Unsupervised multi-source information domain adaptive method and device for rotating machine migration diagnosis, medium and product

Abstract

The invention discloses an unsupervised multi-source information domain self-adaption method and device for rotating machine migration diagnosis, a medium and a product, and relates to the field of mechanical fault diagnosis. The method comprises the following steps: acquiring multi-source information; multi-source information is fused and compressed into a single fusion sample; constructing an information fusion enhanced domain adaptive self-attention network transfer learning model; and inputting the single fusion sample into the information fusion enhanced domain adaptive self-attention network transfer learning model to obtain a transfer diagnosis result of the rotating machinery. According to the method, the problems of negative migration, information loss, high calculation burden and the like in a known multi-source information domain adaptive method can be solved.

Description

Technical Field

[0001] The present application relates to the field of mechanical fault diagnosis, and in particular to an unsupervised multi-source information domain adaptive method, device, medium and product for rotating machinery migration diagnosis. Background Art

[0002] Fault diagnosis is considered to be one of the key technologies to ensure the safe and reliable operation of rotating machinery. With the advantage of adaptive feature extraction, deep learning-based fault diagnosis methods have been widely studied in the past decade. However, deep learning-based fault diagnosis methods have not been effectively and widely used in practical industrial applications due to their limitations such as insufficient labeled samples and the requirement that training samples and test samples must follow the same distribution.

[0003] Transfer diagnosis methods based on unsupervised domain adaptation can overcome the problem that deep learning-based fault diagnosis methods cannot accurately diagnose when there are insufficient labeled samples and different sample distributions. Although unsupervised domain adaptation methods have achieved great success in transfer diagnosis, they still have certain limitations. On the one hand, focusing only on the knowledge of a single source domain cannot fully cover the target domain. This limitation hinders the generalization ability of the transfer diagnosis model, especially in dealing with diagnostic tasks involving more complex fault modes and different working conditions. On the other hand, due to the large and complex data, the selection of a suitable single source domain often depends on subjective experience.

[0004] Multi-source domain adaptation helps address the limitations of the above studies by leveraging data from multiple related but different domains. In recent years, multi-source domain adaptation has provided new insights into transfer diagnosis research and greatly improved the diagnostic versatility under different working conditions. The success of multi-source domain adaptation methods in transfer diagnosis can be attributed to the multi-source feature alignment using single-source information (i.e., signals from the same sensor at different speeds). However, relying solely on single-source information still has flaws. The same fault mode may show completely different signal characteristics under different working conditions, which makes single-source information appear more incomplete compared to multi-source information (e.g., multi-source heterogeneous signals such as vibration and electrical signals, and vibration signals from different axes). Therefore, the performance of the transfer diagnosis model will be limited in the case of incomplete information. Although multi-source information domain adaptation can provide deeper insights into fault modes and greatly improve diagnostic performance by effectively utilizing features from multiple sensor sources, it has the following disadvantages:

[0005] 1) There is a problem of negative transfer of feature distribution between multiple source information domains and target domains.

[0006] 2) Implementing feature alignment across multiple source information domains may reduce the diagnostic ability of the transfer learning model, making the transfer diagnosis performance inferior to that of methods based on a single source information domain.

[0007] 3) Current multi-source information domain adaptation methods ignore the spatial relationship between sensors, which may lead to the loss of critical fault information.

[0008] 4) In the information fusion module, with the increase of information sources, the amount of training data and the scale of model structure will increase significantly, which will seriously affect the computational efficiency of migration diagnosis. Summary of the invention

[0009] In order to solve the above-mentioned shortcomings of the prior art, the present application provides an unsupervised multi-source information domain adaptive method, device, medium and product for rotating machinery migration diagnosis.

[0010] To achieve the above objectives, this application provides the following solutions:

[0011] In a first aspect, the present application provides an unsupervised multi-source information domain adaptation method for rotating machinery migration diagnosis, comprising:

[0012] Acquire multi-source information; multi-source information includes time series signals; time series signals are collected by multiple sensors; time series signals cover various health states and working conditions in the information fusion source domain and target domain;

[0013] fusing and compressing the multi-source information into a single fused sample;

[0014] Constructing an information fusion enhanced domain adaptive self-attention network transfer learning model; the constructed information fusion enhanced domain adaptive self-attention network transfer learning model includes a multi-source feature extractor, a domain adaptation module and a classifier;

[0015] The single fusion sample is input into the information fusion enhanced domain adaptive self-attention network transfer learning model to obtain the transfer diagnosis result of the rotating machinery.

[0016] Optionally, fusing and compressing the multi-source information into a single fused sample comprises:

[0017] Applying principal component analysis to process the multi-source information to obtain first three principal components of the principal component analysis result;

[0018] A channel fusion mechanism is designed, and the first three principal components are fused using the designed channel fusion mechanism to obtain the single fused sample.

[0019] Optionally, the designed channel fusion mechanism is expressed as:

[0020]

[0021] In the formula, represents the i-th sample of the first single-source information domain, represents the i-th sample of the second single-source information domain, represents the i-th sample of the third single-source information domain, S1′ represents the first single-source information domain, S2′ represents the second single-source information domain, S3′ represents the third single-source information domain, and R S Represents the information of the red channel value in the image sample, G S Represents the information of the green channel value in the image sample, B S Information representing the blue channel value in the image sample, represents the i-th information fusion sample.

