Fault Diagnosis Method Based on Imperfect Information Conditions of Heterogeneous Mapping in Orthogonal Spaces

Abstract

This invention discloses a fault diagnosis method based on orthogonal spatial heterogeneous mapping under imperfect information conditions, relating to the field of fault diagnosis technology for key components of power systems. A hierarchical fusion multidimensional fault feature extraction network based on a cross-domain attention mechanism is proposed to extract features from a complete dataset of fault cases, forming a multidimensional feature vector encoder. For incomplete and missing data, an enhanced orthogonal spatial heterogeneous network embedding multidimensional feature vectors is proposed. By introducing a multidimensional feature vector encoder and an identity mapping residual block, and by integrating fault feature frequencies and a generative adversarial reconstruction loss function, the data reconstruction process is optimized, enhancing data integrity. For missing and incorrect labels, a label self-correction strategy based on spatial differentiation is proposed. This strategy not only considers spatial information feature matching between different fault types and different sensors but also proposes spatially differentiated feature frequency projection, thereby improving the accuracy of data labels.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of fault diagnosis of key components of power systems, and more particularly to a fault diagnosis method under imperfect information based on orthogonal space heterogeneous mapping. BACKGROUND

[0002] As the core component of special equipment, the stability and safety of the key components of the power system are crucial for the operation of key fields such as aviation, military and special industry. However, in the actual operation process, the fault diagnosis technology of the key components of the power system faces many challenges. First, due to factors such as sensor degradation, data acquisition equipment failure or multi-rate sampling, the collected operation data often has missing or incomplete data. This data incompleteness not only increases the complexity of fault diagnosis, but also may cause the failure of traditional diagnosis methods. Secondly, high-quality fault diagnosis models usually rely on a large amount of labeled data, and data labeling requires a lot of expert experience and resources, which is low in efficiency and high in cost. In addition, there are problems of label missing or error in some scenarios, which further limits the promotion and application of traditional data-driven methods.

[0003] In view of the above problems, the rapid development of artificial intelligence technology and big data technology provides a new solution for the fault diagnosis of key components of the power system. In particular, intelligent algorithms based on deep learning, generative adversarial networks and attention mechanisms have shown significant advantages in handling high-dimensional data, incomplete information and complex label correction. These methods can automatically extract deep features from multi-dimensional, multi-source data, and through data reconstruction and label self-correction strategies, reduce the dependence on complete data and accurate labels. The fault diagnosis framework integrating these technologies not only can efficiently deal with data missing, label error and other practical problems, but also can significantly improve the diagnosis accuracy and reliability under complex operating conditions.

[0004] Therefore, how to provide a fault diagnosis method under imperfect information based on orthogonal space heterogeneous mapping is a problem that those skilled in the art need to solve. SUMMARY

[0005] Therefore, the present application provides a fault diagnosis method under imperfect information based on orthogonal space heterogeneous mapping, which constructs an enhanced orthogonal space heterogeneous network embedding multi-dimensional feature vectors through cross-domain attention mechanism and generative adversarial network, effectively reconstructs incomplete information, and intelligently corrects labels through spatial feature information matching and spatial feature frequency differentiation projection, significantly improves the comprehensiveness and accuracy of power system fault diagnosis, and provides strong technical support for the safe operation of special equipment.

[0006] In order to achieve the above purpose, the present application adopts the following technical solutions:

[0007] The fault diagnosis method based on orthogonal space isomorphism mapping under imperfect information condition comprises:

[0008] S1: judging the integrity of the fault data set, which is divided into complete data set, incomplete data set and missing data set;

[0009] S2: inputting the complete data set into the cross-domain attention mechanism network, and inputting the output of the cross-domain attention mechanism network into the hierarchical fusion multi-dimensional fault feature extraction network to form the data feature encoder output after processing;

[0010] S3: inputting the incomplete data set, the missing data set and the output of the data feature encoder into the enhanced orthogonal space isomorphism network of the embedded multi-dimensional generator, efficiently reconstructing the data missing and the incomplete data to obtain the reconstructed complete data set, and the reconstructed complete data set and the complete data set jointly forming the standard data set, and the standard data set is input into the hierarchical fusion multi-dimensional fault feature extraction network based on the cross-domain attention mechanism again to generate data features related to different fault types and sensors;

[0011] S4: performing spatial information position coding on the data features obtained after processing the standard data set to form a spatial relative position coding system of different fault types and sensors;

[0012] S5: a label self-correction strategy based on spatial differentiation is proposed, which matches the spatial information features of different fault types and sensors by using the spatial relative position coding system of different fault types and sensors, and projects the spatial features of the matched same fault type and sensor information; and finally obtains the fault classification result.

[0013] Preferably, the cross-domain attention mechanism network comprises a channel attention network and a spatial attention network in sequence, the channel attention network comprises three parallel branches, and the three parallel branches are spliced to be provided with a convolution layer, an activation function and a convolution layer, the first branch comprises a convolution layer, an asymmetric convolution block, a batch normalization and an activation function, the second branch comprises a convolution layer, two asymmetric convolution blocks, a batch normalization and an activation function, and the third branch comprises a convolution layer and an average pooling layer; the spatial attention network comprises an average pooling layer and a maximum pooling layer, and the average pooling layer and the maximum pooling layer are spliced to be provided with a plurality of asymmetric convolution blocks and a sigmoid activation function in sequence;

[0014] The processing process of the cross-domain attention mechanism network is:

[0015] ;

[0016] wherein, represents the output of the cross-domain attention mechanism network, represents a learnable weight parameter, represents sensor data, represents element-wise multiplication operation, represents ReLU activation function, represents asymmetric convolution block with kernel size k, represents 1x1 convolution operation, represents average pooling operation, represents max pooling operation, represents Sigmoid activation function, represents concatenation operation of features, represents sum operation on all feature channels, represents nonlinear activation function.

