Production equipment open set fault diagnosis model and method based on multi-source domain weighted migration

Abstract

The invention relates to the technical field of internet big data and new generation information, in particular to a production equipment open set fault diagnosis model and method based on multi-source domain weighted migration, and the method comprises the steps: S1, obtaining a data set of a target domain and n source domains; s2, performing feature extraction on the data of the target domain and the n source domains to obtain deep features of the data of the target domain and the n source domains; s3, training based on the deep features of the data of the n source domains to obtain an auxiliary domain classifier; s4, inputting the deep features of the data of the target domain into the auxiliary domain classifier to obtain similarity weights between the data of the target domain and the data of the n source domains; s5, respectively training n source classifiers based on the deep features of the data of the n source domains; s6, performing weighted combination on the n source classifiers based on similarity weights between the data of the target domain and the data of the n source domains to obtain a complementary classifier; and S7, inputting the data of the target domain into the complementary classifier, and outputting a corresponding prediction fault type label. According to the invention, the accuracy of open set fault diagnosis of the production equipment can be improved.

Description

Technical Field

[0001] The present invention relates to the field of Internet big data and new generation information technology, and particularly to an open-set fault diagnosis model and method for production equipment based on multi-source domain weighted transfer. Background Art

[0002] With the deep penetration of industrial automation and intelligent manufacturing, production equipment has become the core carrier to support the efficient operation of the modern industrial system. In key fields such as automobile manufacturing, consumer electronics, and precision instruments, the stable operation of the equipment cluster directly determines the production line rhythm, product quality, and delivery capacity, and is the "skeletal system" to ensure the continuous output of social productivity.

[0003] However, industrial production equipment is long-term in a complex working condition of high load and multi-task parallel, and is prone to cause progressive performance decline due to mechanical wear, electrical component aging, control system disorder, or human operation error. These latent faults only show slight vibration abnormalities, energy consumption fluctuations, or precision offsets in the initial stage, but rapidly deteriorate when the production load suddenly increases or the process parameters mutate, resulting in line-wide chain shutdowns, batch product scrap, and even safety accidents. Therefore, building a multi-dimensional and highly sensitive fault diagnosis system to accurately identify the characteristic laws of typical fault modes such as wear, fatigue, and overheating has great engineering value for realizing predictive maintenance, minimizing unplanned downtime, and maintaining the economic resilience of the industry.

[0004] The fault diagnosis method based on domain adaptation technology (DA) is the most widely used method to determine the fault types of production equipment at present. This method maps the diagnostic knowledge extracted from multi-source heterogeneous data to a unified feature space, and uses the shared features of data by similar machines for fault diagnosis. However, these methods can only identify the fault types included in the source domain, and it will be difficult to correctly identify when there are specific fault types in the target domain. The source domain data and the target domain data are obtained from different working conditions of production equipment. The two share some of the same fault types, but there will also be some specific fault types in the target domain. This is the open-set fault diagnosis problem.

[0005] The applicant found in the research that a large amount of labeled data can be obtained from multiple source domains that follow different distributions, and the fault types missing in a single source domain can be supplemented. The fault diagnosis performance of the model in the target task can be improved by using the method of multi-source information fusion. However, there is currently a lack of an effective strategy to make full use of multi-source information, and to optimize the fault diagnosis accuracy of the target domain through weighted fusion based on the similarity between the source domain and the target domain.

[0006] In summary, for the existing diagnosis methods of open-set fault types in production equipment, it is difficult to correctly identify the specific fault types that appear in the target domain, and the open-set fault diagnosis problem cannot be solved. Although the single-source domain open-set DA fault diagnosis can partially solve the open-set fault diagnosis problem, the accuracy of its diagnosis results is still not ideal. Therefore, it is necessary to make full use of multi-source information for fusion and establish an effective multi-source domain weighted fusion strategy to improve the accuracy of fault type identification in the open-set fault diagnosis of production equipment. Summary of the Invention

[0007] Aiming at the deficiencies of the above-mentioned existing technologies, the technical problem to be solved by the present invention is: how to provide an open-set fault diagnosis model and method for production equipment based on multi-source domain weighted transfer, which can realize complete fault prediction for the target domain by learning complementary knowledge in multiple source domains, solve the problem of identifying fault types in the target domain that are not available in a single source domain, enhance the fault diagnosis performance in the cross-condition and cross-machine transfer diagnosis tasks of production equipment, and thus improve the accuracy of open-set fault diagnosis of production equipment.

[0008] To solve the above technical problems, the present invention adopts the following technical solutions:

[0009] An open-set fault diagnosis model and method for production equipment based on multi-source domain weighted transfer, comprising:

[0010] S1: Obtain the data sets of the target domain and n source domains; the data set of the target domain includes the data of the target domain; the data sets of the source domains include the data of the source domains and their corresponding labels;

[0011] S2: Extract features from the data of the target domain and n source domains through a feature extractor to obtain the deep features of the data of the target domain and n source domains;

[0012] S3: Train a classifier based on the deep features of the data of n source domains to obtain an auxiliary domain classifier;

[0013] S4: Input the deep features of the data of the target domain into the auxiliary domain classifier, and output the similarity weights between the data of the target domain and the data of n source domains;

[0014] S5: Train n classifiers respectively based on the deep features of the data of n source domains to obtain n source classifiers;

[0015] S6: Weightedly combine the n source classifiers based on the similarity weights between the data of the target domain and the data of n source domains to obtain a complementary classifier;

[0016] S7: Input the data of the target domain into the complementary classifier, and output the corresponding predicted fault type labels.

[0017] Preferably, in step S1, the time-series signal data of the target domain and the source domain are converted into wavelet time-frequency diagrams through continuous wavelet transform, and then the subsequent steps are executed.

[0018] Preferably, in step S2, a feature extractor is constructed through the following steps:

[0019] S201: Construct an encoder G based on an autoencoder framework f (x j ,θ f ) and a decoder G y (x j ,θ y );

[0020] S202: Input the data x of the target domain and the source domain j into the encoder G f (x j ,θ f ) to extract the corresponding deep features;

[0021] S203: Obtain the posterior distribution p(Z|X) of the deep features through two fully connected layers, and sample from the posterior distribution to obtain the

[0022] S204: Input into the decoder G y (x j ,θ y ) for reconstruction to obtain the reconstructed data

[0023] S205: Calculate the mean square error loss function and KL divergence based on the original data x j and the reconstructed data to train the encoder G f (x j ,θ f ) and the decoder G y (x j ,θ y );

[0024] S206: Repeat steps S202 and S205 to iteratively train the encoder G f (x j ,θ f ) and the decoder G y (x j ,θ y ), and use the trained encoder G f (x j ,θ f ) as the feature extractor.