[0022] Optionally, a domain adaptive self-attention network transfer learning model enhanced by information fusion is constructed, including:

[0023] Construct an initial multi-source feature extractor, an initial domain adaptation module, and an initial classifier respectively;

[0024] Acquire multi-sensor data, fuse and compress the multi-sensor data, and obtain information fusion samples of different health conditions in a source domain and information fusion samples of different health conditions in a target domain;

[0025] Divide the information fusion samples of different health conditions in the source domain and the information fusion samples of different health conditions in the target domain into a training set and a test set;

[0026] The initial multi-source feature extractor, the initial domain adaptation module and the initial classifier are jointly trained using the training set, and the model performance is evaluated using the test set; during the training process, the feature distributions of the information fusion source domain and the target domain are gradually aligned through the optimization of the adversarial loss; the training process continues until the initial multi-source feature extractor and the initial domain adaptation module are combined to extract domain invariant features, and the classifier reaches the expected accuracy in the target domain, and then the multi-source feature extractor, the domain adaptation module and the classifier are obtained;

[0027] Based on the combination of the multi-source feature extractor, the domain adaptation module and the classifier, an information fusion enhanced domain adaptive self-attention network transfer learning model is formed.

[0028] Optionally, the multi-source feature extractor is a two-dimensional convolutional neural network based on a parameter-independent self-attention mechanism; the multi-source feature extractor consists of three two-dimensional convolutional layers, three maximum pooling layers, three modules based on a parameter-independent self-attention mechanism and one fully connected layer;

[0029] The parameter-independent self-attention mechanism module is used to generate a feature weight map based on the feature maps output by three two-dimensional convolutional layers and three maximum pooling layers, and to obtain an enhanced feature map based on the feature weight map.

[0030] Optionally, the domain adaptation module includes: a domain discriminator module and a feature distribution difference measurement module;

[0031] The domain discriminator module determines feature weights based on test errors by introducing a data-level weight allocation mechanism; the feature distribution difference measurement module introduces conditional distribution alignment to achieve domain adaptation.

[0032] Optionally, the loss function of the domain discriminator module after the data-level weight allocation mechanism is introduced is expressed as L D (θ f ,θ d ):

[0033]

[0034] In the formula, represents the normalized weight of the i-th sample in the source domain, D(*) represents the domain discriminator, represents the characteristics of the i-th sample in the source domain, represents the feature of the i-th sample in the target domain, θ d represents the training parameters of the domain discriminator module D, represents the number of batch samples in the source domain, Indicates the number of batch samples in the target domain;

[0035] The loss function of the feature distribution difference measurement module that introduces conditional distribution alignment is expressed as L J (θ f ,θ c ):

[0036]

[0037] In the formula, represents the predicted label of the i-th sample in the source domain, represents the predicted label of the jth sample in the target domain, k(·) represents the Gaussian kernel function, represents the predicted label of the source domain, represents the predicted label of the target domain, represents the jth feature of the source domain, represents the i-th feature of the target domain, represents the predicted label of the jth sample in the source domain, represents the predicted label of the i-th sample in the target domain.

[0038] In a second aspect, the present application provides a computer device comprising: a memory, a processor, and a computer program stored on the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of any of the above-described unsupervised multi-source information domain adaptive methods for rotating machinery migration diagnosis.

[0039] In a third aspect, the present application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of any of the above-mentioned unsupervised multi-source information domain adaptation methods for rotating machinery migration diagnosis.

[0040] In a fourth aspect, the present application provides a computer program product, including a computer program, which, when executed by a processor, implements the steps of any of the above-mentioned unsupervised multi-source information domain adaptive methods for rotating machinery migration diagnosis.

[0041] According to the specific embodiments provided in this application, this application has the following technical effects:

[0042] The present application provides an unsupervised multi-source information domain adaptation method, device, medium and product for migration diagnosis of rotating machinery. By fusing and compressing multi-source information into a single fused sample, the fusion of multi-source information is achieved to solve the negative migration problem of feature distribution between the multi-source information domain and the target domain. In addition, by fusing multi-source information formed by time series signals collected by different sensors, the spatial relationship between sensors is taken into account, and the problem of key fault information loss caused by ignoring the spatial relationship between sensors can be solved. By adopting the information fusion-enhanced domain adaptation attention network (IF-EDAAN) transfer learning model, the migration diagnosis result of the rotating machinery is obtained based on a single fused sample, which can effectively reduce the computational burden and improve the fault diagnosis performance, thereby solving the problems of high computational burden and so on in multi-source information domain adaptation. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0044] Figure 1 This is an application environment diagram of an unsupervised multi-source information domain adaptive method for rotating machinery migration diagnosis in an embodiment of the present application;

[0045] Figure 2 A flowchart of an unsupervised multi-source information domain adaptive method for rotating machinery migration diagnosis provided by an embodiment of the present application;

[0046] Figure 3 A schematic diagram of the structure of the information fusion enhanced domain adaptive self-attention network transfer learning model provided in one embodiment of the present application;

[0047] Figure 4 A schematic diagram of the architecture design of a PFAM module provided in another embodiment of the present application;

[0048] Figure 5 An overall implementation flow chart of an unsupervised multi-source information domain adaptation method for rotating machinery migration diagnosis provided by an embodiment of the present application;

[0049] Figure 6 A schematic diagram of convergence curves of different methods on the information fusion target domain test set in a T8→T3 migration diagnosis scenario provided by an embodiment of the present application;

[0050] Figure 7 A schematic diagram of the computing time and memory consumption results of different unsupervised transfer learning methods provided in another embodiment of the present application.