[0017] Preferably, the output of the cross-domain attention mechanism network enters the hierarchical fusion multi-dimensional fault feature extraction network, which specifically includes:

[0018] The output of the cross-domain attention mechanism network is divided into four sensor data streams, and each sensor data stream enters the hierarchical fusion multi-dimensional fault feature extraction network, which includes multi-branch convolution, convolution pooling and multi-branch pooling.

[0019] The multi-branch convolution processing process is as follows:

[0020] ;

[0021] wherein, represents the output of the multi-branch convolution block, BN and BP represent data normalization and batch normalization operation of the multi-branch convolution layer, represents max pooling operation, , f = 1, 3, represents f x f standard convolution, represents convolution and, represents asymmetric convolution operation with kernel size 3 and g repetition times;

[0022] The output after convolution pooling is calculated as follows:

[0023] ;

[0024] wherein, is the output after convolution pooling, represents asymmetric convolution operation with kernel size 1 and repetition 3 times;

[0025] The multi-branch pooling processing process is as follows:

[0026] ;

[0027] wherein, is the output of the multi-branch pooling block, represents the average pooling operation repeats .

[0028] Preferably, the enhanced orthogonal space isomerization network embedded with a multi-dimensional generator in S3 includes two parts of a hybrid generator and a regularization discriminator:

[0029] The hybrid generator captures the time series characteristics of the data using a recurrent neural network model, maps the data to the latent space and generates new data through a variational autoencoder, and optimizes the feature representation and network training through the SE module and the identity mapping residual connection.

[0030] Initialization is performed through the hidden state of the recurrent neural network model:

[0031] ;

[0032] wherein, is the initial hidden state of the recurrent neural network model RNN, is the input parameter of the recurrent neural network model, represents the initialization of the weight matrix;

[0033] The recurrent neural network model takes the incomplete data set, the missing data set and the output of the data feature encoder as input, performs recursive processing, and uses the hidden state of the previous time and the latent variable of the current time to calculate the hidden state of the current time :

[0034] ;

[0035] wherein, represents the weight matrix of the hidden layer, which is responsible for the weighted transformation of the hidden state of the previous time, represents the weight matrix of the input layer, which is used to weight the latent variable of the current time ; is the ReLU activation function, is the parameter of the recurrent neural network model, is the bias term;

[0036] Through such recursive calculation, the recurrent neural network model gradually captures the time series characteristics of the data, and finally integrates the time series information through summation operation to obtain as the output result generated by the recurrent neural network model:

[0037] ;

[0038] wherein, N represents the total number of time steps processed by the recurrent neural network for the integrated time series information;

[0039] The data processed by the recurrent neural network model is mapped to the latent space Z through the variational autoencoder, and new data is generated from the latent space using the decoder. The operation process of the variational autoencoder is as follows:

[0040] ;

[0041] wherein, represents the variable and parameter set of the variational autoencoder, is the variable of the latent space, representing the feature representation of the input data in the latent space after encoding, is the parameter set in the variational autoencoder, represents the weight of the encoder part of the variational autoencoder, represents the bias of the encoder part of the variational autoencoder; is the sampling noise generated by the standard normal distribution, represents a nonlinear activation function, represents the weight of the decoder, represents the weight and bias of the decoder, used to decode the latent space variable m to generate new data;

[0042] Through the operation of the variational autoencoder, the mapping from the original feature representation to the latent space and the generation of new data are realized, further mining the data features and increasing the data diversity. The generated data feature map enters the SE module, and the SE module adjusts the importance of each channel using a weighting coefficient through a fully connected layer, so that the model focuses more on key features. The formula is:

[0043] ;

[0044] wherein, represents processing the input feature map , and the feature map adjusted by the channel weight is obtained by calculation, represents the output of the generator at the current moment, represents the parameter set of the SE module, represents the weight matrix of the fully connected layer in the SE module, is the bias term, represents the global average pooling operation on the input feature, and S is the Sigmoid activation function;

[0045] The features processed by the SE module are further optimized through the identity mapping residual connection, and the expression is:

[0046]

[0047] where, represents the identity mapping residual, is the output of the generator at the last time step, which is generated by adding noise and the input feature to generate data, is the parameter corresponding to the network layer; the residual block adds the output of the generator to its input to form a residual, which helps to alleviate the gradient vanishing problem, accelerate network training and optimization, and ensure that the generator learns and generates data stably and efficiently in multiple rounds of calculation.

[0048] In the regularized discriminator, the feature encoder encodes the incomplete data set and the missing data set to obtain the feature representation The final feature encoding representation is obtained by weighted summation:

[0049] ;

[0050] where, represents the feature representation generated by different feature encoders, is the weight coefficient of each feature encoder, indicating the importance of different features in discrimination, is the total number of feature encoders, represents the final feature encoding representation, is the set of all related parameters;

[0051] The regularized discriminator performs gradient calculation on the incomplete data set and the missing data set, and the gradient calculation formula is: ;

[0052] where, represents the gradient of the input data , is the feature representation function in the gradient calculation related operation, is the corresponding parameter;

[0053] The input feature is activated by the sigmoid activation function to obtain the true and false probability:

[0054] :

[0055] where, is the output of the regularized discriminator, representing the probability that the data is real data, S is the Sigmoid activation function, which ensures that the output value is between 0 and 1, is the regularization coefficient, which is used to control the normalization of the gradient, is a hyperparameter used to control the size of the regularization term.

[0056] Preferably, S4 comprises:

[0057] Computing spatial position encoding function with expression:

[0058] ;

[0059] ;

[0060] wherein, denotes the spatial position encoding function, is the Fourier transform feature representation, denotes the weight matrix of the feature, is the sensor feature, is the standard deviation of the Gaussian kernel, is the number of position matrices considered, is the fault type index, is the dimension index, is the total number of vector dimensions;

[0061] On the basis of obtaining the spatial position encoding function result, further calculate the discriminative features between different sensors and different fault types, the discriminative features are calculated by the following formula:

[0062] ;

[0063] wherein, denotes the position encoding input, denotes the multi-layer perception, denotes the Gaussian error linear unit, denotes the sensor state at the previous time;

[0064] And the calculation of different fault types is related to the intermediate variable , first calculate :

[0065] ;

[0066] wherein, denotes the feature alignment vector of the sensor feature and the fault prototype, denotes the sensor spatial feature, denotes the fault prototype feature;

[0067] Then calculate different fault types :

[0068]

[0069] ​The multi-layer perceptron (MLP) here extracts and transforms the input data, and the GELU introduces non-linear characteristics to the calculation results, making it and better distinguish different sensors and fault types.