[0025] Preferably, in step S205, the formula of the mean square error loss function is expressed as:

[0026]

[0027] In the formula: E G represents the reconstruction loss, m S and m T respectively represent the number of data in the source domain and the target domain, θ f 、θ y represent the parameters to be optimized for the encoder and decoder, x j represents the original data, represents the reconstructed data, S and T respectively represent the data sets of the source domain and the target domain;

[0028] The formula of the KL divergence is expressed as:

[0029]

[0030] In the formula: σ j 、μ j represent the mean and variance after the distribution transformation; σ is the noise intensity.

[0031] Preferably, in step S3, when training the auxiliary domain classifier, the cross-entropy loss function is calculated by combining the deep features of all the source domain data with the corresponding data labels to optimize the parameters of the auxiliary domain classifier;

[0032] The formula of the cross-entropy loss function of the auxiliary domain classifier is expressed as:

[0033]

[0034] In the formula: θ d represents the parameter to be optimized for the auxiliary domain classifier, d j represents the domain label.

[0035] Preferably, in step S4, the deep features of the target domain data are input into the auxiliary domain classifier, and the deep features are mapped to the domain label space {0, 1, 2,..., n} through the softmax function, and the probability that the target domain data is assigned to different source domains is used as the similarity weight between the target domain data and the data of each source domain;

[0036] The formula is expressed as:

[0037]

[0038] In the formula: ω n includes n to {ω1, ω2,..., ω n} respectively represent the similarity weights between the target domain samples and each source domain sample, and k and b are the weights and biases of the auxiliary domain classifier respectively.

[0039] Preferably, in step S205, the encoder G f (x, θ f ) and the decoder G y (x, θ y ) are trained by incorporating the similarity weights ω n between the data in the target domain and the data in each source domain into the mean squared error loss function and KL divergence of the encoder G f (x, θ f ) and the decoder G y (x, θ y );

[0040] The updated formula of the mean squared error loss function is expressed as:

[0041]

[0042] In the formula: m S and m T respectively represent the number of data in the source domain and the target domain, θ f , θ y represent the parameters to be optimized of the encoder and decoder, x j represents the original data, represents the reconstructed data, and S and T respectively represent the data sets of the source domain and the target domain;

[0043] The updated formula of the KL divergence is expressed as:

[0044]

[0045] In the formula: σ j , μ j represent the mean and variance after the distribution transformation; σ is the noise intensity.

[0046] Preferably, in step S5, the formula of the source classifier corresponding to each source domain is expressed as:

[0047]

[0048] In the formula: represents the source classifier of the p-th source domain, θ c represents the parameter to be trained of the source classifier, k c and b c respectively represent the weights and biases of the source classifier.

[0049] Preferably, when training the source classifier, the cross-entropy loss function is calculated by combining the deep features of the data corresponding to the source domain with the labels of the corresponding data to optimize the parameters of the source classifier;

[0050] The formula for the cross-entropy loss function of the source classifier is expressed as:

[0051]

[0052] In the formula: y j represents the label of the data x j .

[0053] Preferably, in step S6, the calculation formula of the complementary classifier is expressed as:

[0054]

[0055] In the formula: represents the combination of the source classifiers of each source domain, represents the source classifier of the p-th source domain, represents the weights and biases of the source classifier of the p-th source domain.

[0056] Compared with the prior art, the production equipment open-set fault diagnosis model and method based on multi-source domain weighted transfer in the present invention have the following beneficial effects:

[0057] The present invention realizes the complete fault prediction of the target domain by learning complementary knowledge in multiple source domains, solves the problem of identifying fault types in the target domain that are not included in a single source domain, enhances the fault diagnosis performance in the cross-condition and cross-machine transfer diagnosis tasks of production equipment, and improves the accuracy of open-set fault diagnosis of production equipment. First, a feature extraction model is constructed using the framework of an autoencoder. Each source domain dataset no longer cares about the fault types within the domain, but uses the global differences between source domains as the classification basis of the model. At the same time, the encoder is synchronously optimized through the classification loss of the domain classifier to extract deep features representing the differences between domains. Then, the target domain data is input into the trained domain classifier to obtain the probability that the fault types in the target domain are also included in the i-th source domain, and an auxiliary domain classifier is constructed accordingly. Secondly, a domain fusion module based on the auxiliary domain classifier is used to reduce the feature distribution difference between the source domain and the target domain. The (preprocessed) source domain data x S and the target domain data x TAt the same time, as input, the source domain and target domain data are synchronously processed through a shared encoder and decoder, thereby reducing the difference in feature distribution between the two. The variational principle of the variational autoencoder is used to optimize the decoder, and the generalization ability of the model is improved by adding noise and KL divergence, enhancing the model's ability to extract domain-invariant features. In addition, through the adaptive weighted learning strategy based on the domain classifier, the similarity degree between different target domain data and the source domain data can be distinguished, enabling the domain adaptation model to extract more important features, avoiding negative transfer caused by specific categories in the target domain, and promoting positive transfer from the source domain to the target domain. Finally, the present invention uses the decision fusion module based on the auxiliary domain classifier to weight and fuse the classification results of multiple source classifiers using the similarity scores of the fault data. Only the samples sharing the same category as the source domain have high confidence in the classification results using the source classifier. By adaptively adjusting the contributions of the source domain and target domain data to model training, more refined feature alignment is achieved, improving the open-set fault diagnosis accuracy of production equipment in cross-condition and cross-machine migration scenarios. Description of the Drawings

[0058] To make the objectives, technical solutions, and advantages of the invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings, where:

[0059] Figure 1 It is the network structure diagram of the fault type perception strategy.

[0060] Figure 2 and Figure 3 They are the network structure diagram and flowchart of the open-set fault diagnosis model and method (MSWDAN) for production equipment based on multi-source domain weighted migration.

[0061] Figure 4 It is the feature visualization result.

[0062] Figure 5 It is the ablation experiment result of MSWDAN.

[0063] Figure 6 They are the confusion matrices of MSDAN && MSWDAN.

