A Fault Diagnosis Method Based on Multi-dimensional Knowledge Transfer

Abstract

The present invention discloses a fault diagnosis method based on multi-dimensional knowledge transfer, the method comprising: obtaining a source domain data set and a multi-target domain data set; constructing a dual-branch neural network model based on data knowledge and semantic knowledge guidance, and training it, optimizing the model parameters with the optimization goal of minimizing the total loss function of the dual-branch neural network model; wherein the model comprises M+1 data knowledge guided branch networks and M+1 semantic knowledge guided branch networks; and using the trained dual-branch neural network model to perform fault diagnosis tasks in different target domains. The present invention can solve the negative transfer problem caused by forced alignment distribution in traditional transfer learning, and realize effective knowledge transfer from the source domain to multiple target domains from the perspective of multi-dimensional knowledge fusion, interaction, and driving; it can be applied to industrial operation and maintenance, rapid deployment and migration of industrial fault diagnosis tasks in various transfer learning scenarios, and has good learning effect and generalization ability.

Description

Technical Field

[0001] The present invention relates to the field of industrial fault diagnosis technology, specifically to the field of industrial fault diagnosis technology for rapid deployment and migration of cluster equipment in a multi-target domain migration scenario, and in particular to a fault diagnosis method based on multi-dimensional knowledge migration. Background Art

[0002] In recent years, driven by the rapid development of data acquisition and processing technology, deep learning has received great attention and research in various fields. Deep neural networks have shown their advantages over traditional machine learning algorithms in classification and regression. However, the excellent performance of deep neural networks mainly depends on sufficient labeled data. It also requires that the training samples and test samples follow the same distribution. However, in the actual industrial operation and maintenance process, there are often domain differences between the training data set and the test data set due to the switching of machine operation status. At the same time, in large-scale industrial processes, it is costly to obtain enough labeled training samples. To this end, transfer learning provides a feasible solution to solve the distribution difference by extracting and migrating knowledge from the source domain (original stored data) to the target domain (newly acquired data).

[0003] Existing transfer learning methods learn domain-invariant features through distribution alignment or adversarial learning. However, in practical applications, due to different working condition switching, the collected data may come from different distributions (i.e., different domains), and it is expected to transfer knowledge from a labeled domain to other domains with insufficient labels. To this end, the motivation of multi-target transfer learning is to build a model from the source domain and generalize it to multiple different target domains. Existing transfer learning methods for multi-target domains are not well studied and can be mainly divided into methods based on distribution alignment, methods based on feature decoupling, and methods based on pseudo-label iterative updates.

[0004] Among them, the feature-based method aligns the data distribution of cross-domain data by mapping the data to the domain-invariant space. Although this method can reduce the distance between cross-domain data distributions, it may cause negative transfer due to blindly aligning features. The method based on feature decoupling filters domain-specific information by decoupling the public and private features of the data to achieve indirect alignment of features. However, this method may oversimplify the relationship between public and private features and may weaken the ability to capture some key information. The pseudo-label iterative update method gradually selects the target domain that is most similar to the source domain according to the similarity relationship and performs pseudo-labeling operations in sequence. By pseudo-labeling the unlabeled data in the target domain, it iteratively updates the available training data, expands the labeled data set, and gradually realizes the model update; however, this method usually requires multiple iterative updates and is very dependent on the initial prediction effect of the model. If a large classification error occurs in the early stage of the iteration, it is likely to cause the model to collapse.

[0005] In general, existing methods mainly rely on data knowledge to achieve the goal of distribution alignment. However, in multi-target domain scenarios, as the number of target domains increases, the public information that can be extracted will be further reduced, and there is less supervisory information available compared to multi-source domain transfer learning, which requires higher adaptability in building models. Summary of the invention

[0006] The purpose of the present invention is to provide a fault diagnosis method based on multi-dimensional knowledge transfer to solve the problem that there is less available supervision information in multiple target domains and less public information that can be extracted in traditional transfer learning in the prior art. The present invention adopts a data knowledge and semantic knowledge fusion modeling strategy to construct a dual-branch neural network model, and realizes knowledge transfer and correction from the source domain to multiple target domains through the fusion and interaction of knowledge in different dimensions.

[0007] The objective of the present invention is achieved through the following technical solution: A fault diagnosis method based on multi-dimensional knowledge migration, comprising the following steps:

[0008] (1) Obtain source domain data, multi-target domain data, and the semantic attribute matrices of various faults in the source domain and multi-target domain to construct source domain datasets and multi-target domain datasets;

[0009] (2) constructing a dual-branch neural network model based on data knowledge and semantic knowledge guidance; wherein the dual-branch neural network model includes a data knowledge-guided branch network and a semantic knowledge-guided branch network in the source domain and a data knowledge-guided branch network and a semantic knowledge-guided branch network in M ​​target domains, and the data knowledge-guided branch network in the source domain includes a data feature extractor and fault classifier , the semantic knowledge of the source domain guides the branch network including the attribute feature extractor and attribute classifier , the data knowledge of the mth target domain guides the branch network including the data feature extractor and fault classifier , the semantic knowledge of the mth target domain guides the branch network including the attribute feature extractor and attribute classifier ;

[0010] (3) Using the source domain dataset and the multi-target domain dataset to train the two-branch neural network model, taking minimizing the total loss function of the two-branch neural network model as the optimization goal, optimizing the parameters of the two-branch neural network model until the preset training rounds are reached to obtain the trained two-branch neural network model;

[0011] (4) Use the trained dual-branch neural network model to perform fault diagnosis tasks in different target domains.

[0012] Furthermore, the source domain dataset is represented as ,in They represent the i-th data sample in the source domain dataset and the corresponding fault category label and fault semantic attribute matrix label, respectively. represents the total number of samples in the source domain dataset, Indicates that each sample consists of r-dimensional variable measurements, Representation sample The first dimension measurement value of Representation sample The second dimension measurement value of Representation sample The r-th dimension measurement value, each fault semantic attribute matrix label Each is a 0-1 attribute matrix;

[0013] The multi-target domain dataset includes M target domain datasets in total, and the mth target domain dataset is represented as ,in They represent the i-th data sample in the m-th target domain dataset and the corresponding fault category label and fault semantic attribute matrix label, respectively. represents the total number of samples in the mth target domain dataset, and each fault semantic attribute matrix label Each is a 0-1 attribute matrix.

[0014] Furthermore, the dual-branch neural network model specifically includes:

[0015] The data samples in each domain are input into the dual-branch neural network model. The data samples in each domain enter the corresponding data knowledge-guided branch network and the semantic knowledge-guided branch network respectively. First, in the data knowledge-guided branch network, the data samples pass through the data feature extractor to obtain the corresponding data features. The data features pass through the fault classifier to obtain the corresponding predicted fault category. Then, in the semantic knowledge-guided branch network, the data samples pass through the attribute feature extractor to obtain the corresponding attribute features. The attribute features pass through the attribute classifier to obtain the corresponding predicted fault semantic attribute matrix.

[0016] During the reasoning process of the dual-branch neural network model, the training objective functions of the data knowledge-guided branch network and the semantic knowledge-guided branch network corresponding to the source domain and the M target domains are synchronously determined, the training objective function for realizing cross-domain alignment guided by semantic knowledge, the training objective function for realizing information transfer from semantic knowledge to data knowledge, the training objective function for realizing self-correction of semantic knowledge, and the training objective function for correcting inaccurate semantic knowledge using data knowledge.