[0070] Compared with the prior art, the fault diagnosis method based on orthogonal space heterogeneous mapping under imperfect information condition provided by the technical solution has the following beneficial effects:

[0071] Firstly, the application designs an innovative hierarchical fusion multi-dimensional fault feature extraction network based on cross-domain attention mechanism, which can effectively capture multi-dimensional features, including time-frequency domain features, spatial features, and multi-sensor fusion features. Through the joint action of channel and spatial attention mechanism, the network dynamically allocates feature weights, making the model more accurately extract key information related to faults. At the same time, by introducing asymmetric convolution and multi-branch convolution strategy, not only the adaptability of the model to complex fault patterns is improved, but also the consumption of computing resources is significantly reduced, overcoming the problem of feature loss caused by traditional feature extraction methods when dealing with multi-dimensional complex data, thereby improving the reliability and efficiency of diagnosis.

[0072] Secondly, the application proposes an enhanced orthogonal space heterogeneous network embedded with a multi-dimensional generator, which is used for efficient data reconstruction for data missing and incompleteness. The hybrid generator combines the advantages of recurrent neural networks and variational autoencoders, while introducing identity mapping residual blocks to optimize gradient transmission, enhancing the training depth and stability of the model. The regularization discriminator further improves the discrimination ability of the generated data by integrating cross-domain multi-scale feature extraction encoder and gradient penalty mechanism. The design of this generative adversarial network overcomes the problem of unstable training caused by gradient vanishing or excessively deep network structure in traditional adversarial networks, and significantly improves the quality and feature integrity of the generated data, providing more reliable data support for fault diagnosis.

[0073] Thirdly, the application has significant advantages in solving label missing and errors. A label self-repair strategy based on spatial heterogeneous features is proposed, which can accurately capture the spatial feature relationship between sensors and the association characteristics between fault types. On this basis, through self-attention and cross-attention mechanism, a spatial information coding system between sensors and fault types is established. Further, combined with cosine similarity and bidirectional soft allocation strategy, the network can accurately identify and correct error labels, and efficiently fill missing labels. In addition, the spatial differential feature frequency projection network improves the stability and accuracy of label correction by optimizing the feature offset, so that the fault diagnosis can still maintain high performance in complex data environment.

[0074] Finally, the application designs a comprehensive loss optimization strategy, which significantly improves the learning ability and fault diagnosis performance of the model by introducing the synergistic effect of reconstruction loss, matching loss and differential loss. The reconstruction loss optimizes the matching degree of the generated data and the target data in the frequency domain, ensuring the high fidelity of data reconstruction; the matching loss improves the matching efficiency between features through the joint optimization of soft assignment matrix and real matching pair; the differential loss enhances the discrimination ability of feature expression by correcting the offset of feature projection. The innovative design of these loss functions enables the model to more efficiently restore complete data and improve the accuracy and stability of diagnosis.

[0075] Overall, the technical solutions of the application cover the whole process of fault data classification, preprocessing, feature extraction, data enhancement, label self-correction and classification decision, providing an efficient and robust fault diagnosis method. Compared with the prior art, the application has significant technical advantages in data incompleteness processing, feature extraction accuracy and label correction capability. Even in complex application scenarios such as data loss, label error or sensor collection anomaly, the application can still maintain high diagnostic accuracy and robustness. In addition, the application has important practical significance for improving the reliability of equipment operation, reducing operation and maintenance costs, and ensuring safety in industrial scenarios. BRIEF DESCRIPTION OF DRAWINGS

[0076] In order to more clearly illustrate the technical solutions in the embodiments of the application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description only constitute the embodiments of the application, and for those skilled in the art, other drawings can be obtained without creative labor based on the provided drawings.

[0077] Figure 1 The flowchart of the fault diagnosis method under imperfect information condition based on orthogonal space heterogeneous mapping provided by the application;

[0078] Figure 2 The flowchart of the cross-domain attention mechanism network provided by the application;

[0079] Figure 3 The multi-branch convolution structure in the hierarchical fusion multi-dimensional fault feature extraction network provided by the application;

[0080] Figure 4 The convolution pooling structure in the hierarchical fusion multi-dimensional fault feature extraction network provided by the application;

[0081] Figure 5 The multi-branch pooling structure in the hierarchical fusion multi-dimensional fault feature extraction network provided by the application;

[0082] Figure 6 A schematic diagram of an enhanced orthogonal space heterogeneous network embedded with a multi-dimensional generator is provided for the present application.

[0083] Figure 7 A label self-correction network based on spatial differentiation is provided for the present application.

[0084] Figure 8 A key model ablation experiment verification result graph for the method of the present application.

[0085] Figure 9 A learning performance comparison schematic diagram of the method of the present application and an existing incomplete fault information reconstruction and label self-correction method. DETAILED DESCRIPTION

[0086] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0087] The embodiments of the present application disclose a fault diagnosis method under imperfect information condition based on orthogonal space heterogeneous mapping, as shown in Figure 1 , comprising:

[0088] S1: judging the integrity of the fault data set, which is divided into complete data set, incomplete data set and missing data set; the information incomplete data set refers to the case that there are partial missing data in the sensor monitoring data, or the case that part of the labels of the monitoring data are missing. It usually covers the case that part of the data or labels can be obtained, but there are missing fragments due to the limitation of the sensor and other reasons, that is, the information incomplete data is the sensor data or label with partial missing. Correspondingly, the information missing data set refers to the case that the sensor fails to monitor any data, or the case that the labels of the monitoring data are all missing. Compared with the information incomplete, the information missing covers the missing of all data. In the model of the present application, the case that the data and the label are missing at the same time is not considered. Preferably, the data standardization, clipping, splicing and division preprocessing operations are performed on the complete input data of different fault types.