[0064] Figure 7 It is the cross-machine migration fault diagnosis result. Detailed Embodiments

[0065] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. Components of the embodiments of the present invention generally described and illustrated in the figures herein can be arranged and designed in a variety of different configurations. Therefore, the detailed description of the embodiments of the present invention provided herein is not intended to limit the scope of the claimed invention, but is merely representative of selected embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts fall within the scope of protection of the present invention.

[0066] It should be noted that like reference numerals and letters denote like items in the following figures. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. In the description of the present invention, it should be noted that the terms "center", "upper", "lower", "left", "right", "vertical", "horizontal", "inner", "outer", etc. indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, or the orientation or positional relationship in which the inventive product is customarily placed during use. This is only for the convenience of describing the present invention and simplifying the description, and does not indicate or imply that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and thus should not be construed as a limitation of the present invention. In addition, the terms "first", "second", "third", etc. are only used for descriptive distinction and should not be construed as indicating or implying relative importance. In addition, terms such as "horizontal" and "vertical" do not mean that the components are required to be absolutely horizontal or hanging, but can be slightly inclined. For example, "horizontal" only means that its direction is more horizontal relative to "vertical", and does not mean that the structure must be completely horizontal, but can be slightly inclined. In the description of the present invention, it should also be noted that unless otherwise clearly specified and limited, the terms "set", "installed", "connected", and "connected" should be understood in a broad sense. For example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be directly connected, or indirectly connected through an intermediate medium, and can be the communication inside two elements. For those of ordinary skill in the art, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0067] The following will be further described in detail through specific embodiments:

[0068] Embodiment:

[0069] In this embodiment, an open-set fault diagnosis model and method for production equipment based on multi-source domain weighted migration are disclosed.

[0070] AsFigure 1 and Figure 2 As shown, the open-set fault diagnosis model and method for production equipment based on multi-source domain weighted transfer include:

[0071] S1: Obtain the data sets of the target domain and n source domains;

[0072] The data set of the target domain includes the data of the target domain q represents the number of data in the target domain; the data set of the source domain includes the data of the source domain and the corresponding labels m represents the number of data in the source domain;

[0073] S2: Extract the features of the data in the target domain and all source domains through a feature extractor to obtain the deep features of the data in the target domain and n source domains;

[0074] S3: Train a classifier based on the deep features of the data in n source domains to obtain an auxiliary domain classifier;

[0075] S4: Input the deep features of the data in the target domain into the auxiliary domain classifier, and output the similarity weight ω between the data in the target domain and the data in n source domains n ;

[0076] S5: Train n classifiers respectively based on the deep features of the data in n source domains to obtain n source classifiers;

[0077] S6: Weightedly combine the n source classifiers based on the similarity weight ω between the data in the target domain and the data in n source domains n to obtain a complementary classifier;

[0078] S7: Input the data in the target domain into the complementary classifier, and output the corresponding predicted fault type label.

[0079] The present invention realizes complete fault prediction for the target domain by learning complementary knowledge in multiple source domains, solves the problem of identifying fault types in the target domain that are not included in a single source domain, enhances the fault diagnosis performance in the cross-condition and cross-machine migration diagnosis tasks of production equipment, and improves the accuracy of open-set fault diagnosis for production equipment. First, a feature extraction model is constructed using the framework of an autoencoder. Each source domain dataset no longer cares about the fault types within the domain, but uses the global differences between source domains as the classification basis for the model. At the same time, the encoder is synchronously optimized through the classification loss of the domain classifier to extract deep features representing inter-domain differences. Then, the target domain data is input into the trained domain classifier to obtain the probability that the fault types in the target domain are also included in the i-th source domain, and an auxiliary domain classifier is constructed based on this. Second, a domain fusion module based on the auxiliary domain classifier is used to reduce the feature distribution difference between the source domain and the target domain. The (preprocessed) source domain data x S and the target domain data x T are used as inputs simultaneously, and the source domain and target domain data are synchronously processed through a shared encoder and decoder, thereby reducing the feature distribution difference between the two. The variational principle of the variational autoencoder is used to optimize the decoder, and the generalization ability of the model is improved by adding noise and KL divergence, enhancing the ability of the model to extract domain-invariant features. In addition, through an adaptive weighted learning strategy based on the domain classifier, the similarity degrees between different target domain data and source domain data can be distinguished, so that the domain adaptation model can extract more important features, avoid negative transfer caused by unique categories in the target domain, and promote positive transfer from the source domain to the target domain. Finally, the present invention uses a decision fusion module based on the auxiliary domain classifier to weight-fuse the classification results of multiple source classifiers using the similarity scores of fault data. Only the samples sharing the same categories as the source domain have high confidence in the classification results using the source classifiers. By adaptively adjusting the contributions of the source domain and target domain data to model training, more refined feature alignment is achieved, and the accuracy of open-set fault diagnosis for production equipment in cross-condition and cross-machine migration scenarios is improved.

[0080] To better introduce the technical solution of the present invention, this embodiment is described through the following several parts.

[0081] Due to the existence of fault types in the target domain that are unknown in the source domain, traditional DA methods directly align the feature distributions of source domain data with those of target domain data, which may lead to negative transfer during model training. To reduce the impact of target domain-specific fault data in the domain adaptation process, the present invention constructs a weighted learning module to assign greater weights to shared fault features and promote positive transfer of domain adaptation. Each model trained with source domain and target domain data can only identify the labeled fault types in the source domain, and the specific fault types will be assigned to the wrong categories in this model. Therefore, weighted fusion can be performed at the decision layer through multiple models to improve the fault diagnosis accuracy of the target domain. As Figure 2 shown, it mainly consists of three parts: 1) data preprocessing; 2) domain adaptation module based on an auxiliary domain classifier; 3) decision fusion module based on an auxiliary domain classifier.

[0082] I. Definition of Target Domain and Source Domain

[0083] In this embodiment, a rolling bearing on a production device is taken as an example for illustration. For the open-set fault diagnosis problem of rolling bearings based on multiple source domains, it is assumed that there are n source domains and a single target domain. The source domain is represented as where x j and y j represent the data and corresponding labels in the source domain respectively, and m represents the number of samples in the source domain. The data in the source domain can come from rolling bearings under different working conditions, and the fault categories they contain may also be different. For example, each source domain contains specific fault categories, which is a more reasonable assumption for the actual working process of rolling bearings. The target domain is represented as q represents the number of samples in the target domain. The present invention considers a more practical unsupervised fault diagnosis scenario, so the data in the target domain lacks corresponding labels, which are the labels predicted by the model. The data in the target domain comes from rolling bearings under working conditions different from those in the source domain, that is, At the same time, the fault categories in the target domain cannot be included in any single source domain, but can be covered by the union of all source domain fault categories, that is,

[0084] Therefore, the problem studied by the present invention can be summarized as: under the condition of by extracting domain-invariant features between the source domain and the target domain and using complementary label information of multiple source domains, the fault diagnosis of the target domain is realized.