[0017] Furthermore, the training objective functions of the data knowledge-guided branch network and the semantic knowledge-guided branch network corresponding to the source domain and the M target domains specifically include:

[0018] The data samples in the source domain dataset are input into the data knowledge-guided branch network and the semantic knowledge-guided branch network corresponding to the source domain in the dual-branch neural network model. First, in the data knowledge-guided branch network, the data samples pass through the data feature extractor to obtain the corresponding data features, which are expressed as:

[0019]

[0020] In the formula, represents the data samples in the source domain dataset, , Represents data samples The corresponding data features are ; The data characteristics After passing through the fault classifier, the corresponding predicted fault category is obtained, which is achieved by maximizing the mutual information between the data features and the fault category label, expressed as:

[0021]

[0022] In the formula, Represents a collection of data features in the source domain and a collection of fault category labels The size of the mutual information between , and Represents the fault category label of a single sample in the source domain and data characteristics The probability distribution of express and The joint probability distribution of express about The conditional probability distribution of ; at the same time, it is required to minimize the mutual information between the data features and the input original data samples, expressed as:

[0023]

[0024] In the formula, Represents a collection of data features in the source domain and the collection of data samples in the source domain dataset The size of the mutual information between Represents a single data sample in the source domain and data characteristics The joint probability distribution of express about The conditional probability distribution of Indicates a The reference normal distribution of ; From the above, the data knowledge corresponding to the source domain guides the training objective function of the branch network, which is expressed as:

[0025]

[0026] In the formula, for The weight coefficient of

[0027] Similarly, the semantic knowledge corresponding to the source domain guides the training objective function of the branch network, which is expressed as:

[0028]

[0029] In the formula, Represents data samples The corresponding attribute characteristics, Represents a collection of attribute features in the source domain and the set of fault semantic attribute matrix labels The size of the mutual information between , , Represents a collection of attribute features in the source domain and the collection of data samples in the source domain dataset The size of the mutual information between for The weight coefficient of

[0030] Similarly, the training objective function of the data knowledge-guided branch network corresponding to the mth target domain is obtained, which is expressed as:

[0031]

[0032] In the formula, Represents the set of data features in the mth target domain and a collection of fault category labels The size of the mutual information between Represents the data sample in the mth target domain dataset The corresponding data features are , , , Represents the set of data features in the mth target domain and the set of data samples in the mth target domain dataset The size of the mutual information between for The weight coefficient of

[0033] Similarly, the training objective function of the semantic knowledge-guided branch network corresponding to the mth target domain is obtained, which is expressed as:

[0034]

[0035] In the formula, Represents the set of attribute features in the mth target domain and the set of fault semantic attribute matrix labels in the source domain The size of the mutual information between Represents the data sample in the mth target domain dataset The corresponding attribute characteristics, Represents the set of attribute features in the mth target domain and the set of data samples in the mth target domain dataset The size of the mutual information between for The weight coefficient of .

[0036] Furthermore, the expression of the training objective function for achieving semantic knowledge-guided cross-domain alignment is:

[0037]

[0038] In the formula, represents the attribute features extracted from the lth target domain. represents the Euclidean distance.

[0039] Furthermore, the training objective function for realizing information transfer from semantic knowledge to data knowledge is specifically determined by the following method:

[0040] Input the data sample X from the source domain or target domain into the corresponding data knowledge guided branch network and semantic knowledge guided branch network in the dual-branch neural network model to extract the corresponding data features and attribute characteristics , calculate attribute features and data characteristics The mutual information between them is used as the training objective function for realizing the information transfer from semantic knowledge to data knowledge, where the mutual information is calculated by calculating the attribute features in the same batch of training samples. and data characteristics The KL divergence loss between the joint distribution sampling and marginal distribution product of these two types of features is obtained, expressed as:

[0041]

[0042] In the formula, Represents attribute characteristics and data characteristics The size of the mutual information between Represents the KL divergence loss between different elements, Represents attribute characteristics and data characteristics The joint distribution of these two types of features is Represents attribute characteristics and data characteristics The product of the marginal distributions of these two categories of features;

[0043] Assuming that the number of training samples in a batch is b, the joint distribution sampling is represented by the attribute characteristics of the same samples in the same batch of training samples and data characteristics The splicing is specifically realized by the neural network T; the marginal distribution product is expressed as the attribute characteristics of the samples in the same batch of training samples The average value of the data features corresponding to other samples The splicing is realized by neural network T;

[0044] Attribute characteristics and data characteristics The mutual information between The mutual information of the neural network T The estimate is obtained and written into the calculation form of the loss function, expressed as:

[0045]

[0046] In the formula, represents the neural network structure, represents the feature concatenation operation, Represents the attribute features corresponding to the i-th data sample, represents the data features corresponding to the i-th data sample, Represents the average value of the data features corresponding to the remaining data samples except the i-th data sample.

[0047] Furthermore, the training objective function for realizing self-correction of semantic knowledge is specifically determined by the following method:

[0048] First, the fault semantic attribute matrix predicted by each target domain is sharpened, which is expressed as:

[0049]

[0050] In the formula, represents the sharpened fault semantic attribute matrix corresponding to the jth data sample in the mth target domain, represents the predicted fault semantic attribute matrix corresponding to the jth data sample in the mth target domain;

[0051] Then, the predicted semantic residual is calculated based on the sharpened fault semantic attribute matrix and the corresponding true fault semantic attribute matrix label in the source domain, which is expressed as:

[0052]

[0053] In the formula, Represents the jth data sample in the mth target domain The corresponding predicted semantic residual, Represents the data samples in the source domain is the real fault semantic attribute matrix label corresponding to the samples of the same fault category;

[0054] Then the calculated prediction semantic residual is judged according to the set residual threshold to determine the final prediction semantic residual size, which is expressed as:

[0055]

[0056] In the formula, is the residual threshold;

[0057] The training objective function for self-correction of semantic knowledge is determined according to the final predicted semantic residual size, and its expression is:

[0058]

[0059]

[0060] In the formula, represents the predicted semantic residual variance in the mth target domain, It represents the average value of the predicted semantic residuals of all training samples in a batch in the mth target domain.

[0061] Furthermore, the training objective function for correcting inaccurate semantic knowledge by using data knowledge is specifically determined by the following method:

[0062] First, the relationship between data samples in the multi-target domain in the data feature space guided by data knowledge is obtained, and its expression is:

[0063]

[0064] In the formula, Represents the i-th data sample in the m-th target domain and the jth data sample The relationship between the data features, Represents the i-th data sample in the m-th target domain The corresponding data features are Represents the jth data sample in the mth target domain The corresponding data features are Indicates the distance threshold between preset data features;

[0065] Then, the metric function of similar samples is determined according to the relationship between data samples in the data feature space guided by data knowledge, which is expressed as:

[0066]

[0067] In the formula, represents the metric function of similar samples in the mth target domain, represents the parameters of the attribute feature extractor of the mth target domain, and the superscript T represents the transpose of a vector or matrix;

[0068] Secondly, the metric function of dissimilar samples is determined according to the relationship between data samples in the data feature space guided by data knowledge, which is expressed as:

[0069]

[0070] In the formula, Represents the metric function of dissimilar samples in the mth target domain;

[0071] Finally, the consistency constraint index is determined according to the measurement function of similar samples and the measurement function of dissimilar samples, and is used as the training objective function to correct inaccurate semantic knowledge using data knowledge, which is expressed as:

[0072]

[0073] In the formula, Represents the consistency constraint indicator of the mth target domain.