[0089] S2: inputting the complete data set into the cross-domain attention mechanism network, and assigning spatial domain weights and channel domain weights to different sensor data. The output of the cross-domain attention mechanism network is input into the hierarchical fusion multi-dimensional fault feature extraction network, and after processing, the data feature encoder output is formed;

[0090] S3: input the incomplete data set, the missing data set and the output of the data feature encoder into the enhanced orthogonal space heterogeneous network embedded in the multi-dimensional generator, efficiently reconstruct the data missing and incompleteness problem to obtain a reconstructed complete data set. The reconstructed complete data set and the original complete data set jointly constitute a standard data set. The standard data set is input into the cross-domain attention mechanism network and the hierarchical fusion multi-dimensional fault feature extraction network again to generate data features of different fault types and sensors;

[0091] S4: encode the features obtained by processing the standard data set to form a spatial relative position coding system of different fault types and sensors;

[0092] S5: propose a label self-correction strategy based on spatial differentiation, match the spatial information features of different fault types and sensors by using the spatial relative position coding system of different fault types and sensors, project the spatial features of the matched same fault type and sensor information, and finally obtain a fault classification result.

[0093] In this embodiment, the preprocessed complete data set enters the cross-domain attention mechanism network, which allocates weights in the channel and spatial dimensions to emphasize key features and suppress secondary information. As shown in Figure 2 , the structure sequentially processes the input data through the channel attention network and the spatial attention component network. The channel attention network includes three parallel branches, and the three parallel branches are spliced to be provided with a convolution layer, an activation function and a convolution layer. The first branch includes a convolution layer, an asymmetric convolution block, a batch normalization and an activation function. The second branch includes a convolution layer, two asymmetric convolution blocks, a batch normalization and an activation function. The third branch includes a convolution layer and an average pooling layer. The spatial attention network includes an average pooling layer and a maximum pooling layer, and the average pooling layer and the maximum pooling layer are spliced to be sequentially provided with a plurality of asymmetric convolution blocks and a sigmoid activation function;

[0094] The processing process of the cross-domain attention mechanism network is as follows:

[0095] ;

[0096] wherein, represents the output of the cross-domain attention mechanism network, represents a learnable weight parameter, represents sensor data, represents an element-wise multiplication operation, represents a ReLU activation function, represents an asymmetric convolution block with a kernel size of k, represents a 1x1 convolution operation, represents an average pooling operation, represents max-pooling operation, represents Sigmoid activation function, represents concatenation operation of features, represents sum operation on all feature channels, represents nonlinear activation function.

[0097] In order to deeply extract the features of each fault type, the output of the cross-domain attention mechanism network enters the hierarchical fusion multi-dimensional fault feature extraction network, which specifically includes:

[0098] The hierarchical fusion multi-dimensional fault feature extraction network includes multi-branch convolution, convolution pooling and multi-branch pooling;

[0099] The output of the cross-domain attention module is divided into four sensor data streams, and each sensor data stream enters the hierarchical fusion multi-dimensional fault feature extraction network;

[0100] The multi-branch convolution structure diagram is as shown in Figure 3 , and the processing process is:

[0101] ;

[0102] wherein, represents the output of the multi-branch convolution block, BN and BP represent data normalization and batch normalization operation of the multi-branch convolution layer respectively, represents max-pooling operation, , f =1,3, represents f ×f standard convolution, represents convolution and, represents asymmetric convolution operation with kernel size of 3 and g repetition times;

[0103] The convolution pooling structure diagram is as shown in Figure 4 , and the output is calculated as follows:

[0104] ;

[0105] wherein, is the output after convolution pooling, represents asymmetric convolution operation with kernel size of 1 and repetition of 3 times;

[0106] The multi-branch pooling structure diagram is as shown in Figure 5 , and the processing process is:

[0107] ;

[0108] wherein, is the output of the multi-branch pooling block, represents average pooling operation repeated times.

[0109] The data feature encoder formed above will enter the enhanced orthogonal space heterogeneous network embedded in the multi-dimensional generator, introduce an identity mapping residual block, and optimize the data reconstruction process by integrating the fault feature frequency with the reconstruction loss function of the adversarial loss, thereby enhancing the integrity of the data. The detailed reasoning process is shown in Figure 6 .

[0110] The enhanced orthogonal space heterogeneous network embedded in the multi-dimensional generator mainly includes a hybrid generator and a regularization discriminator. The detailed steps of the two parts are as follows:

[0111] The core idea of the hybrid generator is to use a recurrent neural network to capture the time series features of the data, combine a variational autoencoder to map the data to a latent space and generate new data, and then optimize the feature representation and network training through an SE module and an identity mapping residual connection.

[0112] The hidden state of the recurrent neural network model is initialized as follows:

[0113] ;

[0114] wherein, is the initial hidden state of the recurrent neural network model RNN, is the input parameter of the recurrent neural network model, denotes the initialization of the weight matrix;

[0115] The recurrent neural network model takes an incomplete data set, a missing data set, and a data feature encoder obtained by a hierarchical fusion multi-dimensional fault feature extraction network as input, performs recursive processing, and uses the hidden state at the previous time and the latent variable at the current time to calculate the hidden state at the current time :

[0116] ;

[0117] wherein, denotes the weight matrix of the hidden layer, which is responsible for weighting transformation of the hidden state at the previous time, denotes the weight matrix of the input layer, which is used for weighting the latent variable at the current time ; is a ReLU activation function, is the parameter of the recurrent neural network model, is a bias term;

[0118] Through such recursive calculation, the RNN gradually captures the time series features of the data. Finally, the time series information is integrated by summation operation to obtain The output generated by the recurrent neural network model provides a time series-based feature representation for subsequent processing:

[0119]

[0120] wherein, is the integrated time series information, i.e., the output generated by the recurrent neural network model, and N represents the total number of time steps processed by the recurrent neural network;

[0121] The data processed by the recurrent neural network is mapped to the latent space Z through the variational autoencoder, and new data is generated from the latent space using the decoder. The operation process of the variational autoencoder is as follows:

[0122]

[0123] wherein, represents the variable and parameter set of the variational autoencoder, is the variable of the latent space, representing the feature representation of the input data in the latent space after encoding, is the parameter set in the variational autoencoder, represents the weight of the encoder part of the variational autoencoder, represents the bias of the encoder part of the variational autoencoder; is the sampling noise generated from the standard normal distribution, represents a nonlinear activation function, represents the weight of the decoder, represents the weight and bias of the decoder, used to decode the variable m in the latent space to generate new data.