[0085] II. Data Preprocessing

[0086] In an actual industrial environment, it is inevitable that there is noise in the collected data. Although CNN can also achieve good performance in a noisy environment, compared with methods such as wavelet transform, CNN is slightly rough in signal processing. Wavelet analysis is not only a powerful signal denoising tool, but also has good feature learning efficiency and interpretability.

[0087] Before inputting the data into the constructed model, the present invention uses Continuous Wavelet Transform (CWT) to preprocess the original time series signal collected from the rolling bearing. First, an equal-length intercept strategy is adopted to segment and sample the original vibration signal, with every 1024 digital signals taken as a sample, and then each sample is converted into a wavelet time-frequency diagram using continuous wavelet transform.

[0088] III. Feature Extractor

[0089] The domain adaptation module adopts a framework based on autoencoder (AE) to design the encoder G f (x j ,θ f ) and the decoder G y (x j ,θ y ). The source domain data x S and the target domain data x T after preprocessing are taken as inputs at the same time, and the source domain and target domain data are synchronously processed through the shared encoder and decoder to reduce the difference in feature distribution between the two.

[0090] Specifically, the feature extractor is constructed through the following steps:

[0091] S201: Construct the encoder G f (x j ,θ f ) and the decoder G y (x j ,θ y ) based on the autoencoder framework; θ f and θ y represent the parameters to be optimized in the encoder and decoder;

[0092] S202: Input the data x j of the target domain and the source domain into the encoder G f (x j ,θ f ) to extract the corresponding deep features;

[0093] The basic components of the encoder include a convolutional layer, a batch normalization layer, and an activation layer. In the convolutional layer, the wavelet time-frequency map is used to extract features through a convolutional kernel of a specified size. The feature values are limited between (0, 1) through batch normalization, and then processed using an activation function to represent the output of the neurons in the (l-1)-th layer, represent the size of the convolutional kernel in the l-th layer, then the output of the l-th layer network can be expressed as:

[0094]

[0095] where: f(·) represents the activation function, M and m represent the number of convolutional kernels and the number of data respectively, is the bias of the l-th layer convolution, represents the size of the convolutional kernel in the l-th layer;

[0096] By stacking these basic components with the pooling layer, a deeper feature extractor can be constructed, thereby converting the original vibration signal of the rolling bearing into a high-dimensional feature representation.

[0097] To align the feature distributions between domains, the present invention uses the variational principle based on the variational autoencoder to optimize the decoder. The basic idea is to improve the generalization ability of the model by adding noise.

[0098] S203: Obtain the posterior distribution p(Z|X) of the deep features through two fully connected layers, and sample from the posterior distribution to obtain the

[0099] The formula for the posterior distribution p(Z|X) is expressed as:

[0100] p(Z|X) = (μ, σ 2 );

[0101] μ = f1(X), σ = f2(X);

[0102]

[0103] where: μ and σ are the mean and variance (also called the noise intensity) obtained from the fully connected network f(·) respectively, ε is randomly sampled from N(0, 1), X is the deep feature, and Z is the feature after distribution transformation;

[0104] S204: Input into the decoder G y (x j , θ y ) for reconstruction to obtain the reconstructed data

[0105] The decoder G y (xj , θ y ) has a structure opposite to that of the encoder and replaces the convolutional layer with a transposed convolutional layer:

[0106]

[0107] where: ε is randomly sampled from N(0, 1), and σ is the noise intensity;

[0108] S205: Based on the original data x j and the reconstructed data as well as the noise intensity σ, calculate the mean squared error loss function and KL divergence to train the encoder G f (x j , θ f ) and the decoder G y (x j , θ y );

[0109] The reconstructed rolling bearing vibration signal should be close to the original vibration signal. The mean squared error loss function is used to measure the error between the reconstructed signal and the original signal. Therefore, the formula of the mean squared error loss function is expressed as:

[0110]

[0111] where: E G represents the reconstruction loss, m S and m T represent the number of data in the source domain and the target domain respectively, θ f , θ y represent the parameters to be optimized of the encoder and the decoder, x j represents the original data, represents the reconstructed data, and S and T represent the datasets of the source domain and the target domain respectively;

[0112] Since the reconstruction process continuously reduces the influence of noise, and the noise intensity σ is calculated by the neural network. Therefore, during the training process, the loss function continuously optimizes the noise intensity σ to make it approach 0, resulting in the weakening of noise and affecting the learning ability of the model. To this end, by adding KL divergence, all posterior distributions p(Z|X j ) are aligned to the standard normal distribution to align the feature distributions of the source domain and the target domain, while ensuring the generalization ability of the model. Therefore, the formula of KL divergence is expressed as:

[0113]

[0114] where: σ j , μ j represent the mean and variance after the distribution transformation.

[0115] S206: Repeat steps S202 and S205 to iteratively train the encoder G f (x j , θ f ) and the decoder G y (x j , θ y ). Use the trained encoder G f (x j , θ f ) as the feature extractor.

[0116] IV. Auxiliary Domain Classifier

[0117] Through the above steps, the feature distribution alignment between the source domain and the target domain can be achieved. However, since the target domain contains fault categories that do not exist in the source domain, this domain adaptation process may lead to negative transfer. For example, the shared categories in the target domain may be finally classified as the specific categories in the target domain through the domain adaptation process. To enhance positive transfer and improve the classification performance, an auxiliary domain classifier is used to discriminate the similarity between the samples in the target domain and the source domain samples, so as to assign a greater weight to the samples with higher similarity in the target domain and reduce the attention to the samples with low similarity.

[0118] Specifically, input the deep features of the data in the target domain into the auxiliary domain classifier, and map the deep features to the domain label space {0, 1, 2,..., n} through the softmax function to obtain the probability that the data in the target domain is assigned to different source domains as the similarity weight between the data in the target domain and the data in each source domain;

[0119] The formula is expressed as:

[0120]

[0121] where: ω n includes n to {ω1, ω2,..., ω n} respectively representing the similarity weights between the target domain samples and the samples in each source domain, and k and b are the weights and biases of the auxiliary domain classifier respectively.