[0074] Furthermore, the step (3) specifically includes:

[0075] The data samples in the source domain data set and the data samples in the M target domain data sets are input into the dual-branch neural network model. First, the data samples are sent to the data knowledge guided branch network, and then pass through the data feature extractor and fault classifier of the corresponding domain to obtain the corresponding data features. and , and the predicted fault category; then the data sample is sent to the semantic knowledge guided branch network, and passes through 1 attribute feature extractor of the source domain with shared parameters and M+1 attribute classifiers of the corresponding domains to obtain the corresponding attribute features and , and the predicted fault semantic attribute matrix;

[0076] Then according to the input data sample and its corresponding data features and The basic training loss of the branch network guided by data knowledge is calculated based on the fault category labels corresponding to the source domain dataset and the multi-target domain dataset; according to the input data samples and the corresponding attribute features and and the corresponding fault semantic attribute matrix labels in the source domain dataset and the multi-target domain datasets to calculate the basic training loss of the semantic knowledge-guided branch network; the cross-domain feature alignment loss is calculated according to the attribute features of the data samples in the source domain dataset and the multi-target domain dataset; for each target domain, the information transfer loss from semantic knowledge to data knowledge is calculated according to the data features and attribute features obtained in the domain; for each target domain, the predicted semantic residual constraint and predicted semantic residual variance constraint of each sample are calculated according to the predicted fault semantic attribute matrix in each target domain and the corresponding real fault semantic attribute matrix label in the source domain to obtain the self-correction loss of semantic knowledge; the similarity of samples is calculated in the data feature space according to the data features of samples in different target domains, and the semantic knowledge correction loss guided by data knowledge is calculated based on this; the total loss function of the dual-branch neural network model is calculated according to the basic training loss of the data knowledge-guided branch network, the basic training loss of the semantic knowledge-guided branch network, the cross-domain feature alignment loss, the information transfer loss from semantic knowledge to data knowledge, the self-correction loss of semantic knowledge and the semantic knowledge correction loss guided by data knowledge;

[0077] Taking minimizing the total loss function of the two-branch neural network model as the optimization goal, the SGD optimizer is used to adjust and optimize the parameters of the two-branch neural network model until the preset training rounds are reached to obtain the trained two-branch neural network model.

[0078] Furthermore, the calculation formula of the basic training loss of the data knowledge-guided branch network is:

[0079]

[0080] In the formula, Represents the basic training loss of the branch network guided by data knowledge;

[0081] The calculation formula of the basic training loss of the semantic knowledge-guided branch network is:

[0082]

[0083] In the formula, Represents the basic training loss of the semantic knowledge-guided branch network;

[0084] The calculation formula of the cross-domain feature alignment loss is:

[0085]

[0086] In the formula, represents the cross-domain feature alignment loss;

[0087] The calculation formula for the information transfer loss from semantic knowledge to data knowledge is:

[0088]

[0089] In the formula, represents the information transfer loss from semantic knowledge to data knowledge, and They represent the data features and attribute features corresponding to the i-th data sample extracted by the data knowledge-guided branch network and the semantic knowledge-guided branch network corresponding to the source domain, respectively. It represents the average value of data features corresponding to the remaining samples except the i-th data sample extracted from the branch network guided by the data knowledge corresponding to the source domain. and They represent the data features and attribute features corresponding to the j-th data sample extracted by the data knowledge-guided branch network and the semantic knowledge-guided branch network corresponding to the m-th target domain, respectively. It represents the average value of data features corresponding to the remaining samples except the jth data sample extracted by the branch network guided by the data knowledge corresponding to the mth target domain;

[0090] The calculation formula of the self-correction loss of the semantic knowledge is:

[0091]

[0092] In the formula, A self-correction loss that represents semantic knowledge;

[0093] The calculation formula of the semantic knowledge correction loss guided by the data knowledge is:

[0094]

[0095] In the formula, Representation data knowledge guided semantic knowledge correction loss;

[0096] The calculation formula of the total loss function of the dual-branch neural network model is:

[0097]

[0098] In the formula, Represents the total loss function of the two-branch neural network model.

[0099] Compared with the prior art, the present invention has the following beneficial effects:

[0100] (1) The present invention solves the problem that there is less supervisory information available in multiple target domains during the migration process, and less public information can be extracted, thereby achieving effective knowledge migration from the source domain to multiple target domains;

[0101] (2) This paper proposes for the first time a modeling strategy that integrates data knowledge and semantic knowledge, integrates prior attribute knowledge into the established model, and realizes transfer learning from a multi-dimensional knowledge-driven perspective by utilizing the guidance, commonality, and bidirectionality of knowledge.

[0102] (3) The present invention designs a dual-branch neural network model guided by data knowledge and semantic knowledge, realizes cross-domain knowledge sharing through feature alignment guided by semantic knowledge, and realizes knowledge interaction and correction through an information transfer module from semantic knowledge to data knowledge, a self-correction module of semantic knowledge, and a correction module for inaccurate semantic knowledge based on data knowledge;

[0103] (4) This invention defines a new data knowledge and semantic knowledge fusion interaction strategy, forming a generalizable framework that can realize the migration and correction of cross-domain knowledge, and can be used for the rapid deployment and migration of models in industrial operation and maintenance;

[0104] (5) The present invention solves the problem of negative transfer that is easily caused by forced distribution alignment in traditional transfer learning, and realizes effective knowledge transfer from the source domain to multiple target domains from the perspective of multi-dimensional knowledge fusion, interaction, and driving. The multi-dimensional knowledge-driven modeling idea can be applied to the new generation of industrial operation and maintenance, and industrial fault diagnosis tasks with rapid deployment and migration in various transfer learning scenarios, with good learning effects and generalization capabilities. BRIEF DESCRIPTION OF THE DRAWINGS

[0105] Figure 1 It is a flow chart of the fault diagnosis method based on multi-dimensional knowledge migration of the present invention;

[0106] Figure 2 It is a structural schematic diagram of a dual-branch neural network model of the present invention;

[0107] Figure 3 An architectural flow chart for implementing semantic knowledge-guided cross-domain alignment of the present invention;

[0108] Figure 4 A flow chart of the architecture for transferring information from semantic knowledge to data knowledge of the present invention;

[0109] Figure 5 The architecture flow chart of the present invention for correcting inaccurate semantic knowledge by using data knowledge;

[0110] Figure 6 It is a training flow chart of the dual-branch neural network model of the present invention. DETAILED DESCRIPTION

[0111] Exemplary embodiments will be described in detail herein, examples of which are shown in the accompanying drawings. When the following description refers to the drawings, the same numbers in different drawings represent the same or similar elements unless otherwise indicated. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present invention. Instead, they are merely examples of devices and methods consistent with some aspects of the present invention as detailed in the appended claims.

[0112] The terms used in the present invention are only for the purpose of describing specific embodiments and are not intended to limit the present invention. The singular forms "a", "the" and "the" used in the present invention and the appended claims are also intended to include plural forms unless the context clearly indicates other meanings. It should also be understood that the term "and / or" used herein refers to and includes any or all possible combinations of one or more associated listed items.

[0113] It should be understood that although the terms first, second, third, etc. may be used in the present invention to describe various information, these information should not be limited to these terms. These terms are only used to distinguish the same type of information from each other. For example, without departing from the scope of the present invention, the first information may also be referred to as the second information, and similarly, the second information may also be referred to as the first information. Depending on the context, the word "if" as used herein may be interpreted as "at the time of" or "when" or "in response to determining".

[0114] The present invention is described in detail below in conjunction with the accompanying drawings. In the absence of conflict, the features of the following embodiments and implementations can be combined with each other.

[0115] The following is a detailed description of the fault diagnosis method based on multi-dimensional knowledge transfer of the present invention, taking a data set in a real thermal power plant as an example. The process is as follows: Figure 1 As shown, the method specifically comprises the following steps:

[0116] (1) Obtain source domain data, multi-target domain data, and the semantic attribute matrices of various faults in the source domain and multi-target domains to construct source domain datasets and multi-target domain datasets.