[0124] Through the operation of the variational autoencoder, the mapping from the original feature representation to the latent space and the generation of new data are realized, further mining the data features and increasing the data diversity. The generated data feature map enters the SE module, which adjusts the importance of each channel using a weighting coefficient through a fully connected layer, making the model more focused on key features, with the formula:

[0125]

[0126] wherein, represents processing the input feature map , and the feature map adjusted by the channel weight is obtained by calculation, represents the output of the generator at the current time, represents the parameter set of the SE module, including the weights, biases, and other parameters of the components such as the fully connected layer in the module. represents the weight matrix of the fully connected layer in the SE module, is the bias term,​​​ represents a global average pooling operation on the input feature, S is a sigmoid activation function;

[0127] The feature processed by the SE module is further optimized through an identity mapping residual connection, and the expression is:

[0128] ;

[0129] wherein, represents an identity mapping residual, is the output of the generator at the previous moment, which generates data by adding noise and the input feature ; The residual block adds the output of the generator to its input to form a residual, which helps to alleviate the gradient disappearance problem, accelerate network training and optimization, and ensure that the generator learns and generates data stably and efficiently in multiple rounds of calculation;

[0130] In the regularized discriminator, the feature encoder encodes the incomplete data set and the missing data set to obtain the feature representation , and the final feature encoding representation is obtained by weighted summation:

[0131] ;

[0132] wherein, represents the feature representation generated by different feature encoders, is the weight coefficient of each feature encoder, indicating the importance of different features in discrimination, is the total number of feature encoders, represents the final feature encoding representation, is the set of all related parameters;

[0133] Then, the discriminator performs gradient calculation on the incomplete data set and the missing data set to normalize the features and ensure the stability of the gradient. This process involves the gradient calculation formula: .

[0134] wherein, represents the gradient of the input data , is the feature representation function in the gradient calculation related operation, is the corresponding parameter.

[0135] The input feature is activated by a sigmoid activation function to obtain the true and false probability:

[0136]

[0137] wherein, is the output of the regularized discriminator, representing the data is the probability of real data, S is the Sigmoid activation function, ensuring that the output value is between 0 and 1, is the regularization coefficient, used to control the normalization of the gradient, is a hyperparameter used to control the size of the regularization term.

[0138] In the data reconstruction link, in order to make the frequency characteristics of the generator output data meet the target requirements, the overall reconstruction loss function is defined as follows by combining the adversarial loss and the mean square error loss of the frequency characteristics:

[0139]

[0140] where G represents the hybrid generator, responsible for generating simulated data; D is the regularized discriminator, used to distinguish between real and fake data; x represents real data, and z is the noise vector input to the generator; and are the expectations of the real data and the noise vector, respectively; and is the logarithmic probability of the regularized discriminator's judgment of the data as real or fake; F is a function that converts data to the frequency domain; is the generator output data, is its frequency domain feature, is the target frequency feature; measures the difference between the two, and is a hyperparameter that weighs the proportion of frequency feature difference loss.

[0141] Through the collaborative work of the hybrid generator and the regularized discriminator, the enhanced orthogonal space heterogeneous network can effectively reconstruct the missing data, significantly improve the integrity of the data, and provide high-quality data support for fault diagnosis, ultimately improving the performance and reliability of the diagnosis model.

[0142] Through the above data reconstruction, incomplete information and missing information are input to the hybrid generator to form a reconstructed complete data set, which, together with the original complete data set, forms a standard data set. In order to better correct the wrong labels and missing labels, the standard data set will re-enter the cross-domain attention mechanism network, the hierarchical fusion multi-dimensional fault feature extraction network, form a standard data set feature encoder, and serve as the input sequence of the spatial information position encoding architecture between different sensors and fault types.

[0143] The specific construction of the spatial relative position encoding system for different fault types and sensors includes:

[0144] First, calculate the spatial position encoding function , whose expression is:

[0145] ;

[0146] ;

[0147] wherein, represents a spatial position encoding function, is a characteristic representation of Fourier transform, represents a weight matrix of the characteristic, is a sensor characteristic, is a standard deviation of a Gaussian kernel, is a number of position matrices considered, is a fault type index, is a dimension index, is a total number of vector dimensions;

[0148] On the basis of obtaining the spatial position encoding function result, further calculation of discriminative features between different sensors and between different fault types is carried out. The discriminative features are calculated by the following formula:

[0149] ;

[0150] wherein, represents a position encoding input, represents a multi-layer perception, represents a Gaussian error linear unit, represents a sensor state at a previous moment;

[0151] And the calculation of the fault type related features is related to the intermediate variable , and is calculated first:

[0152]

[0153] Then, different fault types are calculated by :

[0154]

[0155] The multi-layer perception MLP here extracts and converts the input data, and the activation function GELU introduces a nonlinear property to the calculation result, so that and can better distinguish the characteristics of different sensors and fault types.

