embodiment, a software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can take the form of a computer program product implemented on one or more computer usable storage media (including but not limited to magnetic disk storage and optical storage, etc.) containing computer usable program code.

[0121] The present application is described with reference to flowcharts and / or block diagrams according to the method, device (system) and computer program product of the embodiments of the present application. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, and the combination of the flows and / or blocks in the flowcharts and / or block diagrams can be realized by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device produce a device that implements the functions specified in the flowcharts and / or block diagrams. Figure 1 The functions specified in one flow or multiple flows and / or blocks Figure 1 The device that implements the functions specified in one block or multiple blocks.

[0122] These computer program instructions can also be stored in a computer readable storage medium that can guide the computer or other programmable data processing device to work in a specific way, so that the instructions stored in the computer readable storage medium produce a product including instruction devices, which implement the functions specified in the flowcharts and / or block diagrams. Figure 1 The functions specified in one flow or multiple flows and / or blocks Figure 1 The device that implements the functions specified in one block or multiple blocks.

[0123] These computer program instructions can also be loaded into a computer or other programmable data processing device, so that a series of operational steps are performed on the computer or other programmable data processing device to generate a computer-implemented process, thus the instructions executed on the computer or other programmable data processing device provide the function of implementing the processes specified in the flowchart Figure 1 one flow or multiple flows and / or the functions specified in the block Figure 1 one block or multiple blocks.

[0124] Those of ordinary skill in the art can understand that all or part of the flow of the above-mentioned embodiment method can be completed by instructing the relevant hardware through a computer program, and the program can be stored in a computer-readable storage medium. When the program is executed, it can include the flow of the above-mentioned embodiment of each method. The storage medium can be a magnetic disc, an optical disc, a read-only memory (ROM), a random access memory (RAM), and the like.

[0125] The above only describes the preferred embodiments of the present application and is not intended to limit the present application. For those skilled in the art, the present application can have various modifications and changes. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A fault migration diagnosis method based on balanced mixed adversarial and smooth rejection labels, characterized by, The method comprises the following steps: obtaining a vibration signal of a rolling bearing; obtaining a vibration signal of a rolling bearing; wherein in the optimization of the balanced mixed adversarial distribution and the smoothed suppression label refinement network, the fault types of the target domain dataset are a subset of the source domain dataset; The source domain sample feature enhances the target domain sample feature, and the data features in the source domain and the target domain are mixed in the domain, and on the basis of the soft pseudo label of the target domain, the uncertainty prediction is balanced and suppressed by combining the supplementary entropy; The network is composed of a domain adversarial neural network (DANN), which is composed of three parts: a feature extractor , a domain discriminator , and a label classifier , whose parameters are , respectively; is used to learn domain-invariant features to confuse ; then tries to distinguish source domain samples from target samples; and minimize the difference between the source domain and the target domain in a minimax game; finally is used to classify different labels; The balanced mixed adversarial distribution adopts a balanced adversarial distribution strategy, uses the source domain sample feature to enhance the target domain sample feature, and achieves the purpose of inter-domain balance, and the formula of the balanced adversarial distribution is as follows: wherein, is source domain labeled data; is target domain unlabeled data, represents entropy-aware weights, is used to measure the difficulty of domain adversarial alignment for samples on the boundary and samples not on the boundary between different classes, represents class-level weights calculated through target domain data, represents a variable trade-off parameter, which decreases with the increase of the number of iterations; The balanced mixed adversarial distribution also adopts a mixed adversarial distribution strategy, that is, the data features in the source domain and the target domain are mixed in the domain, and the formula of the mixed adversarial distribution is as follows: wherein, represents the mixed source domain after mixing, containing samples and labels and the number of samples , represents the mixed target domain after mixing, containing samples and labels and the number of samples , is the soft pseudo label of the target domain sample , is the trade-off mixing ratio, randomly selected from a Beta distribution; The smoothed suppression label refinement adopts a smoothed suppression label target function as follows: wherein, an index representing a true classification, is a hyperparameter; Setting a threshold The target function of the smoothing suppression label is refined when All will also be less than The prediction confidence of the sample is low, and it will be excluded from the target function. The target function of the smoothing suppression label refinement of the target domain data is as follows: 。 2. The fault migration diagnosis method based on balanced mixed adversarial and smoothed suppression label according to claim 1, wherein when the index corresponding to the real classification is less than or equal to the set threshold, all soft pseudo labels are less than or equal to the set threshold, and the prediction confidence of the sample is low, the sample is excluded from the loss function of the smoothed suppression label refinement of the target domain data.

3. The fault migration diagnosis method based on balanced mixed adversarial and smoothed suppression label according to claim 2, wherein the supplementary entropy is used to balance and suppress the information of the uncertain category in the source domain, and the loss function of the smoothed suppression label refinement of the source domain data is obtained. According to the loss function of the smoothed suppression label refinement of the target domain data and the loss function of the smoothed suppression label refinement of the source domain data, the loss function of the smoothed suppression label refinement is obtained by combining the adjustable parameter.