[0117] In this embodiment, the measurement data under different working conditions are obtained, and the measurement data under different working conditions, i.e., the power of thermal power units, are used as different data domains. Among them, the source domain data set contains 2700 samples, including 9 different types of faults, 430 different measurement variables, and the semantic attribute matrix corresponding to each fault category. The target domain uses data sets under different working conditions, including three sub-data sets A, B, and C, i.e., three target domains A, B, and C. Each sub-data set has 90 labeled samples, and the test set has 2700 samples each. According to the source domain thermal power data and multi-target domain thermal power data collected above, the source domain thermal power data set and the target domain thermal power data set are constructed, as shown in Table 1.

[0118] Table 1: Source domain thermal power data set and target domain thermal power data set

[0119]

[0120] Specifically, the source domain dataset is represented as ,in They represent the i-th data sample in the source domain dataset and the corresponding fault category label and fault semantic attribute matrix label, respectively. represents the total number of samples in the source domain dataset, Indicates that each sample consists of r-dimensional variable measurements, Representation sample The first dimension measurement value of Representation sample The second dimension measurement value of Representation sample The r-th dimension measurement value, each fault semantic attribute matrix label is a 0-1 attribute matrix. The multi-target domain dataset contains M target domain datasets, and the mth target domain dataset is represented as ,in They represent the i-th data sample in the m-th target domain dataset and the corresponding fault category label and fault semantic attribute matrix label, respectively. represents the total number of samples in the mth target domain dataset, and each fault semantic attribute matrix label are all 0-1 attribute matrices, but the fault semantic attribute matrix corresponding to the samples in the target domain is unknown during the entire model training phase. During the training process, the labels first use the attribute knowledge of the source domain. The data samples in the above source domain dataset and M target domain datasets are used as input data for the subsequent constructed dual-branch neural network model. In this embodiment, there are 3 target domains, that is, M=3. For the same type of fault, the fault semantic attribute matrix labels in different target domains are different and are also different from the fault semantic attribute matrix labels in the source domain.

[0121] (2) Construct a dual-branch neural network model based on data knowledge and semantic knowledge. The dual-branch neural network model includes a data knowledge-guided branch network and a semantic knowledge-guided branch network in the source domain and a data knowledge-guided branch network and a semantic knowledge-guided branch network in M ​​target domains. The data knowledge-guided branch network in the source domain includes a data feature extractor and fault classifier , the semantic knowledge of the source domain guides the branch network including the attribute feature extractor and attribute classifier , the data knowledge of the mth target domain guides the branch network including the data feature extractor and fault classifier , the semantic knowledge of the mth target domain guides the branch network including the attribute feature extractor and attribute classifier ,like Figure 2 shown.

[0122] Specifically, the data samples in each domain are input into the dual-branch neural network model, and the data samples in each domain enter the corresponding data knowledge guided branch network and the semantic knowledge guided branch network respectively. First, in the data knowledge guided branch network, the data samples pass through the data feature extractor to obtain the corresponding data features, and the data features pass through the fault classifier to obtain the corresponding predicted fault category; then, in the semantic knowledge guided branch network, the data samples pass through the attribute feature extractor to obtain the corresponding attribute features, and the attribute features pass through the attribute classifier to obtain the corresponding predicted fault semantic attribute matrix. In the reasoning process of the dual-branch neural network model, the training objective functions of the data knowledge guided branch network and the semantic knowledge guided branch network corresponding to the source domain and the M target domains are synchronously determined, which are used to realize the training objective function of semantic knowledge guided cross-domain alignment, the training objective function of semantic knowledge to data knowledge information transfer, the training objective function of semantic knowledge self-correction, and the training objective function of using data knowledge to correct inaccurate semantic knowledge. Among them, cross-domain alignment guided by semantic knowledge belongs to the semantic knowledge guided branch network, and the information transfer from semantic knowledge to data knowledge, the self-correction of semantic knowledge, and the use of data knowledge to correct inaccurate semantic knowledge belong to the interactive information content between the data knowledge guided branch network and the semantic knowledge guided branch network.

[0123] Furthermore, the data samples in the source domain data set are input into the data knowledge-guided branch network and the semantic knowledge-guided branch network corresponding to the source domain in the dual-branch neural network model. First, in the data knowledge-guided branch network, the data samples pass through the data feature extractor to obtain the corresponding data features, which are expressed as:

[0124]

[0125] In the formula, represents the data samples in the source domain dataset, , Represents data samples The corresponding data features are The data characteristics After passing through the fault classifier, the corresponding predicted fault category is obtained. Here, we hope to achieve this by maximizing the mutual information between data features and fault category labels, which can be expressed as:

[0126]

[0127] In the formula, Represents a collection of data features in the source domain and a collection of fault category labels The size of the mutual information between , and Represents the fault category label of a single sample in the source domain and data characteristics The probability distribution of express and The joint probability distribution of express about At the same time, in order to ensure the simplicity of data feature representation, it is required to minimize the mutual information between data features and input original data samples, which can be expressed as:

[0128]

[0129] In the formula, Represents a collection of data features in the source domain and the collection of data samples in the source domain dataset The size of the mutual information between Represents a single data sample in the source domain and data characteristics The joint probability distribution of express about The conditional probability distribution of Indicates a Therefore, the training objective function of the branch network guided by the data knowledge corresponding to the source domain can be expressed as:

[0130]

[0131] In the formula, for Then, in the semantic knowledge-guided branch network, the data sample passes through the attribute feature extractor to obtain the corresponding attribute features, which are expressed as:

[0132]

[0133] In the formula, Data Sample The corresponding attribute characteristics, This attribute features After the attribute classifier, the corresponding predicted fault semantic attribute matrix is ​​obtained. Similarly, similar methods and steps as above are used to obtain the training objective function of the semantic knowledge-guided branch network corresponding to the source domain, which is expressed as:

[0134]

[0135] In the formula, Represents a collection of attribute features in the source domain and the set of fault semantic attribute matrix labels The size of the mutual information between , Represents a collection of attribute features in the source domain and the collection of data samples in the source domain dataset The size of the mutual information between for The weight coefficient of .

[0136] Similarly, according to the above steps, the training objective functions of the data knowledge-guided branch network and the semantic knowledge-guided branch network corresponding to the M target domains can be obtained. The training objective function of the data knowledge-guided branch network corresponding to the mth target domain is expressed as:

[0137]

[0138] In the formula, Represents the set of data features in the mth target domain and a collection of fault category labels The size of the mutual information between Represents the data sample in the mth target domain dataset The corresponding data features are , , , Represents the set of data features in the mth target domain and the set of data samples in the mth target domain dataset The size of the mutual information between for The semantic knowledge corresponding to the mth target domain guides the training objective function of the branch network, which is expressed as:

[0139]

[0140] In the formula, Represents the set of attribute features in the mth target domain and the set of fault semantic attribute matrix labels in the source domain The size of the mutual information between Represents the data sample in the mth target domain dataset The corresponding attribute characteristics, Represents the set of attribute features in the mth target domain and the set of data samples in the mth target domain dataset The size of the mutual information between for Here, since the target domain does not know the accurate semantic knowledge, the fault semantic attribute matrix label of the source domain is still used as the supervision information.

[0141] Furthermore, to achieve cross-domain alignment guided by semantic knowledge, the semantic knowledge of M+1 domains is used to guide the branch network, such as Figure 3 As shown. Assume that the relevant parameters of the attribute feature extractors of the source domain and the target domain can be expressed as , and requires these parameters to be shared across domains. On the other hand, it is required that the attribute features extracted from different domains are aligned with each other, that is, and The difference between them is as small as possible, and the Euclidean distance is usually used to quantify it. Therefore, the training objective function for achieving semantic knowledge-guided cross-domain alignment is determined as:

[0142]

[0143] In the formula, represents the attribute features extracted from the lth target domain. represents the Euclidean distance.