[0156] Further, as shown in Figure 7 , in order to fill in the missing labels and classify, based on the position encoding system of different fault types and sensors, spatial information feature matching is carried out between different fault types and sensors, which specifically includes:

[0157] Based on discriminative features between different sensors and different fault types , a similarity matrix is calculated:

[0158] ;

[0159] where, denotes the similarity matrix, denotes the vector dot product, denotes the Euclidean norm of the vector, , ;

[0160] Convert the similarity matrix to a soft assignment matrix :

[0161] ;

[0162] ;

[0163] ;

[0164] where, and denote the similarity matrix after row and column softmax processing, respectively, denotes the relevance index of different fault types and different sensors, and the soft assignment matrix represents the relationship between different sensors and fault types measured and weighted by the time-frequency domain feature similarity.

[0165] Based on calculate the spatial information feature matching:

[0166] ;

[0167] where, denotes the set of high-confidence pairs, denotes the threshold, and for each matching pair , if , it indicates a high-confidence match.

[0168] Use the mutual information maximization strategy to further improve the quality and consistency of feature matching:

[0169] ;

[0170] where, denotes the pair set after further screening, and the quality and consistency of the pairs are ensured by optimizing the maximization of information.

[0171] Further, to further improve the label accuracy, the same fault type and sensor information after matching are projected by spatial difference characteristics, and then the fault classification result is obtained, the specific steps are as follows:

[0172] For sensor o and sensor p time-frequency domain feature F o and F p , perform preliminary matching operation to obtain matching pair , on this basis, calculate the offset , and perform spatial feature projection:

[0173] ;

[0174] Among them, indicates projecting the oth sensor information onto the pth sensor;

[0175] Based on the result of spatial feature projection, calculate the correction vector :

[0176] ;

[0177] Among them, and are weight coefficients for balancing the offset and feature difference;

[0178] According to the calculated correction vector , update the matching pair: .

[0179] Further, in the label self-correction process, introduce matching loss and difference loss , ensure the quality of data reconstruction and label self-correction, so as to improve the accuracy of fault diagnosis, the specific process is as follows:

[0180] Calculate the index of ground truth matching, and combine the soft assignment matrix to calculate the matching loss, defined as spatial information feature matching loss:

[0181] ;

[0182] Among them, indicates the index set of ground truth matching, that is, the index set of correct matching pairs, indicates the balance factor for balancing the weight of positive and negative samples, indicates the adjustment factor for adjusting the attention of the model to difficult samples, indicates the total number of matching pairs;

[0183] According to the difference offset and the corrected ground truth offset, the difference loss is defined as:

[0184] ;

[0185] wherein, is a differentiated offset, is a corrected ground truth offset, and K is the number of differentiated projections.

[0186] To verify the practicability of the method, a case of fault diagnosis of a key component of a power system was taken as an example, and multi-channel signal data were collected from a fault simulation test bench of a traction motor. The data covered various operating conditions of the motor in different health states, and four kinds of imperfect information were designed to simulate the common problems of data loss and incomplete labels in the actual fault diagnosis process.

[0187] The signals collected in the experiment included nine channels with a sampling frequency of 25.6 kHz, corresponding to three-axis acceleration signals and three-phase current signals at the driving end (DE) and the fan end (FE). The data set extracted samples from the signals using a sliding window with a length of 1024, and there was no overlap between the windows. At each motor operating speed, a total of 1200 sample data were extracted, of which 80% were used for training the model and 20% were used for testing the model performance. In addition, eight different motor health states were simulated in the experiment.

[0188] To verify the superiority of the spatial differentiated feature projection technology and its self-correction effect in cooperation with the spatial information feature matching, three comparison methods were designed for testing: method A only introduced spatial information feature matching; method B directly used spatial differentiated feature projection but skipped spatial information feature matching; and the method of the application combined the cooperative self-correction mechanism of spatial information feature matching and spatial differentiated feature projection. The three methods were verified under the same network structure and four experimental conditions, and the quantitative results are shown in Table 1. Figure 8

[0189] The experimental results show that method A has lower classification accuracy and evaluation indicators than the method of the application due to only introducing spatial information feature matching. This shows that relying only on spatial information feature matching is insufficient to handle label missing and error problems. The method of the application uses spatial differentiated feature projection technology to calculate the balanced offset between the time-frequency domains of fault features, generates a correction vector, and effectively corrects missing and erroneous labels, thereby significantly improving the classification accuracy. This verifies the effectiveness of the spatial differentiated feature projection technology in dealing with label errors and data loss.

[0190] ​On the other hand, compared with method B, the method of the application performs better in classification accuracy, macro average F1 value and ranking loss. Method B uses spatially differentiated feature projection, but its classification accuracy is not stable because it skips spatial information feature matching. The method of the application significantly improves the robustness and diagnostic accuracy of the model through the synergistic self-correction mechanism of the two, further proving the superiority of the synergistic self-correction.

[0191] To further verify the performance of the method of the application, it is compared with four existing fault diagnosis methods for incomplete information. The four methods include fuzzy clustering framework based method (method C), dynamic Bayesian network based method (method D), convolutional neural network based method (method E) and three-way decision model based method (method F). The experiment uses a hierarchical fusion multi-dimensional fault feature extraction network as the basic architecture, the basic structure is shown in Table 1, the training parameters are shown in Table 2, and the experimental results are summarized in Table 3. Figure 9 .

[0192] Table 1. Summary of network basic structure

[0193]

[0194] Table 2. Summary of training related parameters

[0195]

[0196] The comparison results show that the method of the application is superior to the three existing methods in terms of classification accuracy, macro average F1 value and generalization ability. This shows that the method of the application can perform better in terms of robustness and diagnostic efficiency under imperfect information conditions such as data missing and label errors. In particular, the self-correction mechanism combining spatial feature matching and differentiated feature projection enables the method of the application to more effectively correct incorrect labels and restore high-quality data, thereby significantly improving diagnostic accuracy.

[0197] Through the above experiments and comparative analysis, it can be seen that the method of the application has practicality and superiority in dealing with complex fault diagnosis problems. The synergistic self-correction technology combining spatial information feature matching and differentiated feature projection enables the method of the application to effectively solve the common information missing and incomplete label problems in practical applications, and has high application value and technical advantages.