[0051] Figure 8 A schematic diagram of the structure of a computer device provided in one embodiment of the present application. DETAILED DESCRIPTION

[0052] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0053] In order to make the above-mentioned objects, features and advantages of the present application more obvious and easy to understand, the present application is further described in detail below in conjunction with the accompanying drawings and specific implementation methods.

[0054] The unsupervised multi-source information domain adaptive method for rotating machinery migration diagnosis provided in the embodiment of the present application can be applied to Figure 1In the application environment shown. Among them, the terminal 102 communicates with the server 104 through the network. The data storage system can store the data that the server 104 needs to process. The data storage system can be set up separately, integrated on the server 104, or placed on the cloud or other servers. The terminal 102 can send multi-source information to the server 104. After the server 104 receives the multi-source information, for the multi-source information, the server 104 fuses and compresses the multi-source information into a single fused sample. Construct an information fusion enhanced domain adaptive self-attention network transfer learning model. Input the single fusion sample into the information fusion enhanced domain adaptive self-attention network transfer learning model to obtain the migration diagnosis result of the rotating machinery. Among them, the constructed information fusion enhanced domain adaptive self-attention network transfer learning model includes a multi-source feature extractor, a domain adaptation module and a classifier. The server 104 can feedback the obtained migration diagnosis result to the terminal 102. In addition, in some embodiments, the unsupervised multi-source information domain adaptation method for migration diagnosis of rotating machinery can also be implemented independently by the server 104 or the terminal 102. For example, the terminal 102 can directly perform migration diagnosis on the multi-source information, or the server 104 can obtain the multi-source information from the data storage system and perform migration diagnosis on the multi-source information.

[0055] The terminal 102 may be, but is not limited to, various desktop computers, laptop computers, smart phones, tablet computers, IoT devices, and portable wearable devices. The IoT devices may be smart speakers, smart TVs, smart air conditioners, smart vehicle-mounted devices, etc. The portable wearable devices may be smart watches, smart bracelets, head-mounted devices, etc. The server 104 may be implemented as an independent server or a server cluster consisting of multiple servers, or may be a cloud server.

[0056] In an exemplary embodiment, Figure 2 As shown, an unsupervised multi-source information domain adaptive method for rotating machinery migration diagnosis is provided. The method is executed by a computer device, and can be executed by a computer device such as a terminal or a server alone, or by a terminal and a server together. In the embodiment of the present application, the method is applied to Figure 1 The server 104 in the example is used as an example to illustrate, including the following steps 201 to 203. Among them:

[0057] Step 200: Acquire multi-source information. The multi-source information includes time series signals. The time series signals are collected by multiple sensors. The time series signals cover various health states and working conditions of the information fusion source domain and the target domain.

[0058] Step 201: Fuse and compress multi-source information into a single fused sample.

[0059] Step 202: Construct a domain adaptive self-attention network transfer learning model enhanced by information fusion. Figure 3 As shown, the constructed information fusion enhanced domain adaptive self-attention network transfer learning model includes a multi-source feature extractor, a domain adaptation module and a classifier.

[0060] Step 203: Input the single fusion sample into the information fusion enhanced domain adaptive self-attention network transfer learning model to obtain the transfer diagnosis result of the rotating machinery.

[0061] Implementing the above steps 200 to 203 can solve the problems of negative transfer, information loss and high computational burden existing in the known multi-source information domain adaptation methods.

[0062] In another exemplary embodiment of the present application, in order to mine spatial information, it is necessary to fuse and compress multi-source information into a single fused sample. Based on this, in this embodiment, the implementation process of the above step 201 is as follows: First, principal component analysis (PCA) is applied to process multi-source information from different sensors. In order to obtain three RGB (red, green, blue) channel values, only the first three principal components are retained. Then, the enhanced signal is randomly sampled and normalized. Again, information fusion is performed based on the first three principal components to compress multi-source data and enhance fault characteristics.

[0063] In actual application, considering the spatial relationship between multiple sensors, in this embodiment, a channel fusion mechanism is also designed to fuse the first three principal components using the designed channel fusion mechanism to obtain a single fused sample. The designed channel fusion mechanism is expressed as:

[0064]

[0065] In the formula, represents the i-th sample of the first single-source information domain, represents the i-th sample of the second single-source information domain, represents the i-th sample of the third single-source information domain, S1′ represents the first single-source information domain, S2′ represents the second single-source information domain, S3′ represents the third single-source information domain, and R S Represents the information of the red channel value in the image sample, G S Represents the information of the green channel value in the image sample, B S Information representing the blue channel value in the image sample, represents the i-th information fusion sample.

[0066] In another exemplary embodiment of the present application, the process of constructing the information fusion enhanced domain adaptive self-attention network transfer learning model in the above step 202 can be described as:

[0067] (1) An initial multi-source feature extractor, an initial domain adaptation module, and an initial classifier are constructed respectively to initialize the information fusion enhanced domain adaptive self-attention network.