[0144] Furthermore, to realize the information transfer from semantic knowledge to data knowledge, the specific implementation method is as follows: Figure 4As shown in the figure, the input data samples are respectively passed through the data knowledge guided branch network and the semantic knowledge guided branch network to obtain the corresponding data features and attribute features, and can achieve cross-domain alignment guided by semantic knowledge. To this end, the transferable information extracted by the semantic knowledge branch is passed to the data knowledge branch by guiding the mutual information of the data space and semantic space features. The data sample X from the source domain or the target domain is input into the corresponding data knowledge guided branch network and the semantic knowledge guided branch network in the dual-branch neural network model to extract the corresponding data features. and attribute characteristics , calculate attribute features and data characteristics The mutual information between them is used as the training objective function for realizing the information transfer from semantic knowledge to data knowledge, where the mutual information can be approximately represented by the KL divergence loss between the joint distribution sampling and the marginal distribution product sampling of the two types of features in the same batch of training samples:

[0145]

[0146] In the formula, Represents attribute characteristics and data characteristics The size of the mutual information between Represents the KL divergence loss between different elements, Represents attribute characteristics and data characteristics The joint distribution between these two types of features, Represents attribute characteristics and data characteristics The product of the marginal distributions of these two types of features. Since the expression and calculation of the mutual information here is very complicated, it can be assisted by the neural network T. Specifically, assuming that the number of training samples in a batch is b, the joint distribution sampling can be expressed as the attribute features of the same samples in the same batch of training samples and data characteristics The marginal distribution product can be expressed as the attribute characteristics of the samples in the same batch of training samples. The average value of the data features corresponding to other samples The attribute characteristics mentioned are and data characteristics The mutual information between The mutual information of the neural network T can be The estimation can be written into the calculation form of the loss function, which is expressed as:

[0147]

[0148]

[0149] In the formula, represents the neural network structure, represents the feature concatenation operation, Represents the attribute features corresponding to the i-th data sample, represents the data features corresponding to the i-th data sample, represents the average value of the data features corresponding to the remaining data samples except the i-th data sample. Further, assuming and They represent the data features and attribute features corresponding to the i-th data sample extracted by the data knowledge-guided branch network and the semantic knowledge-guided branch network corresponding to the source domain, respectively, and Represents the average value of data features corresponding to the remaining samples except the i-th data sample extracted from the branch network guided by the data knowledge corresponding to the source domain; and They represent the data features and attribute features corresponding to the j-th data sample extracted by the data knowledge-guided branch network and the semantic knowledge-guided branch network corresponding to the m-th target domain, respectively. It represents the average value of the data features corresponding to the samples other than the jth data sample extracted by the branch network guided by the data knowledge corresponding to the mth target domain. In this way, the loss of the above neural network T can be Rewritten as:

[0150]

[0151] Furthermore, for the self-correction of semantic knowledge, the self-correction of the semantic knowledge-guided branch network is achieved mainly from two aspects. On the one hand, for each target domain, the residual between the predicted fault semantic attribute matrix and the real fault semantic attribute matrix label of the source domain should be as small as possible. Specifically, the fault semantic attribute matrix predicted by the multi-target domain is first sharpened, and the predicted fault semantic attribute matrix corresponding to the jth data sample in the mth target domain is sharpened as follows:

[0152]

[0153] In the formula, represents the predicted fault semantic attribute matrix corresponding to the jth data sample in the mth target domain, represents the sharpened fault semantic attribute matrix corresponding to the jth data sample in the mth target domain. Then, the predicted semantic residual is calculated based on the sharpened fault semantic attribute matrix and the corresponding true fault semantic attribute matrix label in the source domain, which is expressed as:

[0154]

[0155] In the formula, Represents the jth data sample in the mth target domain The corresponding predicted semantic residual, Represents the data samples in the source domain is the real fault semantic attribute matrix label corresponding to the sample of the same fault category, where the predicted semantic residual is close to 1 and is not included in the loss, indicating inconsistent semantics in the source domain, while the predicted semantic residual is close to 0 and is included in the loss, indicating that it may be consistent semantics in the source domain. The residual threshold is set to , the predicted semantic residual size in the specific calculation is:

[0156]

[0157] Therefore, the specific training objective function can be written as:

[0158]

[0159] In the formula, express The square of the modulus length. On the other hand, the fault semantic attribute matrix of each fault category predicted in the same target domain should be nearly consistent, because under the same working condition, the semantic description of the same type of fault is consistent. Combined with the above-mentioned prediction semantic residual representation, such an objective function can be expressed as the variance of the prediction semantic residual in the same target domain should be as small as possible. The specific training objective function can be written as:

[0160]

[0161] In the formula, represents the predicted semantic residual variance in the mth target domain, It represents the average value of the predicted semantic residuals of all training samples in a batch in the mth target domain.

[0162] Furthermore, data knowledge is used to correct inaccurate semantic knowledge. Specifically, the feature extraction guided by semantic knowledge in multi-target domains is corrected through relatively accurate data knowledge guidance. In multi-target domains, the use of fault semantic attribute knowledge from the source domain may lead to inaccurate feature extraction, while data knowledge uses accurate fault categories as supervision to correct inaccurate semantic knowledge. The specific idea is to use the constraint that sample neighbor relationships must maintain consistency in the semantic knowledge-guided branch network, that is, samples with similar fault categories in the data knowledge-guided branch network should also be similar in the semantic knowledge-guided branch network, and vice versa. The corresponding specific implementation methods are as follows: Figure 5As shown, it is expected that the difference between similar samples is as small as possible, and the difference between dissimilar samples is as large as possible. First, the relationship between data samples in the multi-target domain in the data feature space guided by data knowledge is obtained, and its expression is:

[0163]

[0164] In the formula, Represents the i-th data sample in the m-th target domain and the jth data sample The relationship between the data features of , that is, the relationship between the i-th data sample and the j-th data sample in the m-th target domain in the data feature space guided by data knowledge, Represents the i-th data sample in the m-th target domain The corresponding data features are Represents the jth data sample in the mth target domain The corresponding data features are Represents the preset distance threshold between data features. Then, the metric function of similar samples is determined based on the relationship between data samples in the data feature space guided by data knowledge, which is expressed as:

[0165]

[0166] In the formula, represents the metric function of similar samples in the mth target domain, represents the parameters of the attribute feature extractor of the mth target domain, and the superscript T represents the transpose of a vector or matrix. Secondly, the metric function of dissimilar samples is determined according to the relationship between data samples in the data feature space guided by data knowledge, which is expressed as:

[0167]

[0168] In the formula, Represents the metric function of dissimilar samples in the mth target domain. Finally, the consistency constraint index is determined according to the metric function of similar samples and the metric function of dissimilar samples, and is used as the training objective function for correcting inaccurate semantic knowledge using data knowledge. Since the consistency constraint index expects the metric function value of similar samples to be as small as possible and the metric function of dissimilar samples to be as large as possible, the consistency constraint index is expressed as:

[0169]

[0170] In the formula, Represents the consistency constraint indicator of the mth target domain.

[0171] It should be understood that the use of data knowledge to guide branch networks and semantic knowledge to guide branch networks and the interaction between data knowledge and semantic knowledge can achieve self-correction of semantic knowledge and use data knowledge to correct inaccurate semantic knowledge, which is an interaction between knowledge.