[0198] The embodiments in the specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be mutually referred to. For the device disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple, and the relevant parts can be referred to the method part.

[0199] The foregoing description of the disclosed embodiments enables a person skilled in the art to make or use the application. Modifications of these embodiments will occur to persons of skill in the art, and that the appended claims are intended to cover all such modifications that do not depart from the true spirit and scope of the application. Therefore, the application is not limited to the embodiments shown but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A fault diagnosis method under imperfect information condition based on orthogonal space isomerism mapping, characterized in that, The method comprises the following steps: S1: integrity judgment is performed on the fault data set, which is divided into a complete data set, an incomplete data set and a missing data set; S2: the complete data set is input into a cross-domain attention mechanism network, and the output of the cross-domain attention mechanism network is input into a hierarchical fusion multi-dimensional fault feature extraction network, and after processing, a data feature encoder output is formed; S3: the incomplete data set, the missing data set and the output of the data feature encoder are input into an enhanced orthogonal space heterogeneous network of an embedded multi-dimensional generator, efficient data reconstruction is performed on the data missing and incompleteness problem, a reconstructed complete data set is obtained, and the reconstructed complete data set and the complete data set jointly constitute a standard data set, and the standard data set is input into the hierarchical fusion multi-dimensional fault feature extraction network based on the cross-domain attention mechanism network again to generate data features related to different fault types and sensors; S4: spatial information position coding is performed on the data features obtained after processing of the standard data set to form a spatial relative position coding system of different fault types and sensors; S5: a label self-correction strategy based on spatial differentiation is proposed, spatial information feature matching is performed on different fault types and sensors by using the spatial relative position coding system of different fault types and sensors, and spatial feature projection is performed on the matched same fault type and sensor information; and finally, a fault classification result is obtained; The enhanced orthogonal space heterogeneous network of the embedded multi-dimensional generator in S3 comprises a hybrid generator and a regularization discriminator: The hybrid generator uses a recurrent neural network model to capture the time sequence characteristics of data, combines a variational autoencoder to map the data to a latent space and generate new data, and then optimizes the feature representation and network training through an SE module and an identity mapping residual connection; The hidden state of the recurrent neural network model is used for initialization: ; wherein, is an initial hidden state of a recurrent neural network model RNN, is an input parameter of the recurrent neural network model, denotes an initialization of a weight matrix; The recurrent neural network model takes as input an incomplete data set, a missing data set, and an output of a data feature encoder, performs recursive processing, and utilizes a hidden state of a previous time point and a latent variable of a current time point to calculate a hidden state of the current time point : ; wherein, denotes the weight matrix of the hidden layer, responsible for the weighted transformation of the hidden state at the previous time step, denotes the weight matrix of the input layer, used for the weighting of the latent variable at the current time step; is the ReLU activation function, are parameters of the recurrent neural network model, is the bias term; Through recursive calculation, the recurrent neural network model gradually captures the time series characteristics of the data, integrates the time series information through summation operation, and obtains As the output result generated by the recurrent neural network model: ; wherein, is the integrated time series information, N denotes the total number of time steps processed by the recurrent neural network. The data processed by the recurrent neural network model is mapped to a latent space by a variational autoencoder Z And a decoder is used to generate new data from the latent space, and the operation process of the variational autoencoder is as follows: ; wherein, denote the set of variables and parameters of the variational autoencoder, is the variable of the latent space, representing the encoded feature representation of the input data in the latent space, is the set of parameters in the variational autoencoder, denote the weights of the encoder part of the variational autoencoder, denote the bias of the encoder part of the variational autoencoder; is the sampling noise, generated from a standard normal distribution, denote the non-linear activation function, denote the weights of the decoder, denote the weights and bias of the decoder, used to decode the variable m of the latent space to generate new data; The generated data feature map enters the SE module, the SE module adjusts the importance of each channel through a fully connected layer using a weighting coefficient to make the model focus on key features, and the formula is: ; wherein, denotes processing the input feature map , and obtaining the feature map adjusted by channel weights through calculation, denotes the output of the generator at the current moment, denotes a parameter set of the SE module, denotes a weight matrix of the fully connected layer in the SE module, is a bias term, denotes a global average pooling operation on the input feature, and S is a Sigmoid activation function. The features processed by the SE module are further optimized through the identity mapping residual connection, and the expression is: ； wherein, denotes the identity mapping residual, is the output of the generator at the previous time step, which is generated by adding noise to the input features to generate data, are the parameters of the corresponding network layers; In the regularized discriminator, the feature encoder encodes the incomplete dataset and the missing dataset to obtain a feature representation The final feature encoding representation is obtained by weighted summation ； wherein: denote feature representations generated by different feature encoders, are weight coefficients of each feature encoder, indicating the importance of different features in discrimination, is the total number of feature encoders, denotes the final feature encoding representation, is the set of all relevant parameters; The regularization discriminator performs gradient calculation on an incomplete data set, a missing data set, and the gradient calculation formula is: ; wherein, represents a feature representation function for input data gradient, is a feature representation function in gradient computation related operations, is a corresponding parameter; The input features are activated through a sigmoid activation function to obtain true and false probabilities: ； where, is the output of the regularized discriminator, representing data is the probability of real data, S is the Sigmoid activation function, ensuring that the output values are between 0 and 1, is the regularization coefficient, used to control the normalization of the gradient, is a hyperparameter, used to control the size of the regularization term.