[0068] (2) Acquire multi-sensor data, fuse and compress the multi-sensor data, and obtain information fusion samples of different health conditions in the source domain and information fusion samples of different health conditions in the target domain.

[0069] (3) The information fusion samples of different health conditions in the source domain and the information fusion samples of different health conditions in the target domain are divided into a training set and a test set.

[0070] Among them, the method of fusing and compressing multi-sensor data refers to the fusion method of multi-source information mentioned above. By fusing and compressing multi-sensor data, information fusion samples of different health conditions from the source domain and the target domain can be similarly obtained. On this basis, the multi-source information fusion source domain (domain with labels for training models) and the target domain (domain without labels that require models for prediction) are divided to obtain training sets and test sets, and the obtained training sets and test sets are used as data inputs for the information fusion-enhanced domain adaptation attention network (IF-EDAAN) transfer learning model proposed in the above application.

[0071] (4) The initial multi-source feature extractor, initial domain adaptation module, and initial classifier are jointly trained using the training set, and the model performance is evaluated using the test set. During the training process, the feature distributions of the information fusion source domain and the target domain are gradually aligned through the optimization of the adversarial loss. The training process continues until the initial multi-source feature extractor and the initial domain adaptation module jointly extract domain-invariant features, and the classifier achieves the desired accuracy in the target domain. The resulting multi-source feature extractor, domain adaptation module, and classifier can effectively adapt to the information fusion target domain, achieving seamless feature migration and accurate classification across domains.

[0072] (5) Based on the combination of multi-source feature extractor, domain adaptation module and classifier, a domain adaptive self-attention network transfer learning model with information fusion enhancement is formed.

[0073] In another exemplary embodiment of the present application, the above step (1) is essentially to initialize the information fusion enhanced domain adaptive self-attention network. Based on this, the implementation process can refer to the following steps.

[0074] 1) Construct an initial multi-source feature extractor. In this embodiment, the present application designs a two-dimensional convolutional neural network (CNN) based on a parameter-free attention mechanism (PFAM) as an initial multi-source feature extractor (PFAM-CNN). Specifically, the initial multi-source feature extractor consists of three two-dimensional convolutional layers, three maximum pooling layers, three PFAM modules and a fully connected layer (FC). Figure 3 As shown in the figure, in the multi-source feature extractor finally constructed, Conv1&BatchNorm&ReLU, Conv2&BatchNorm&ReLU and Conv3&BatchNorm&ReLU represent three two-dimensional convolutional layers, MaxPool2D represents the maximum pooling layer, PFAM module represents the PFAM module, and Stride represents the stride. The PFAM module receives the feature map generated by the previous convolutional layer and pooling layer. The architecture design of the PFAM module is shown in the figure. Figure 4 As shown in Figure 2, it is assumed that the feature map (denoted as f) generated by the previous convolutional layer and pooling layer is the input of the PFAM module, where C represents the number of channels, and W and H represent the width and height of the feature map, respectively.

[0075] First, calculate the mean m of each channel c in the feature map f c and variance have:

[0076]

[0077] Among them, f c,i,j Represents the original eigenvalue of channel c at position index (i, j).

[0078] Then, generate the feature weight map, which includes:

[0079]

[0080] where c is the channel index, i∈{1,2,...,H} is the index in the height dimension, j∈{1,2,...,W} is the index in the width dimension, ε is a small constant to avoid the uncertainty of division by zero, and y c,i,j Represents the feature weight of channel c at position index (i, j).

[0081] Finally, we get the enhanced feature map It is described as follows:

[0082] α=σ(y).

[0083]

[0084] Among them, y represents the feature weight map of all channels and spatial dimensions, s(·) represents the sigmoid activation function, and α represents the enhancement coefficient of the feature map f.

[0085] By generating a feature weight map, the two-dimensional PFAM-CNN proposed in this application can effectively weight features of different spatiotemporal dimensions, so that important domain-invariant features receive more attention while suppressing noise and irrelevant features. This adaptive feature weight adjustment method using the PFAM module can effectively optimize the feature results of information fusion, thereby improving the effectiveness and accuracy of domain-invariant feature extraction. Finally, PFAM-CNN extracts the feature F of the information fusion source domain and target domain, as shown below:

[0086] F = Concatenate(F S ,F T )=G((D S ,D T ),θ f ).

[0087] Among them, θ f represents the training parameters of the multi-source feature extractor proposed in this application, D S and D T are the input samples of the information fusion source domain and target domain respectively. S and F T Representative D S and D T The features of , G(·) represents the multi-source feature extractor.

[0088] 2) Construct an initial domain adaptation module. The constructed initial domain adaptation module includes a domain discriminator module D and a feature distribution difference measurement module joint maximum mean discrepancy measurement method (Joint Maximum Mean Discrepancy, JMMD). Among them:

[0089] a) Domain discriminator module.