[0172] (3) Use the source domain dataset and the multi-target domain dataset to train the two-branch neural network model, with minimizing the total loss function of the two-branch neural network model as the optimization goal, and optimize the parameters of the two-branch neural network model until the preset training rounds are reached to obtain the trained two-branch neural network model.

[0173] Specifically, Figure 6 As shown in the figure, the data samples in the source domain data set and the data samples in the M target domain data sets are input into the dual-branch neural network model. First, the data samples are sent to the data knowledge guided branch network, and then pass through the data feature extractor and fault classifier of the corresponding domain to obtain the corresponding data features. and , and the predicted fault category; then the data sample is sent to the semantic knowledge guided branch network, and passes through 1 attribute feature extractor of the source domain with shared parameters and M+1 attribute classifiers of the corresponding domains to obtain the corresponding attribute features and , and the predicted fault semantic attribute matrix; then according to the input data sample and its corresponding data features and The basic training loss of the branch network guided by data knowledge is calculated based on the fault category labels corresponding to the source domain dataset and the multi-target domain dataset; according to the input data samples and the corresponding attribute features and The basic training loss of the semantic knowledge-guided branch network is calculated based on the fault semantic attribute matrix labels corresponding to the source domain dataset and the multi-target domain dataset; the cross-domain feature alignment loss is calculated based on the attribute features of the data samples in the source domain dataset and the multi-target domain dataset; for each target domain, the information transfer loss from semantic knowledge to data knowledge is calculated based on the data features and attribute features obtained in the domain; for each target domain, the predicted semantic residual constraint and predicted semantic residual variance constraint of each sample are calculated based on the predicted fault semantic attribute matrix in each target domain and the corresponding real fault semantic attribute matrix label in the source domain to obtain the self-correction loss of semantic knowledge; according to the samples in different target domains, the self-correction loss of semantic knowledge is calculated based on the predicted fault semantic attribute matrix in each target domain and the corresponding real fault semantic attribute matrix label in the source domain. The similarity of samples is calculated in the data feature space based on the data features, and the semantic knowledge correction loss guided by data knowledge is calculated based on this; the total loss function of the two-branch neural network model is calculated according to the basic training loss of the branch network guided by data knowledge, the basic training loss of the branch network guided by semantic knowledge, the cross-domain feature alignment loss, the information transfer loss from semantic knowledge to data knowledge, the self-correction loss of semantic knowledge and the semantic knowledge correction loss guided by data knowledge; with minimizing the total loss function of the two-branch neural network model as the optimization goal, the SGD optimizer is used to adjust and optimize the parameters of the two-branch neural network model until the preset training rounds are reached to obtain the trained two-branch neural network model.

[0174] Furthermore, the calculation formula of the basic training loss of the data knowledge-guided branch network is:

[0175]

[0176] In the formula, Represents the base training loss that guides the branch network with data knowledge.

[0177] Furthermore, the calculation formula of the basic training loss of the semantic knowledge-guided branch network is:

[0178]

[0179] In the formula, Represents the base training loss that guides the branch network with semantic knowledge.

[0180] Furthermore, the calculation formula for the cross-domain feature alignment loss is:

[0181]

[0182] In the formula, represents the cross-domain feature alignment loss.

[0183] Furthermore, the calculation formula for the information transfer loss from semantic knowledge to data knowledge is:

[0184]

[0185] In the formula, Represents the information transfer loss from semantic knowledge to data knowledge.

[0186] Furthermore, the calculation formula of the self-correction loss of semantic knowledge is:

[0187]

[0188] In the formula, Self-correction loss representing semantic knowledge.

[0189] Furthermore, the calculation formula of the semantic knowledge correction loss guided by data knowledge is:

[0190]

[0191] In the formula, Representation data knowledge guided semantic knowledge correction loss.

[0192] Furthermore, the calculation formula of the total loss function of the two-branch neural network model is:

[0193]

[0194] In the formula, Represents the total loss function of the two-branch neural network model.

[0195] (4) Use the trained dual-branch neural network model to perform fault diagnosis tasks in different target domains.

[0196] Specifically, the data samples in the target domain to be predicted are input into the data knowledge-guided branch network and the semantic knowledge-guided branch network corresponding to the target domain in the trained dual-branch neural network model to obtain the predicted fault category and fault semantic attribute matrix, and then determine the result of fault diagnosis.

[0197] It should be understood that the method of the present invention can be applied to various industrial objects, and the following verification is carried out using any industrial object as an example.

[0198] For example, a thermal power data set is used for verification. The thermal power data set is divided into four sub-data sets with different working conditions according to different powers. The data sets are used as four different domains. The fault categories between different domains are consistent, but the fault semantic attribute matrices in different domains are partially different due to the differences between domains. They can be divided into cross-domain consistent attribute parts and cross-domain inconsistent attribute parts. The cross-domain consistent fault semantic attributes are shown in Table 2, which specifically shows the same attributes of fault descriptions between different domains; the different attributes between domains are shown in Table 3.

[0199] Table 2: Fault attributes that are consistent across domains

[0200]

[0201] Table 3: Cross-domain inconsistent fault attributes

[0202]

[0203] The 0 / 1 in Table 2 and Table 3 indicates whether the fault is related to the corresponding semantics. If the fault is related to the corresponding semantics, the attribute value is 1, otherwise it is 0.

[0204] The four groups of data sets are used as the source domain A and the target domains A, B, and C in turn. The method of the present invention and six different methods, including the fine-tuning method (Fine-Tuning), the adversarial multi-target domain adaptation method (AMDA), the multi-discriminator multi-target domain adversarial adaptation method (MTDA), the multi-target domain adaptation method based on information theory (ITA), the heterogeneous graph multi-target domain adaptation method (HGDA) and the multi-target domain adaptation method based on graph curriculum learning (CGCT), are used to perform fault diagnosis tasks under different target domain training sample settings. The performance of the method of the present invention and the other six methods are compared, and the average value of the classification accuracy of different categories is used as the evaluation index. The results are shown in Table 4, where D12, D13 and D14 all indicate that the first data set is used as the source domain data set, and the remaining second, third, and fourth data sets are used as the target domain A, B, and C data sets, with a total of 2700 samples for training.

[0205] Table 4: Fault diagnosis performance evaluation indicators of different methods in thermal power data set

[0206]

[0207] As shown in Table 4 above, the performance evaluation index result value represents the classification effect of transfer learning from the source domain to a specific target domain. It can be seen from Table 4 that compared with other methods, the method of the present invention has the highest classification accuracy on average, and compared with the second best method, the classification accuracy is improved by an average of 3.97%.

[0208] Exemplarily, the Tennessee-Eastman process data set is used for verification. The Tennessee-Eastman process data set is divided into three sub-data sets with different working conditions according to different powers, and is used as three different domains. The fault categories between different domains are consistent, but the fault attribute matrices in different domains are partially different due to the differences between domains, and can be divided into cross-domain consistent attribute parts and cross-domain inconsistent attribute parts. Here, the three groups of data sets are used as the source domain A and the target domain A, B in turn, and the method of the present invention and six different methods, namely Fine-Tuning, AMDA, MTDA, ITA, HGDA and CGCT, are used to perform fault diagnosis tasks under different target domain training sample settings. The performance of the method of the present invention and the other six methods is compared, and the average value of the classification accuracy of different categories is used as the evaluation index. The results are shown in Table 5, where D12 and D13 both represent that the first data set is used as the source domain data set, and the remaining second and third data sets are used as target domains A, B, with a total of 2700 samples for training.