2. The method for fault diagnosis under the condition of imperfect information based on orthogonal space isomerism mapping according to claim 1, characterized in that, The cross-domain attention mechanism network comprises a channel attention network and a spatial attention network in sequence, the channel attention network comprises three parallel branches, convolution layers, activation functions and convolution layers are arranged after splicing of the three parallel branches, the first branch comprises a convolution layer, an asymmetric convolution block, a batch normalization and an activation function, the second branch comprises a convolution layer, two asymmetric convolution blocks, a batch normalization and an activation function, and the third branch comprises a convolution layer and an average pooling layer; the spatial attention network comprises an average pooling layer and a maximum pooling layer, a plurality of asymmetric convolution blocks and a sigmoid activation function are sequentially arranged after splicing of the average pooling layer and the maximum pooling layer; The processing process of the cross-domain attention mechanism network is as follows: ； wherein, represents the output of the cross-domain attention mechanism network, represents a learnable weight parameter, represents sensor data, represents an element-wise multiplication operation, represents a ReLU activation function, represents an asymmetric convolution block with kernel size k represents a 1x1 convolution operation, represents an average pooling operation, represents a max pooling operation, represents a Sigmoid activation function, represents a concatenation operation of features, represents a sum operation over all feature channels, represents a nonlinear activation function.​ 3. The method for fault diagnosis under the condition of imperfect information based on orthogonal space isomerism mapping according to claim 1, characterized in that, The output of the cross-domain attention mechanism network enters the hierarchical fusion multi-dimensional fault feature extraction network, and the hierarchical fusion multi-dimensional fault feature extraction network comprises the following steps: The output of the cross-domain attention mechanism network is divided into four sensor data streams, and each sensor data stream enters a hierarchical fusion multi-dimensional fault feature extraction network, which includes multi-branch convolution, convolution pooling and multi-branch pooling. The multi-branch convolution processing process is as follows: ； wherein, denotes a multi-branch convolutional block output, BN and BP represent data normalization and batch normalization operations for the multi-branch convolutional layer, respectively, denotes a max-pooling operation, f = 1, 3, denotes a f x f standard convolution, represents a convolution sum, denotes an asymmetric convolution operation with a kernel size of 3 and g number of repetitions; The output after convolution pooling is calculated as follows: ; wherein, is the output after convolutional pooling, denotes an asymmetric convolution operation with a kernel size of 1 repeated 3 times; The multi-branch pooling processing process is as follows: ; wherein, is a multi-branch pooling block output, denotes an average pooling operation repeated times.

4. The method for fault diagnosis under the condition of imperfect information based on orthogonal space isomerism mapping according to claim 1, characterized in that S4 It includes: Computing a spatial position encoding function whose expression is: ； ； wherein, represents a spatial position encoding function, is a characteristic representation of a Fourier transform, represents a weight matrix of the characteristic, is a sensor characteristic, is a standard deviation of a Gaussian kernel, is the number of position matrices considered, is a failure type index, is a dimension index, is the total number of vector dimensions; distinguishing features between different sensors and between different fault types, distinguishing features are calculated by the following equation: ； wherein, represents a position encoding input, represents a multi-layer perceptron, represents a Gaussian error linear unit, represents a previous time step sensor state; Computing : ； wherein, represents a sensor feature aligned with a feature of a fault prototype vector, represents a sensor space feature, represents a fault prototype feature; Further by Computing different fault types : 。 5. The method for fault diagnosis under the condition of imperfect information based on orthogonal space isomerism mapping according to claim 4, characterized in that, By using different fault types and sensor spatial relative position coding system, spatial information feature matching is performed on different fault types and sensors, which specifically includes: Based on discriminative features between different sensors and different fault types , a similarity matrix is computed: ； wherein, denotes a similarity matrix, denotes a vector dot product, denotes the Euclidean norm of a vector, , ; Converting a similarity matrix into a soft assignment matrix : ； ； ； wherein, with respectively represent the similarity matrix after row and column softmax processing, represents the correlation index of different fault types and different sensors, and the soft assignment matrix represents the relationship between different sensors and fault types measured and weighted by the time-frequency domain feature similarity. Based on Computing spatial information feature matches: ； wherein, denotes a set of high-confidence pairings, denotes a threshold, for each matching pair if denotes a high-confidence match; The mutual information maximization strategy is adopted to further improve the quality and consistency of feature matching: ； wherein, represents the pair set after further screening, and the quality and consistency of the pairing are ensured by optimizing the maximization of information.

6. The method for fault diagnosis under the condition of imperfect information based on orthogonal space isomerism mapping according to claim 1, characterized in that, The spatial feature projection is performed on the matched same fault type and sensor information, and finally the fault classification result is obtained, and the specific steps are as follows: For sensor o and sensor p, time-frequency domain features F o and F p are obtained, and a preliminary matching operation is performed to obtain a matching pair On this basis, an offset is calculated and spatial feature projection is performed: ； wherein represents the projection of the oth sensor information onto the pth sensor; Based on the results of the spatial feature projection, a correction vector is calculated : ； wherein, and are weight coefficients for balancing the offset and feature difference; the computed correction vector , update the matching pair: .

7. The method for fault diagnosis under the condition of imperfect information based on orthogonal space isomerism mapping according to claim 1, characterized in that, Introducing reconstruction loss in data reconstruction process Introducing matching loss in sticky note self-correction process and differentiation loss The specific process is as follows: In the data reconstruction link, the total reconstruction loss function is defined as follows: ； wherein, G denotes the hybrid generator, responsible for generating the simulated data; D is the regularized discriminator, used to distinguish the data true or false; x represents the real data, z is the noise vector input to the generator; and are the expectations of the real data and the noise vector, respectively; and are the logarithmic probabilities of the data true or false judged by the regularized discriminator; F is the function of converting the data to the frequency domain; is the generator output data, is the frequency domain feature thereof, is the target frequency feature; measures the difference between the two, which is a hyperparameter for balancing the proportion of the frequency feature difference loss. Computing an index of ground truth matches and combining with soft assignment matrix Computing spatial information feature matching loss : ； wherein, denotes a set of indices of ground truth matches, i.e. a set of indices of correctly matched pairs, denotes a balancing factor for balancing the weights of positive and negative samples, denotes an adjustment factor for adjusting the model's focus on hard-to-classify samples, denotes the total number of matched pairs; According to the differentiated offset and the corrected ground true offset, the differentiated loss is defined as: ； wherein, is a differential offset, is a corrected ground true offset, K is the number of differential projections.

Images