[0122] When training the auxiliary domain classifier, calculate the cross-entropy loss function through the deep features of all the data in the source domain combined with the labels of the corresponding data to optimize the parameters of the auxiliary domain classifier;

[0123] The formula of the cross-entropy loss function of the auxiliary domain classifier is expressed as:

[0124]

[0125] where: θ d represents the parameters to be optimized of the auxiliary domain classifier, d jRepresents a domain label.

[0126] Specifically, the similarity weights obtained by the auxiliary domain classifier are attached to the loss function of the domain adaptation model to obtain the loss function of the open-set adaptive weighted network. That is, when training the encoder G f (x j , θ f ) and the decoder G y (x j , θ y ), the similarity weights ω n between the data in the target domain and the data in each source domain are added to the mean squared error loss function and the KL divergence of the encoder G f (x j , θ f ) and the decoder G y (x j , θ y ) to participate in the training;

[0127] The updated formula of the mean squared error loss function is expressed as:

[0128]

[0129] In the formula: m S and m T respectively represent the number of data in the source domain and the target domain, θ f , θ y represent the parameters to be optimized of the encoder and the decoder, x j represents the original data, represents the reconstructed data, S and T respectively represent the data sets of the source domain and the target domain;

[0130] The updated formula of the KL divergence is expressed as:

[0131]

[0132] In the present invention, by adding the similarity weights to the calculation of the mean squared error loss function and the KL divergence of the encoder and the decoder to participate in the training, an adaptive weighted learning strategy based on a domain classifier is constructed, so that the similarity degree between different target domain data and source domain data can be distinguished, so that the domain adaptation model can extract more important features and promote the positive transfer from the source domain to the target domain.

[0133] V. Source classifier

[0134] In this embodiment, the fault classification of the target domain is realized according to the complementary learning of multiple complementary source domains. First, since the data in all source domains have been fully labeled, a plurality of source classifiers can be established accordingly. That is, after the training of the domain adaptation model is completed, the feature extractor G f (xj , θ f Gradient freezing of ) and then connecting the fault classifier The fault classifier uses the softmax function to obtain the final classification result. Therefore, the formula of the source classifier corresponding to each source domain is expressed as:

[0135]

[0136] In the formula: represents the source classifier of the p-th source domain, θ c represents the parameters to be trained of the source classifier, k c and b c represent the weight and bias of the source classifier respectively.

[0137] When training the source classifier, the cross-entropy loss function is calculated by combining the deep features of the data in the corresponding source domain with the labels of the corresponding data to optimize the parameters of the source classifier;

[0138] The formula of the cross-entropy loss function of the source classifier is expressed as:

[0139]

[0140] In the formula: y j represents the label of the data x j .

[0141] VI. Complementary Classifier

[0142] Combine multiple source classifiers Since only the samples sharing the same class with the source domain have high confidence in the classification results using the source classifier, the calculated similarity weights are used to weigh the classification confidence of different source classifiers, so as to obtain a complementary classifier with higher classification accuracy.

[0143] The calculation formula of the complementary classifier is expressed as:

[0144]

[0145] In the formula: represents the combination of the source classifiers of each source domain, represents the source classifier of the p-th source domain, represents the weight and bias of the source classifier of the p-th source domain.

[0146] Figure 3 The overall diagnosis process of the production equipment open-set fault diagnosis model and method based on multi-source domain weighted transfer (subsequently also called the MSWDAN model) is shown as follows and can be specifically divided into the following 5 steps:

[0147] 1) Data preprocessing: The original time series signal of the rolling bearing is converted into a wavelet time-frequency diagram by continuous wavelet transform.

[0148] 2) Training the auxiliary domain classifier: The auxiliary domain classifier is trained by multiple source domain samples, and the loss function is used to optimize the model parameters.

[0149] 3) Training the domain adaptation model: Each domain adaptation model corresponds to a source classifier. The ultimate goal of the domain adaptation model is to train a G f (x, θ f ) for the classifier that can extract domain-invariant features. During the training process, the auxiliary domain classifier is used to extract more relevant features in the target domain and the source domain. The loss functions E G (θ f , θ y ) and E kl (θ f , θ y ) are used to optimize the model parameters.

[0150] 4) Training the fault classifier: Freeze the parameters of the feature extractor G f (x, θ f ), only use the results of its forward propagation, do not perform backpropagation on it, and input the extracted features into the classifier. The loss function is used to optimize the model parameters.

[0151] 5) Predicting the fault type: In the test stage, the source classifier obtains the final fault diagnosis result through the complementary classifier of weighted fusion decision.

[0152] The optimization objectives during the model training process are summarized as follows:

[0153]

[0154] VII. Experimental description

[0155] To better illustrate the advantages of the technical solution of the present invention, the following experiments are disclosed in this embodiment.

[0156] Two cases are used in this experiment to verify the open-set fault diagnosis model and method for production equipment based on multi-source domain weighted migration (MSWDAN model) proposed by the present invention, including the cross-condition migration diagnosis and cross-machine migration diagnosis cases of rolling bearings.

[0157] Case 1: Cross-condition migration diagnosis case of rolling bearings

[0158] 1. Dataset description

[0159] In the cross-condition migration fault diagnosis experiment of rolling bearings, the dataset used was obtained from the Bearing Data Center of Case Western Reserve University (CWRU). This dataset was collected from a 2-horsepower motor test bench, and single-point faults were introduced into the test bearings through electrical discharge machining. The fault locations included the inner ring, outer ring, and balls. Therefore, there were four health conditions: (1) normal condition (NC), (2) inner ring fault (IF), (3) outer ring fault (OF), and (4) ball fault (BF). The fault diameters were 7 mil, 14 mil, 21 mil, 28 mil, and 40 mil. The vibration data was collected by accelerometers, which were placed at the 12 o'clock positions on the motor drive end and the fan end, and the sampling frequency was 12 kHz. The vibration data was recorded under loads of 0, 1, 2, and 3 horsepower respectively. Therefore, four domains were formed according to these four working conditions, denoted as P1, P2, P3, and P4 respectively. The equal-length truncation strategy was adopted to segment and sample the original vibration signals, with 1024 data points as a sample, and the continuous wavelet transform was used to convert the time-series samples into two-dimensional time-frequency diagrams. Finally, each category in each dataset contained 1000 samples.