[0090] Considering the impact of the differences between information fusion source domain samples on transfer diagnosis, this application proposes a data-level weight allocation mechanism. This mechanism assumes that it is difficult for the domain discriminator module D to discriminate samples with high similarity between the information fusion source domain and the target domain, but it is easier to discriminate samples with large differences. This weight allocation mechanism aims to redefine the contribution of each sample in transfer diagnosis. The weight of each sample in the information fusion source domain is determined by the domain discriminator module D based on the domain prediction error. Calculated, among which, represents the characteristics of the i-th sample in the source domain. Based on this, we have:

[0091]

[0092]

[0093] in, and represents the weight and normalized weight of the i-th sample in the information fusion source domain, m S and S Denote the mean and standard deviation of the source domain weight set, respectively. Specifically, the domain discriminator module D consists of three FC layers. During the training process using the source domain dataset, this weight distribution mechanism forces the model to pay more attention to samples similar to the target domain, thereby promoting better migration diagnosis in the target domain.

[0094] After introducing the above sample weight allocation mechanism, the loss function L of the domain discriminator module D is D (θ f ,θ d ) is expressed as follows:

[0095]

[0096] Among them, in the formula, represents the normalized weight of the i-th sample in the source domain, D(*) represents the domain discriminator, represents the characteristics of the i-th sample in the source domain, represents the feature of the i-th sample in the target domain, θ d represents the training parameters of the domain discriminator module D, represents the number of batch samples in the source domain, Represents the number of batch samples of the target domain.

[0097] b) Feature distribution difference measurement module.

[0098] A distance metric-based method, namely JMMD, is used to promote the extraction of domain-invariant spatiotemporal features. This method can maximize and more effectively transfer the knowledge obtained from the information fusion source domain to the information fusion target domain. JMMD introduces conditional distribution alignment to achieve more effective domain adaptation. Based on this, the loss function of the feature distribution difference measurement module that introduces conditional distribution alignment is expressed as:

[0099]

[0100] Among them, in the formula, represents the predicted label of the i-th sample in the source domain, represents the predicted label of the jth sample in the target domain, k(·) represents the Gaussian kernel function, represents the predicted label of the source domain, represents the predicted label of the target domain, represents the jth feature of the source domain, represents the i-th feature of the target domain, represents the predicted label of the jth sample in the source domain, represents the predicted label of the i-th sample in the target domain.

[0101] 3) Classifier.

[0102] For the classifier module composed of FC layers, it can output the diagnostic accuracy of the information fusion source domain and the information fusion target domain, which are:

[0103] P i =Softmax(a i ·F i ′+o i ).

[0104] Among them, P i is the category prediction probability of the information fusion source domain or target domain sample, Softmax(·) represents the softmax classifier, θ c =(a i ,o i ) is the training parameter of the classifier, F i ′ is the feature of the i-th sample from the information fusion source domain or target domain. Given a sample x i The predicted probability P i , its predicted class index in the information fusion source domain or target domain is the corresponding vector P i,j The position index with the highest probability is expressed as follows:

[0105]

[0106] Among them, P i,j Represents sample x i The probability of being predicted as category j in the information fusion source domain or target domain.

[0107] Therefore, the average diagnostic accuracy A is expressed as:

[0108]

[0109] Among them, y i Represents the sample x in the source domain or target domain i The true label, I(·) represents the indicator function, N b Represents the number of samples in a batch in the information fusion source domain or target domain.

[0110] In addition, the classifier loss L of the information fusion source domainC (θ f ,θ c ) is expressed as follows:

[0111]

[0112] in, It is a sample The probability of being predicted as category j in the information fusion source domain, It is a sample The real label. is the total number of categories.

[0113] In another exemplary embodiment of the present application, the initial IF-EDAAN transfer learning model constructed above is trained. This training process includes: training an initial multi-source feature extractor (denoted as G) to extract domain-invariant features. Training an initial domain adaptation module to achieve domain-invariant feature alignment. Training an initial classifier (denoted as C) to achieve the output of the diagnostic accuracy of the information fusion target domain transfer diagnosis training set.

[0114] In the training process of the IF-EDAAN transfer learning model, the goal of domain adaptation is to optimize the initial multi-source feature extractor G to extract domain-invariant spatiotemporal features F, so that the classifier C can achieve accurate migration diagnosis. The loss function of the proposed IF-EDAAN transfer learning model is expressed as L(θ f ,θ c ,θ d ):

[0115] L(θ f ,θ c ,θ d )=L C (θ f ,θ c )-λL D (θ f ,θ d )+μL J (θ f ,θ c ).

[0116]

[0117] Where λ and μ are regularization coefficients, t represents the current iteration time during training, K represents the control change rate, and m represents the maximum value of the regularization coefficient.

[0118] Finally, the parameter θ of the IF-EDAAN transfer learning model f ,θ d and θ c The optimization update will be performed as follows:

[0119]

[0120] Among them, b is the initial learning rate of the model, N p is the total number of training cycles, is the current training cycle number, θ f ′ is the updated parameter θ f ,θ c ′ is the updated parameter θ c ,θ d ′ is the updated parameter θ d ,← represents optimization update, L C is the classifier loss function L of the information fusion source domain C (θ f ,θ c ), L D is the loss function L of the domain discriminator module D (θ f ,θ d ), LJ is the loss function L of the feature distribution difference measurement module that introduces conditional distribution alignment J (θ f ,θ c ). The initial learning rate b can be adjusted during training according to the following formula:

[0121]

[0122] Finally, the trained initial IF-EDAAN transfer learning model is tested. Specifically, after the initial IF-EDAAN transfer learning model is trained, the multi-source information fusion test set of the source domain and the target domain will be input into the trained IF-EDAAN model, and they will only pass through the multi-source feature extractor and classifier modules before the classifier outputs the diagnostic classification results.