[0209] Table 5: Fault diagnosis performance evaluation indicators of different methods on the Tennessee-Eastman process dataset

[0210]

[0211] As shown in Table 5 above, the performance evaluation index result value represents the classification effect of transfer learning from the source domain to a specific target domain. It can be seen from Table 5 that compared with other methods, the method of the present invention has the smallest root mean square error (RMSE) on average, and the classification accuracy is improved by an average of 5.44% compared with the second best method.

[0212] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A fault diagnosis method based on multi-dimensional knowledge transfer, characterized in that: The following steps are involved: (1) Obtain source domain data, multi-target domain data, and semantic attribute matrices of various faults in the source domain and multi-target domain to construct source domain datasets and multi-target domain datasets; The source domain data set and the multi-target domain data set are measurement data under different working conditions; (2) constructing a dual-branch neural network model based on data knowledge and semantic knowledge guidance; wherein the dual-branch neural network model includes a data knowledge-guided branch network and a semantic knowledge-guided branch network in the source domain and a data knowledge-guided branch network and a semantic knowledge-guided branch network in M ​​target domains, and the data knowledge-guided branch network in the source domain includes a data feature extractor and fault classifier The semantic knowledge of the source domain guides the branch network including the attribute feature extractor and attribute classifier The data knowledge of the mth target domain guides the branch network including the data feature extractor and fault classifier The semantic knowledge of the mth target domain guides the branch network including the attribute feature extractor and attribute classifier (3) using the source domain data set and the multi-target domain data set to train the two-branch neural network model, taking minimizing the total loss function of the two-branch neural network model as the optimization goal, optimizing the parameters of the two-branch neural network model until a preset training round is reached to obtain a trained two-branch neural network model; the step (3) specifically includes: The data samples in the source domain data set and the data samples in the M target domain data sets are input into the dual-branch neural network model. First, the data samples are sent to the data knowledge guided branch network, and then pass through the data feature extractor and fault classifier of the corresponding domain to obtain the corresponding data features. and And the predicted fault category; then the data sample is sent to the semantic knowledge guided branch network, and passes through 1 attribute feature extractor of the source domain with shared parameters and M+1 attribute classifiers of the corresponding domains to obtain the corresponding attribute features and and the predicted fault semantic attribute matrix; Then according to the input data sample and its corresponding data features and The basic training loss of the branch network guided by data knowledge is calculated based on the fault category labels corresponding to the source domain dataset and the multi-target domain dataset; according to the input data samples and the corresponding attribute features and and the corresponding fault semantic attribute matrix labels in the source domain dataset and the multi-target domain datasets to calculate the basic training loss of the semantic knowledge-guided branch network; the cross-domain feature alignment loss is calculated according to the attribute features of the data samples in the source domain dataset and the multi-target domain dataset; for each target domain, the information transfer loss from semantic knowledge to data knowledge is calculated according to the data features and attribute features obtained in the domain; for each target domain, the predicted semantic residual constraint and predicted semantic residual variance constraint of each sample are calculated according to the predicted fault semantic attribute matrix in each target domain and the corresponding real fault semantic attribute matrix label in the source domain to obtain the self-correction loss of semantic knowledge; the similarity of samples is calculated in the data feature space according to the data features of samples in different target domains, and the semantic knowledge correction loss guided by data knowledge is calculated based on this; the total loss function of the dual-branch neural network model is calculated according to the basic training loss of the data knowledge-guided branch network, the basic training loss of the semantic knowledge-guided branch network, the cross-domain feature alignment loss, the information transfer loss from semantic knowledge to data knowledge, the self-correction loss of semantic knowledge and the semantic knowledge correction loss guided by data knowledge; Taking minimizing the total loss function of the two-branch neural network model as the optimization goal, the SGD optimizer is used to adjust and optimize the parameters of the two-branch neural network model until the preset training rounds are reached to obtain the trained two-branch neural network model; (4) Use the trained dual-branch neural network model to perform fault diagnosis tasks in different target domains.

2. The fault diagnosis method based on multi-dimensional knowledge transfer according to claim 1 is characterized in that: The source domain dataset is represented as where x s,i ,y s,i , represents the i-th data sample in the source domain dataset and the corresponding fault category label and fault semantic attribute matrix label, respectively. s represents the total number of samples in the source domain dataset, Indicates that each sample consists of r-dimensional variable measurements, Represents sample x s,i The first dimension measurement value of Represents sample x s,i The second dimension measurement value of Represents sample x s,i The r-th dimension measurement value, each fault semantic attribute matrix label Each is an attribute matrix consisting of 0 or 1; The multi-target domain dataset includes M target domain datasets in total, and the mth target domain dataset is represented as in They represent the i-th data sample in the m-th target domain dataset and the corresponding fault category label and fault semantic attribute matrix label, respectively. represents the total number of samples in the mth target domain dataset, and each fault semantic attribute matrix label Each is an attribute matrix consisting of 0 or 1.

3. The fault diagnosis method based on multi-dimensional knowledge transfer according to claim 1 is characterized in that: The dual-branch neural network model specifically includes: The data samples in each domain are input into the dual-branch neural network model. The data samples in each domain enter the corresponding data knowledge-guided branch network and the semantic knowledge-guided branch network respectively. First, in the data knowledge-guided branch network, the data samples pass through the data feature extractor to obtain the corresponding data features. The data features pass through the fault classifier to obtain the corresponding predicted fault category. Then, in the semantic knowledge-guided branch network, the data samples pass through the attribute feature extractor to obtain the corresponding attribute features. The attribute features pass through the attribute classifier to obtain the corresponding predicted fault semantic attribute matrix. During the reasoning process of the dual-branch neural network model, the training objective functions of the data knowledge-guided branch network and the semantic knowledge-guided branch network corresponding to the source domain and the M target domains are synchronously determined, the training objective function for realizing cross-domain alignment guided by semantic knowledge, the training objective function for realizing information transfer from semantic knowledge to data knowledge, the training objective function for realizing self-correction of semantic knowledge, and the training objective function for correcting inaccurate semantic knowledge using data knowledge.

4. The fault diagnosis method based on multi-dimensional knowledge transfer according to claim 3 is characterized in that: The training objective functions of the data knowledge-guided branch network and the semantic knowledge-guided branch network corresponding to the source domain and the M target domains specifically include: The data samples in the source domain dataset are input into the data knowledge-guided branch network and the semantic knowledge-guided branch network corresponding to the source domain in the dual-branch neural network model. First, in the data knowledge-guided branch network, the data samples pass through the data feature extractor to obtain the corresponding data features, which are expressed as: Where, X s represents the data samples in the source domain dataset, Represents data sample X s The corresponding data features are The data characteristics After passing through the fault classifier, the corresponding predicted fault category is obtained, which is achieved by maximizing the mutual information between the data features and the fault category label, expressed as: In the formula, Represents a collection of data features in the source domain and the set of fault category labels Y s The size of the mutual information between p(y s )and Represents the fault category label y of a single sample in the source domain s and data characteristics The probability distribution of Represents y s and The joint probability distribution of Represents y s about The conditional probability distribution of ; at the same time, it is required to minimize the mutual information between the data features and the input original data samples, expressed as: In the formula, Represents a collection of data features in the source domain and the set of data samples X in the source domain dataset s The size of the mutual information between Represents a single data sample x in the source domain s and data characteristics The joint probability distribution of express About x s The conditional probability distribution of Indicates a The reference normal distribution of ; From the above, the data knowledge corresponding to the source domain guides the training objective function of the branch network, which is expressed as: In the formula, for The weight coefficient of Similarly, the semantic knowledge corresponding to the source domain guides the training objective function of the branch network, which is expressed as: In the formula, Represents data sample X s The corresponding attribute characteristics, Represents a collection of attribute features in the source domain and the set of fault semantic attribute matrix labels The size of the mutual information between Represents a collection of attribute features in the source domain and the set of data samples X in the source domain dataset s The size of the mutual information between for The weight coefficient of Similarly, the training objective function of the data knowledge-guided branch network corresponding to the mth target domain is obtained, which is expressed as: In the formula, Represents the set of data features in the mth target domain and the set of fault category labels Y t m The size of the mutual information between Represents the data sample in the mth target domain dataset The corresponding data features are Represents the set of data features in the mth target domain and the set of data samples in the mth target domain dataset The size of the mutual information between for The weight coefficient of Similarly, the training objective function of the semantic knowledge-guided branch network corresponding to the mth target domain is obtained, which is expressed as: In the formula, Represents the set of attribute features in the mth target domain and the set of fault semantic attribute matrix labels in the source domain The size of the mutual information between Represents the data sample in the mth target domain dataset The corresponding attribute characteristics, Represents the set of attribute features in the mth target domain and the set of data samples in the mth target domain dataset The size of the mutual information between for The weight coefficient of .