4. The fault migration diagnosis method based on balanced mixed adversarial and smoothed suppression label according to claim 3, wherein the loss function of the balanced mixed adversarial distribution comprises the sum of the balanced adversarial distribution loss function, the mixed adversarial distribution loss function, the mixed source domain loss function and the mixed target domain loss function.

5. The fault migration diagnosis method based on balanced mixed adversarial and smoothed suppression label according to claim 4, wherein the class-level weight estimated by the target domain data is combined to obtain the optimized source domain classification loss function, and the final loss function is obtained according to the loss function of the smoothed suppression label refinement, the loss function of the balanced mixed adversarial distribution and the optimized source domain classification loss function. It comprises: a data acquisition module configured to obtain a vibration signal of a rolling bearing; a fault diagnosis module configured to obtain a vibration signal of a rolling bearing; wherein in the optimization of the balanced mixed adversarial distribution and the smoothed suppression label refinement network, the fault types of the target domain dataset are a subset of the source domain dataset; 6. A fault migration diagnosis system based on balanced mixed adversarial and smooth rejection labels, characterized in that, ​ ​ ​ ​ The source domain sample feature is used to enhance the target domain sample feature, and the data features in the source domain and the target domain are mixed in the domain, and on the basis of the soft pseudo label of the target domain, the supplementary entropy is combined to balance and suppress the uncertain prediction. The network is composed of a domain adversarial neural network (DANN), which is composed of three parts: a feature extractor , a domain discriminator , and a label classifier , whose parameters are , , respectively; is used to learn domain-invariant features to confuse ; then tries to distinguish source domain samples from target samples; and minimize the difference between the source domain and the target domain in a minimax game; finally is used to classify different labels; The balance mixed adversarial distribution adopts a balance adversarial distribution strategy, uses the source domain sample feature to enhance the target domain sample feature, and achieves the purpose of domain balance, and the formula of the balance mixed adversarial distribution is as follows: wherein, is source domain labeled data; is target domain unlabeled data, represents an entropy-aware weight, is used to measure the difficulty of domain-adversarial alignment for samples located on the boundary between different classes and samples not located on the boundary, represents a class-level weight calculated through target domain data, represents a variable trade-off parameter which decreases with the increase of the number of iterations; The balance mixed adversarial distribution also adopts a mixed adversarial distribution strategy, that is, the data features in the source domain and the target domain are mixed in the domain, and the formula of the mixed adversarial distribution is as follows: wherein, represents the mixed source domain after mixing, containing samples and labels and the number of samples , represents the mixed target domain after mixing, containing samples and labels and the number of samples , is the soft pseudo label of the target domain sample , is the trade-off mixing ratio, randomly selected from a Beta distribution; The smoothing suppression label refinement adopts a smoothing suppression label target function as follows: wherein, an index representing a true classification, is a hyperparameter; Setting a threshold The target function of the smoothing suppression label is refined when All Will also be less than The prediction confidence of the sample is low, and will be excluded from the target function. The target function of the smoothing suppression label refinement of the target domain data is as follows: 。 7. The fault migration diagnosis system based on balance mixed adversarial and smoothing suppression label according to claim 6, wherein, In the fault diagnosis module, when the soft pseudo label corresponding to the index of the true classification is less than or equal to the set threshold, all the soft pseudo labels are less than or equal to the set threshold, and the prediction confidence of the sample is low, the sample is excluded from the loss function of the smoothing suppression label refinement of the target domain data.

8. The fault migration diagnosis system based on balance mixed adversarial and smoothing suppression label according to claim 7, wherein, In the fault diagnosis module, the supplementary entropy is used to balance and suppress the information of the uncertain category in the source domain, and the loss function of the smoothing suppression label refinement of the source domain data is obtained; According to the loss function of the smoothing suppression label refinement of the target domain data and the loss function of the smoothing suppression label refinement of the source domain data, the loss function of the smoothing suppression label refinement is obtained in combination with the adjustable parameter.

9. The fault migration diagnosis system based on balance mixed adversarial and smoothing suppression label according to claim 8, wherein, In the fault diagnosis module, the loss function of the balance mixed adversarial distribution includes the sum of the balance adversarial distribution loss function, the mixed adversarial distribution loss function, the mixed source domain loss function and the mixed target domain loss function.

10. The fault migration diagnosis system based on balance mixed adversarial and smoothing suppression label according to claim 9, wherein, In the fault diagnosis module, the class-level weight estimated by the target domain data is combined to obtain the optimized source domain classification loss function, and the final loss function is obtained according to the loss function of the smoothing suppression label refinement, the loss function of the balance mixed adversarial distribution and the optimized source domain classification loss function.

Images