[0123] Based on the above description, in the actual application process, the present application first uses multiple sensors to collect the original time series signals of the rotating machinery, covering various health states and working conditions of the information fusion source domain and the target domain. Secondly, information fusion is performed. Before information fusion, the original time series signals collected from multiple sensors are compressed and enhanced using principal component analysis, and then the compressed and enhanced signals are randomly sampled and normalized. The compressed and enhanced samples will be aligned with the three color channels to form image samples based on multi-source information fusion, and the multi-source information fusion source domain and target domain training set and test set are divided. Then, an information fusion enhanced domain adaptive self-attention network transfer learning model is constructed. In the process of constructing the information fusion enhanced domain adaptive self-attention network transfer learning model, it is necessary to construct a multi-source feature extractor, a domain adaptive module and a classifier respectively. Again, the information fusion enhanced domain adaptive self-attention network transfer learning model is trained. The information fusion images of the source domain and the target domain will be input into the multi-source feature extractor. By learning the feature weights, the domain-invariant spatiotemporal features of the information fusion samples will be effectively extracted. Perform domain adaptation and classification. By training the domain discriminator, the feature distribution difference between the information fusion source domain and the information fusion target domain is reduced, further promoting the multi-source feature extractor to extract domain-invariant spatiotemporal features. At the same time, by minimizing the classification loss in the source domain, the accuracy of the classifier in the information fusion source domain can be improved, and excellent migration diagnosis accuracy in the information fusion target domain can be achieved. Finally, the migration diagnosis result is output. The multi-source information fusion test set of the source domain and the target domain will be input into the trained IF-EDAAN transfer learning model, and they will only pass through the multi-source feature extractor and classifier, and the classifier will output the final diagnostic classification result. This application fully considers the impact of different working conditions on diagnosis in industrial scenarios, provides a new method for fault diagnosis, and can provide important technical support for the safe and reliable operation of equipment.

[0124] Further, if Figure 5 In the embodiment shown, it should be noted that in the IF-EDAAN transfer learning model, MK-MMD represents the multi-core maximum mean difference measurement method, JMMD represents the joint maximum mean difference measurement method, S-DANN represents the domain adversarial neural network based on a single source information domain, M-DANN represents the domain adversarial neural network based on multiple source information domains, and DAG-MDAN represents the internal adversarial guided unsupervised multi-domain adaptation network. The diagnostic accuracy of the proposed IF-EDAAN transfer learning model in six different transfer diagnosis scenarios is evaluated. Figure 6The convergence curves of different methods on the information fusion target domain test set in the T8→T3 migration diagnosis scenario are shown. The IF-EDAAN transfer learning model proposed in this application converges the fastest, most stably, and has the highest accuracy of 98.6%. The best accuracy of the DAG-MDAN method is 78.95%, which is still 19.65% lower than the proposed IF-EDAAN transfer learning model. It should be emphasized that the M-DANN method can achieve stable convergence. However, the accuracy of the M-DANN method is poor, with the best accuracy of only 62.05%. These results show that compared with other unsupervised transfer learning methods, the method proposed in this application has stronger stability and superior convergence performance.

[0125] like Figure 7 The calculation time and memory consumption results of different unsupervised transfer learning methods shown. By comparing the calculation time of MK-MMD, JMMD, S-DANN, M-DANN and DAG-MDAN methods with the method proposed in this application, it can be seen that the total calculation time required for this application is the least, only 182.71 seconds. One of the reasons for this result is that information fusion can both compress and enhance multi-sensor vibration signals, so that the IF-EDAAN transfer learning model can process multi-source information simultaneously. Another possible reason is that the efficiency of image processing can be better accelerated. In contrast, the DAG-MDAN method requires the most calculation time, a total of 1300.73 seconds. These findings show that the method provided in this application is more computationally efficient than other unsupervised transfer learning methods. The method proposed in this application requires the least memory, a total of 394.1MB. This is because three vibration signal samples are compressed into one image sample through information fusion. Compared with the results of MK-MMD, JMMD, and S-DANN using single source domain input, their memory requirements are 1425.6MB, which is 1031.5MB more than the method provided in this application. This is because different network structure complexities will result in different memory usage. It should be emphasized that the M-DANN method occupies the most memory, reaching 2253.05MB. These findings show that compared with other unsupervised transfer learning methods, the method proposed in this application has a more optimized network structure.