5. The fault diagnosis method based on multi-dimensional knowledge transfer according to claim 4 is characterized in that: The expression of the training objective function for achieving semantic knowledge-guided cross-domain alignment is: In the formula, represents the attribute features extracted from the lth target domain, ||·|| represents the Euclidean distance, Represents the attribute features corresponding to the data samples in the source domain dataset in the semantic knowledge guided branch network.

6. The fault diagnosis method based on multi-dimensional knowledge transfer according to claim 5 is characterized in that: The training objective function for realizing information transfer from semantic knowledge to data knowledge is specifically determined by the following method: Input the data sample X from the source domain or target domain into the corresponding data knowledge guided branch network and semantic knowledge guided branch network in the dual-branch neural network model to extract the corresponding data feature Z d and attribute feature Z k , calculate the attribute feature Z k and data feature Z d The mutual information between them is used as the training objective function for realizing the information transfer from semantic knowledge to data knowledge, where the mutual information is calculated by calculating the attribute feature Z in the same batch of training samples. k and data feature Z d The KL divergence loss between the joint distribution sampling and marginal distribution product of these two types of features is obtained, expressed as: In the formula, I(Z k ,Z d ) represents the attribute feature Z k and data feature Z d The mutual information between KL () represents the KL divergence loss between different elements, Represents attribute feature Z k and data feature Z d The joint distribution of these two types of features is Represents attribute feature Z k and data feature Z d The product of the marginal distributions of these two categories of features; Assuming that the number of training samples in a batch is b, the joint distribution sampling is represented by the attribute feature Z of the same sample in the same batch of training samples k and data feature Z d The splicing is specifically realized by the neural network T; the marginal distribution product is expressed as the attribute feature Z of the sample in the same batch of training samples k The average value of the data features corresponding to other samples The splicing is realized by neural network T; Then the attribute feature Z k and data feature Z d The mutual information size I(Z k ,Z d ) The mutual information of the neural network T The estimate is obtained and written into the calculation form of the loss function, expressed as: In the formula, T() represents the neural network structure, represents the feature concatenation operation, Represents the attribute features corresponding to the i-th data sample, represents the data features corresponding to the i-th data sample, Represents the average value of the data features corresponding to the remaining data samples except the i-th data sample.

7. The fault diagnosis method based on multi-dimensional knowledge transfer according to claim 6 is characterized in that: The training objective function for realizing self-correction of semantic knowledge is specifically determined by the following method: First, the fault semantic attribute matrix predicted by each target domain is sharpened, which is expressed as: In the formula, represents the sharpened fault semantic attribute matrix corresponding to the jth data sample in the mth target domain, represents the predicted fault semantic attribute matrix corresponding to the jth data sample in the mth target domain; Then, the predicted semantic residual is calculated based on the sharpened fault semantic attribute matrix and the corresponding true fault semantic attribute matrix label in the source domain, which is expressed as: In the formula, r m,j Represents the jth data sample in the mth target domain The corresponding predicted semantic residual, Represents the data samples in the source domain is the real fault semantic attribute matrix label corresponding to the samples of the same fault category; Then the calculated prediction semantic residual is judged according to the set residual threshold to determine the final prediction semantic residual size, which is expressed as: Where δ is the residual threshold; The training objective function for self-correction of semantic knowledge is determined according to the final predicted semantic residual size, and its expression is: In the formula, var m represents the predicted semantic residual variance in the mth target domain, It represents the average value of the predicted semantic residuals of all training samples in a batch in the mth target domain.

8. The fault diagnosis method based on multi-dimensional knowledge transfer according to claim 7 is characterized in that: The training objective function for correcting inaccurate semantic knowledge by using data knowledge is specifically determined by the following method: First, the relationship between data samples in the multi-target domain in the data feature space guided by data knowledge is obtained, and its expression is: In the formula, Represents the i-th data sample in the m-th target domain and the jth data sample The relationship between the data features, Represents the i-th data sample in the m-th target domain The corresponding data features are Represents the jth data sample in the mth target domain Corresponding data features, λ represents the preset distance threshold between data features; Then, the metric function of similar samples is determined according to the relationship between data samples in the data feature space guided by data knowledge, which is expressed as: In the formula, represents the metric function of similar samples in the mth target domain, represents the parameters of the attribute feature extractor of the mth target domain, and the superscript T represents the transpose of a vector or matrix; Secondly, the metric function of dissimilar samples is determined according to the relationship between data samples in the data feature space guided by data knowledge, which is expressed as: In the formula, Represents the metric function of dissimilar samples in the mth target domain; Finally, the consistency constraint index is determined according to the measurement function of similar samples and the measurement function of dissimilar samples, and is used as the training objective function to correct inaccurate semantic knowledge using data knowledge, which is expressed as: In the formula, J m Represents the consistency constraint indicator of the mth target domain.

9. The fault diagnosis method based on multi-dimensional knowledge transfer according to claim 8 is characterized in that: The calculation formula of the basic training loss of the data knowledge-guided branch network is: In the formula, L1 represents the basic training loss of the branch network guided by data knowledge; The calculation formula of the basic training loss of the semantic knowledge-guided branch network is: Where L2 represents the basic training loss of the semantic knowledge-guided branch network; The calculation formula of the cross-domain feature alignment loss is: In the formula, L3 represents the cross-domain feature alignment loss; The calculation formula for the information transfer loss from semantic knowledge to data knowledge is: In the formula, L4 represents the information transfer loss from semantic knowledge to data knowledge, and They represent the data features and attribute features corresponding to the i-th data sample extracted by the data knowledge-guided branch network and the semantic knowledge-guided branch network corresponding to the source domain, respectively. It represents the average value of data features corresponding to the remaining samples except the i-th data sample extracted from the branch network guided by the data knowledge corresponding to the source domain. and They represent the data features and attribute features corresponding to the j-th data sample extracted by the data knowledge-guided branch network and the semantic knowledge-guided branch network corresponding to the m-th target domain, respectively. It represents the average value of data features corresponding to the remaining samples except the jth data sample extracted by the branch network guided by the data knowledge corresponding to the mth target domain; The calculation formula of the self-correction loss of the semantic knowledge is: Where L5 represents the self-correction loss of semantic knowledge; The calculation formula of the semantic knowledge correction loss guided by the data knowledge is: Where L6 represents the semantic knowledge correction loss guided by data knowledge; The calculation formula of the total loss function of the dual-branch neural network model is: <h2 style=";text-align:left;direction:ltr">L<h2 style=";text-align:left;direction:ltr"> total <h2 style=";text-align:left;direction:ltr"> (L1+L2+L3+L4+L5+L6) Where, L total Represents the total loss function of the two-branch neural network model.

Images