[0160] Table 1 Dataset Label Settings

[0161]

[0162] In the migration fault diagnosis task of this experiment, the vibration data with fault diameters of 7 mil, 14 mil, and 21 mil collected at the drive end of the bearing was used, and its fault labels were set as shown in Table 1. According to the similarity of the four loads, P1, P3, and P4 were selected as the source domains, and P2 was selected as the target domain. According to the inconsistent domain categories, six migration fault diagnosis tasks were designed to evaluate the effectiveness of the proposed method, as shown in Table 2. The first three tasks were respectively set with missing values within NC, IF, and OF. To ensure the randomness of the fault missing, the last three tasks were randomly set with missing values on NC, IF, and OF, and two types of faults were missing in each domain.

[0163] Table 2 Cross-Condition Migration Fault Diagnosis Tasks

[0164]

[0165] 2. Comparative Algorithms

[0166] To verify the effectiveness and superiority of the proposed MSWDAN of the present invention, performance comparisons were made through existing mainstream DA diagnosis methods.

[0167] 1) DCNN: A traditional deep convolutional neural network model.

[0168] 2) DANN: A gradient reversal layer is introduced in DANN to implement adversarial training between the feature extractor and the domain discriminator, enabling the feature extractor to extract domain-invariant features.

[0169] 3) DAFD: DAFD uses an autoencoder as the backbone network of the model. By mapping the data to the Hilbert space and combining the maximum mean discrepancy, it reduces the distribution error.

[0170] 4) DIDBN: DIDBN connects the restricted Boltzmann machine with the maximum mean discrepancy based on the deep belief network, capable of obtaining the data distribution characteristics with close connections under different working conditions.

[0171] 5) DATLN: DATLN uses a one-dimensional convolutional neural network and a multi-label classifier to identify known fault categories in the target domain. Through adversarial transfer learning, it constructs a decision boundary for newly emerging fault categories to identify new faults, but it cannot determine which specific fault type the new faults belong to.

[0172] 6) PLCWDA: PLCWDA utilizes the pseudo-label mechanism of the target domain to align the within-class feature space and constructs a decision boundary for unknown classes through the method of weighted adversarial learning.

[0173] 7) MRTN: During the adversarial training process, MRTN reduces the within-class refinement differences between each domain through the smooth labels generated by the complementary classifier and determines the weight values of the complementary classifier according to the adaptive loss.

[0174] Since PLCWDA is an open-set fault diagnosis method based on a single source domain, it can only finally divide the unknown classes in the target domain, but cannot determine which specific fault type the unknown classes belong to. Therefore, in this experiment, without changing the main model of PLCWDA, the output values of each softmax classifier on the unknown classes are set to 0, and then the values of the multi-source classifiers are added together to obtain the final fault diagnosis result, thus realizing the identification of all fault types. In this way, the comparability of PLCWDA is stronger. At the same time, to ensure the fairness of the comparison algorithms, when building the above networks, the depth of the model is the same as that of MSWDAN, and all models are compiled using the python3.9 compilation tool under the pytorch2.3 framework and Anaconda. The specific configuration of the experimental equipment is Window11, Intel(R) Core(TM) i5 CPU, 4GRAM, and 1060-super GPU. All models use the Adam algorithm for parameter optimization, with a learning rate of 0.001 and a batch size of 16.

[0175] 3. Analysis of Experimental Results

[0176] To more comprehensively evaluate the performance of the method proposed in the present invention, this experiment compared two comparison scenarios, respectively denoted as source best and source combination. Source best refers to the best accuracy obtained by separately migrating the fault diagnosis knowledge of each source domain to the target domain; source combination refers to combining the training sets of multiple source domains and then using the single-source domain adaptation method for cross-domain fault diagnosis. Each group of experiments was repeated 5 times and the average value of the accuracy was taken.

[0177] Table 3 Cross-condition migration fault diagnosis results

[0178]

[0179] The comparison results are shown in Table 3. Since in the cross-condition migration fault diagnosis task setting, a single source domain always lacks two fault types compared to the target domain, it is difficult to accurately identify these two missing types when diagnosing faults in the target domain. It can be observed that the average accuracy of single-source domain migration fault diagnosis is around 70%. From the comparison results in the source best scenario, it can be seen that the traditional DCNN method achieved a higher fault diagnosis accuracy compared to the other three domain adaptation methods, with an average accuracy 2.67% higher than that of the best domain adaptation method, DATLN. This shows that when the fault types of the source domain and the target domain are inconsistent in the migration fault diagnosis task, directly adopting the global domain adaptation method will lead to negative transfer, thereby reducing the performance of fault diagnosis. In the source combination scenario, it can be observed that MSWDAN has the highest average fault diagnosis accuracy, reaching more than 95%. Although the comparison algorithms improved the fault diagnosis accuracy by mixing the data sets of all source domains, their average accuracy is about 6% lower than that of MSWDAN. This shows that the weighted domain adaptation strategy adopted in the present invention can reduce the negative transfer in the global domain adaptation process. In the comparison results of the three open-set fault diagnosis methods, it can be seen that compared with the traditional DA method, adopting more detailed feature distribution alignment during the model training process is beneficial to improving the fault diagnosis accuracy. Both PLCWDA and MRTN use the idea of generating pseudo-labels for the target domain to achieve more detailed domain adaptation, but the accuracy of PLCWDA is about 3% lower than that of MRTN. This is because MRTN uses a complementary classifier to generate smooth labels in order to avoid errors in explicit label prediction. However, this method has a weak label prediction performance at the initial stage of model training, resulting in the model possibly continuously learning incorrect knowledge during the training process, thereby degrading the fault diagnosis performance of the model. Although the MSWDAN method proposed in the present invention is slightly rough in terms of within-domain class alignment, it avoids the process of guiding model training by incorrect pseudo-labels.

[0180] Then, the t-sne visualization method was used to display the fault features extracted by the above source combination method and open-set fault diagnosis method in task T1. AsFigure 4 As shown in the visualization results of the five source combination methods and the PLCWDA features, four fault categories, namely OF-7, OF-14, IF-7, and IF-21, show obvious distinguishability and have relatively clear boundaries with other fault categories. However, there are varying degrees of confusion in the feature values of the four fault types, namely IF-14, BF-14, IF-21, and OF-21, resulting in a large deviation in the fault diagnosis results. In the visualization results of the MRTN and the MSWDAN model proposed in the present invention, each fault type shows relatively obvious distinguishing boundaries. Only the feature values of the two fault types, IF-14 and OF-21, have a small amount of overlap. However, it can be seen that the features extracted by the MSWDAN are more compact in the visualization performance, indicating that the proposed method has better ability to extract domain-invariant features in the scenario where the domain categories are inconsistent.