[0126] Furthermore, in order to verify the diagnostic ability of the method proposed in this application, an ablation experiment was conducted. Table 1 lists the results of ablation experiments under various migration conditions. IF-EDAAN (without PFAM module), IF-DAAN (without weight distribution module), EDAAN (without IF module) and F-EDAAN (without PCA module). The method proposed in this application is superior to the methods of IF-EDAAN (without PFAM module), IF-DAAN (without weight distribution module), EDAAN (without IF module) and F-EDAAN (without PCA module). In different migration scenarios, the average diagnostic accuracy is the highest, reaching 98.88%, which is an increase of 1.89%, 1.97%, 12.81% and 2.93% respectively. Under the condition of changes in speed and current load, the method proposed in this application can significantly improve the diagnostic accuracy, which is reflected in the migration scenarios of T8→T3 and T8→T1. The method provided in this application significantly improves the average diagnostic accuracy, by 29.88% and 39.48% respectively. This improvement is due to the fact that the method provided by this application can significantly expand and improve the diagnostic knowledge in complex migration scenarios. In contrast, in the simple migration scenarios of T5→T6 and T4→T3, the accuracy of the method provided by this application still achieved an increase of 0.48% and 0.56%. By comparing the diagnostic accuracy of the method provided by this application with that of pure IF-DAAN, the diagnostic accuracy of the method proposed in this application in various migration scenarios was improved by 7.56%, 1.2%, 1.48%, 0.08%, 0.4% and 0.6%, respectively. This is because the PFAM module used in this application can capture the correlation between multi-source sensor data and adaptively increase the weight of domain-invariant features, thereby effectively improving the performance of migration diagnosis. The domain adaptation module can improve the stability of training, which can be reflected in the migration scenarios of T8→T3, T8→T1, T7→T1 and T4→T8. In addition, the application of PCA can effectively retain key fault features and remove redundant features, thereby improving diagnostic accuracy.

[0127] Table 1 Ablation experiment results under various migration conditions

[0128]

[0129]

[0130] In summary, compared with the prior art, this application has the following advantages:

[0131] 1. This application proposes an unsupervised multi-source information domain adaptation method, which fuses and compresses multi-source data into a unified representation through principal component analysis and RGB channel fusion, which can effectively enrich the domain invariant features under different working conditions, and improve the performance by mining potential spatial information from multiple sensors through RGB channel fusion, and realize migration diagnosis between information fusion source domain and target domain.

[0132] 2. This application proposes an architecture of the IF-EDAAN transfer learning model, which fuses and compresses multi-source data into a unified two-dimensional representation. This method can effectively eliminate potential conflicts between multi-sensor data and prevent negative transfer problems.

[0133] 3. This application designs the PFAM module in the IF-EDAAN transfer learning model, which can enhance multi-source feature extraction by assigning feature weights at the feature level. In addition, based on the weight assignment mechanism, the transfer diagnosis performance of the target domain can also be improved.

[0134] 4. This application designs a parameter-independent self-attention mechanism module in the multi-source feature extractor, which can redefine the weights of information fusion samples at the feature level, so that important domain-invariant features receive more attention while suppressing noise and irrelevant features.

[0135] 5. This application proposes a domain adaptation module and a joint maximum mean difference measurement strategy, which can effectively extract and align domain-invariant temporal and spatial features, focus on enhancing discriminative features, and improve migration diagnosis performance.

[0136] 6. The two-dimensional data compression and fusion method proposed in this application can also greatly improve the performance utilization of the graphics processing unit (GPU), thereby reducing the computational burden and improving the training efficiency of the proposed transfer learning model.

[0137] In an exemplary embodiment, a computer device is provided. The computer device may be a server or a terminal. The internal structure diagram thereof may be as follows: Figure 8As shown. The computer device includes a processor, a memory, an input / output interface (Input / Output, referred to as I / O) and a communication interface. The processor, the memory and the input / output interface are connected through a system bus, and the communication interface is connected to the system bus through the input / output interface. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store migration diagnostic data. The input / output interface of the computer device is used to exchange information between the processor and an external device. The communication interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, an unsupervised multi-source information domain adaptive method for rotating machinery migration diagnosis is implemented.

[0138] Those skilled in the art will understand that Figure 8 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components. In an exemplary embodiment, a computer device is provided, including a memory and a processor, wherein a computer program is stored in the memory, and the processor implements the steps in the above-mentioned method embodiments when executing the computer program.

[0139] In an exemplary embodiment, a computer-readable storage medium is provided, storing a computer program, and when the computer program is executed by a processor, the steps in the above method embodiments are implemented.

[0140] In an exemplary embodiment, a computer program product is provided, including a computer program, and when the computer program is executed by a processor, the steps in the above method embodiments are implemented.

[0141] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data must comply with relevant regulations.

[0142] Those of ordinary skill in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to the memory, database or other medium used in the embodiments provided in the present application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM may be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM).

[0143] The database involved in each embodiment provided in this application may include at least one of a relational database and a non-relational database. The non-relational database may include a distributed database based on blockchain, etc., but is not limited thereto. The processor involved in each embodiment provided in this application may be a general-purpose processor, a central processing unit, a graphics processor, a digital signal processor, a programmable logic device, a data processing logic device based on quantum computing, etc., but is not limited thereto.

[0144] The technical features of the above embodiments may be arbitrarily combined. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0145] This article uses specific examples to illustrate the principles and implementation methods of this application. The description of the above embodiments is only used to help understand the method and core ideas of this application. At the same time, for those skilled in the art, according to the ideas of this application, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting this application.