[0181] To further evaluate the effectiveness of the proposed domain fusion module and decision fusion module based on the auxiliary domain classifier, the domain fusion module based on the auxiliary domain classifier was modified into a general domain fusion module to construct an ablation model. That is, the ablation comparison model no longer uses the weighted method in the loss function. Then, the multi-source domain sample combination in the above comparison algorithm was used as the input of the ablation model, and this ablation model is denoted as MSDAN. Similarly, comparative experiments were carried out in the above 6 transfer fault diagnosis tasks, and the experimental results are as Figure 5 shown.

[0182] It can be observed that after replacing the domain fusion module and decision fusion module based on the auxiliary domain classifier proposed in the present invention with the traditional domain adaptation mechanism, the transfer fault diagnosis accuracy rate significantly decreases. Specifically, the average fault diagnosis accuracy rate of the MSWDAN proposed in the present invention is 96.67%, while the average accuracy rate of the modified MSDAN is 90.7%, and the accuracy rate drops by about 7%. Compared with the DCNN without using the adaptation mechanism, the average accuracy rate of the MSDAN also slightly decreases.

[0183] Taking the task with the highest fault diagnosis accuracy rate as an example, a confusion matrix was constructed, as Figure 6 shown. The abscissa and ordinate respectively represent the labels predicted by the model and the true labels. It can be observed that there will be certain deviations in the diagnosis of missing categories by both algorithms. The MSDAN even has a 50% probability of making mistakes when distinguishing fault category 5. However, in contrast, the MSWDAN proposed in the present invention has a significant improvement in the diagnosis performance of missing categories, with only a 12% error rate in fault category 5, and other fault categories can be basically completely identified. This further proves that the proposed domain fusion module and decision fusion module based on the auxiliary domain classifier can reduce the negative transfer generated in the global adaptation process and improve the fault diagnosis accuracy rate.

[0184] Case 2: Diagnostic Case of Rolling Bearing Cross-Machine Migration

[0185] 1. Dataset Description

[0186] In the cross-machine migration diagnosis experiment, the CWRU bearing dataset is used as the source domain, and its motor bearing is a rolling bearing of model 6205. The Paderborn University (PU) bearing dataset is used as the target domain, and the bearing it uses is a rolling bearing of model 6203. The sampling frequency of the PU dataset is 64 kHz, including 6 healthy bearings and 26 faulty bearings. Among the 26 faulty bearings, there are 12 artificially damaged bearings (electrical discharge machining, electric engraving, and drilling) and 14 bearings with accelerated life degradation. The PU dataset contains 3 health states: normal condition (NC), inner race fault (IF), and outer race fault (OF). Select the damage data formed by electrical discharge machining in the PU dataset, and 1000 samples are collected for each of the three health states and divided in the ratio of 7:1:2. 700 samples are used for model training, 100 for validation, and another 200 for testing. Each sample contains 4096 sampling points, and the details are shown in Table 4.

[0187] Table 4 Description of Cross-Machine Migration Dataset

[0188]

[0189] According to the three health states in the PU dataset, 6 cross-machine migration fault diagnosis tasks are also designed in the cross-machine migration diagnosis experiment, as shown in Table 5. In tasks C1, C2, and C3, two fault types are missing in the three CWRU datasets as the source domain. In tasks C4, C5, and C6, one fault type is missing in the three CWRU datasets as the source domain.

[0190] Table 5 Cross-Machine Migration Fault Diagnosis Tasks

[0191]

[0192] 2. Analysis of Experimental Results

[0193] The test accuracies of the 6 cross-machine migration fault diagnosis tasks on the algorithm proposed in the present invention and the comparative algorithm are as Figure 7As shown. The comparison algorithms all use the source combination method as the data input. The comparison results are similar to the experimental results in the cross-condition migration fault diagnosis case of rolling bearings. The MSWDAN algorithm proposed in the present invention has the best diagnostic effect. Among the comparison algorithms, the fault diagnosis results of DCNN and MRTN without using the domain adaptation strategy are comparable. According to the experimental results, the fault diagnosis accuracy in the cross-machine migration fault diagnosis task has decreased to a certain extent. The average accuracy of the MSWDAN algorithm has decreased by 0.82% compared with the cross-condition case, and the average accuracy of DCNN has decreased by 1.47%. This is because the characteristics of the fault signals shown by different types of bearings during operation are quite different. Further analysis reveals that the average accuracy on tasks 4, 5, and 6 is slightly higher than that on tasks 1, 2, and 3. This indicates that in the open-set fault diagnosis task, the number of missing fault categories in the source domain will affect the performance of the domain adaptation method in extracting domain-invariant features.

[0194] In addition, this experiment also conducted a statistical analysis on the calculation times of four models with relatively high fault diagnosis accuracy, namely DCNN, DATLN, MRTN, and the proposed MSWDAN. Table 6 lists the average training time and test time of these four algorithms on tasks 1, 2, 5, and 6 respectively. The training time refers to the time taken for the model to converge, and the test time takes the average test time of a single sample.

[0195] Table 6 Comparative analysis of model calculation times

[0196]

[0197] As observed from Table 6, the training time of the model on Tasks 5 and 6 is longer than that on Tasks 1 and 2. This is mainly because Tasks 5 and 6 only lack one category on each source domain. Therefore, the number of samples in the training set is more than that of Tasks 1 and 2, and the data set loading time and the model training time will also become longer. Since the traditional DCNN method does not have a domain adaptation module and does not need to reduce the distribution error between the source domain and the target domain through methods such as adversarial training, the required computational cost is lower. The training time of the MRTN model is the longest because this method needs to predict pseudo-labels during the domain adaptation process and uses the smoothed pseudo-labels predicted by the multi-source classifier to reduce the prediction error, which significantly increases the computational amount of the model. The training time of the method proposed in the present invention is longer than that of DATLN because an additional domain fusion module based on an auxiliary domain classifier is constructed, and the weighted learning on each fault category increases the computational cost. However, in terms of the test time, the performance of the four algorithms is similar. This is because during the test, the domain fusion module will be frozen. Although the proposed algorithm also adds a weighted decision fusion module, the consumption of this module for computing resources is less and does not increase too much computational cost. The test time of a single sample remains at about 0.03 s, which can meet the application of rolling bearing fault diagnosis.