Claims

1. An unsupervised multi-source information domain adaptation method for rotating machinery migration diagnosis, characterized in that: The unsupervised multi-source information domain adaptive method for rotating machinery migration diagnosis includes: Acquire multi-source information; the multi-source information includes time series signals; the time series signals are collected by multiple sensors; the time series signals cover various health states and working conditions in the information fusion source domain and the target domain; fusing and compressing the multi-source information into a single fused sample; Constructing an information fusion enhanced domain adaptive self-attention network transfer learning model; the constructed information fusion enhanced domain adaptive self-attention network transfer learning model includes a multi-source feature extractor, a domain adaptation module and a classifier; The single fusion sample is input into the information fusion enhanced domain adaptive self-attention network transfer learning model to obtain the transfer diagnosis result of the rotating machinery.

2. The unsupervised multi-source information domain adaptive method for rotating machinery migration diagnosis according to claim 1, characterized in that: The multi-source information is fused and compressed into a single fused sample, including: Applying principal component analysis to process the multi-source information to obtain first three principal components of the principal component analysis result; A channel fusion mechanism is designed, and the first three principal components are fused using the designed channel fusion mechanism to obtain the single fused sample.

3. The unsupervised multi-source information domain adaptive method for rotating machinery migration diagnosis according to claim 2 is characterized in that: The designed channel fusion mechanism is expressed as: In the formula, represents the i-th sample of the first single-source information domain, represents the i-th sample of the second single-source information domain, represents the i-th sample of the third single-source information domain, S1′ represents the first single-source information domain, S2′ represents the second single-source information domain, S3′ represents the third single-source information domain, and R S Represents the information of the red channel value in the image sample, G S Represents the information of the green channel value in the image sample, B S Information representing the blue channel value in the image sample, represents the i-th information fusion sample.

4. The unsupervised multi-source information domain adaptive method for rotating machinery migration diagnosis according to claim 1, characterized in that: Construct a domain adaptive self-attention network transfer learning model enhanced by information fusion, including: Construct an initial multi-source feature extractor, an initial domain adaptation module, and an initial classifier respectively; Acquire multi-sensor data, fuse and compress the multi-sensor data, and obtain information fusion samples of different health conditions in a source domain and information fusion samples of different health conditions in a target domain; Divide the information fusion samples of different health conditions in the source domain and the information fusion samples of different health conditions in the target domain into a training set and a test set; The initial multi-source feature extractor, the initial domain adaptation module and the initial classifier are jointly trained using the training set, and the model performance is evaluated using the test set; during the training process, the feature distributions of the information fusion source domain and the target domain are gradually aligned through the optimization of the adversarial loss; the training process continues until the initial multi-source feature extractor and the initial domain adaptation module jointly extract domain-invariant features, and the classifier reaches the expected accuracy in the target domain, and then the multi-source feature extractor, the domain adaptation module and the classifier are obtained; Based on the combination of the multi-source feature extractor, the domain adaptation module and the classifier, an information fusion enhanced domain adaptive self-attention network transfer learning model is formed.

5. The unsupervised multi-source information domain adaptive method for rotating machinery migration diagnosis according to claim 1, characterized in that: The multi-source feature extractor is a two-dimensional convolutional neural network based on a parameter-independent self-attention mechanism; the multi-source feature extractor consists of three two-dimensional convolutional layers, three maximum pooling layers, three modules based on a parameter-independent self-attention mechanism and one fully connected layer; The parameter-independent self-attention mechanism module is used to generate a feature weight map based on the feature maps output by three two-dimensional convolutional layers and three maximum pooling layers, and to obtain an enhanced feature map based on the feature weight map.

6. The unsupervised multi-source information domain adaptive method for rotating machinery migration diagnosis according to claim 4, characterized in that: The domain adaptation module includes: a domain discriminator module and a feature distribution difference measurement module; The domain discriminator module determines feature weights based on test errors by introducing a data-level weight allocation mechanism; the feature distribution difference measurement module introduces conditional distribution alignment to achieve domain adaptation.

7. The unsupervised multi-source information domain adaptive method for rotating machinery migration diagnosis according to claim 6, characterized in that: The loss function of the domain discriminator module after introducing the data-level weight allocation mechanism is expressed as L D (θ f ,θ d ): In the formula, represents the normalized weight of the i-th sample in the source domain, D(*) represents the domain discriminator, represents the characteristics of the i-th sample in the source domain, represents the feature of the i-th sample in the target domain, θ d represents the training parameters of the domain discriminator module D, represents the number of batch samples in the source domain, Indicates the number of batch samples in the target domain; The loss function of the feature distribution difference measurement module that introduces conditional distribution alignment is expressed as L J (θ f ,θ c ): In the formula, represents the predicted label of the i-th sample in the source domain, represents the predicted label of the jth sample in the target domain, k(·) represents the Gaussian kernel function, represents the predicted label of the source domain, represents the predicted label of the target domain, represents the jth feature of the source domain, F i T represents the i-th feature of the target domain, represents the predicted label of the jth sample in the source domain, represents the predicted label of the i-th sample in the target domain.

8. A computer device comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the unsupervised multi-source information domain adaptive method for rotating machinery migration diagnosis according to any one of claims 1 to 7.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the unsupervised multi-source information domain adaptive method for rotating machinery migration diagnosis described in any one of claims 1 to 7 is implemented.

10. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the unsupervised multi-source information domain adaptive method for rotating machinery migration diagnosis described in any one of claims 1 to 7 is implemented.

Images