[0198] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit them. Those of ordinary skill in the art should understand that any modifications or equivalent replacements to the technical solutions of the present invention without departing from the purpose and scope of the present technical solution shall be covered by the scope of the claims of the present invention.

Claims

1. An open-set fault diagnosis model and method for production equipment based on multi-source domain weighted transfer, characterized in that, Including: S1: Obtain the datasets of the target domain and n source domains; the dataset of the target domain includes the data of the target domain; the datasets of the source domains include the data of the source domains and the corresponding labels; S2: Extract features from the data of the target domain and n source domains through a feature extractor to obtain the deep features of the data of the target domain and n source domains; S3: Train a classifier based on the deep features of the data of n source domains to obtain an auxiliary domain classifier; S4: Input the deep features of the data of the target domain into the auxiliary domain classifier, and output the similarity weights between the data of the target domain and the data of n source domains; S5: Train n classifiers respectively based on the deep features of the data of n source domains to obtain n source classifiers; S6: Perform weighted combination on the n source classifiers based on the similarity weights between the data of the target domain and the data of n source domains to obtain a complementary classifier; S7: Input the data of the target domain into the complementary classifier and output the corresponding predicted fault type label.

2. The open-set fault diagnosis model and method for production equipment based on multi-source domain weighted migration according to claim 1, characterized in that: In step S1, convert the time-series signal form data of the target domain and source domains into wavelet time-frequency diagrams through continuous wavelet transform, and then perform the subsequent steps.

3. The production equipment open-set fault diagnosis model and method based on multi-source domain weighted migration according to claim 1, characterized in that: In step S2, construct a feature extractor through the following steps: S201: Construct an encoder G based on the auto - encoding framework f (x j ,θ f ) and a decoder G y (x j ,θ y ); S202: Input the data x of the target domain and the source domain j into the encoder G f (x j , θ f ) to extract the corresponding deep features; S203: Obtain the posterior distribution p(Z|X) of the deep features through two fully connected layers, and sample from the posterior distribution to obtain the one after adding noise S204: Apply input decoder G y (x j , θ y ) for reconstruction to obtain reconstructed data S205: Based on the original data x j and the reconstructed data calculate the mean squared error loss function and KL divergence to train the encoder G f (x j , θ f ) and the decoder G y (x j , θ y ); S206: Repeat steps S202 and S205 to iteratively train the encoder G f (x j , θ f ) and the decoder G y (x j , θ y ). Use the trained encoder G f (x j , θ f ) as the feature extractor.

4. The production equipment open-set fault diagnosis model and method based on multi-source domain weighted migration according to claim 3, characterized in that: In step S205, the formula of the mean square error loss function is expressed as: where: E G represents the reconstruction loss, m S and m T represent the number of data in the source domain and the target domain respectively, θ f and θ y represent the parameters to be optimized for the encoder and decoder, x j represents the original data, represents the reconstructed data, and S and T represent the datasets of the source domain and the target domain respectively; The formula of KL divergence is expressed as: where: σ j , μ j represent the mean and variance after the distribution transformation; σ is the noise intensity.

5. The production equipment open-set fault diagnosis model and method based on multi-source domain weighted migration according to claim 3, characterized in that: In step S3, when training the auxiliary domain classifier, calculate the cross-entropy loss function by combining the deep features of the data of all source domains with the corresponding data labels to optimize the parameters of the auxiliary domain classifier; The formula of the cross-entropy loss function of the auxiliary domain classifier is expressed as: Where: θ d represents the parameter to be optimized for the auxiliary domain classifier, d j represents the domain label.

6. The open-set fault diagnosis model and method for production equipment based on multi-source domain weighted migration according to claim 5, characterized in that: In step S4, input the deep features of the data of the target domain into the auxiliary domain classifier, map the deep features to the domain label space {0, 1, 2,..., n} through the softmax function, and obtain the probabilities that the data of the target domain are assigned to different source domains as the similarity weights between the data of the target domain and the data of each source domain; The formula is expressed as: Where: ω n includes n to {ω1, ω2,..., ω n} respectively represent the similarity weights between the target domain samples and each source domain sample, and k and b are the weights and biases of the auxiliary domain classifier, respectively.

7. The production equipment open-set fault diagnosis model and method based on multi-source domain weighted migration according to claim 6, characterized in that: In step S205, the encoder G is trained f (x, θ f ) and the decoder G y (x, θ y ), the similarity weight ω between the data in the target domain and the data in each source domain is n added to the mean square error loss function and the KL divergence of the encoder G f (x, θ f ) and the decoder G y (x, θ y ) to participate in the training; The formula of the updated mean square error loss function is expressed as: Where: m S and m T respectively represent the number of data in the source domain and the target domain, θ f and θ y represent the parameters to be optimized for the encoder and the decoder, x j represents the original data, represents the reconstructed data, and S and T respectively represent the data sets of the source domain and the target domain; The formula of the updated KL divergence is expressed as: where: σ j , μ j represent the mean and variance after the distribution transformation; σ is the noise intensity.

8. The open-set fault diagnosis model and method for production equipment based on multi-source domain weighted migration according to claim 3, characterized in that: In step S5, the formula of the source classifier corresponding to each source domain is expressed as: In the formula: represents the source classifier of the p-th source domain, and θ c represents the parameters to be trained of the source classifier, k c and b c respectively represent the weight and bias of the source classifier.

9. The production equipment open-set fault diagnosis model and method based on multi-source domain weighted migration according to claim 8, characterized in that: When training the source classifier, calculate the cross-entropy loss function by combining the deep features of the data of the corresponding source domain with the corresponding data labels to optimize the parameters of the source classifier; The formula of the cross-entropy loss function of the source classifier is expressed as: where: y j represents the label of data x j .

10. The open-set fault diagnosis model and method for production equipment based on multi-source domain weighted migration according to claim 1, characterized in that: In step S6, the calculation formula of the complementary classifier is expressed as: In the formula: represents the combination of the source classifiers of each source domain, represents the source classifier of the p-th source domain, represents the weights and biases of the source classifier of the p-th source domain.

Images