A causal decoupling representation method for data based on non-uniform superpixel graph neural network

Abstract

The present invention relates to the field of fault diagnosis technology, and specifically to a causal decoupling characterization method based on non-uniform superpixel graph neural network data. The present application can efficiently process and analyze the fault data of the object to be tested, convert complex signals into intuitive time-frequency images through fast continuous wavelet transform, and then refine image features using superpixel segmentation technology, which not only improves the accuracy of feature extraction, but also realizes deep mining and accurate prediction of the causal relationship of the fault of the object to be tested by constructing a relationship graph and a graph convolution network model, providing strong support for real-time monitoring and diagnosis of the fault of the object to be tested, and helping to improve the maintenance efficiency and operation safety of the equipment.

Description

Technical Field

[0001] The present invention relates to the field of fault diagnosis technology, and in particular to a causal decoupling characterization method based on non-uniform superpixel graph neural network data. Background Art

[0002] At present, intelligent fault diagnosis methods based on deep learning have become a popular choice among many diagnostic technologies. Although fault diagnosis methods based on deep learning have achieved certain success, there are still some problems. In terms of feature extraction, fault signals are usually multidimensional, nonlinear and non-stationary. For example, when a fault occurs, the vibration signal may show characteristics such as mutation and waveform distortion. Conventional feature extraction methods have limitations in processing such nonlinear and non-stationary signals and cannot capture the dynamic changes in fault signals. This incomplete feature extraction will lead to the loss of some important information, thereby reducing the accuracy and reliability of fault diagnosis. In addition, although the above-mentioned fault diagnosis methods reduce the requirements for model establishment, the models established in this way have limitations in interpretability and generalization ability, which makes it difficult to establish a causal relationship between signal features and fault causes, which further affects the accuracy and reliability of the diagnosis results, so it is urgent to solve. Summary of the invention

[0003] In order to avoid and overcome the technical problems existing in the prior art, the present invention provides a causal decoupling characterization method based on non-uniform superpixel graph neural network data. The present invention can effectively improve the accuracy of fault diagnosis of the object to be tested.

[0004] To achieve the above object, the present invention provides the following technical solutions:

[0005] A causal decoupling representation method based on non-uniform superpixel graph neural network data includes the following decoupling steps:

[0006] S1. Obtaining fault data of the object to be tested including causal features and their corresponding fault types, and processing each fault data of the object to be tested using a fast continuous wavelet transform to generate a corresponding time-frequency image;

[0007] S2, performing super-pixel segmentation on each time-frequency image to obtain a plurality of super-pixels;

[0008] S3, obtaining feature vectors of each superpixel in the same time-frequency image, and calculating the cosine similarity between the superpixels through the feature vectors;

[0009] S4, constructing a relationship graph representing the connection relationship between each superpixel in the time-frequency image based on each cosine similarity in the same time-frequency image;

[0010] S5. Based on the graph convolutional network, a causal relationship inference model is constructed, and each relationship graph is used to train the causal relationship inference model to obtain a causal relationship inference model that meets the set requirements, and the model is recorded as the optimal model;

[0011] S6. According to the process from step S1 to step S4, the real-time fault data of the object to be tested is processed to obtain a corresponding relationship diagram; the relationship diagram is input into the optimal model to predict the fault type corresponding to the real-time fault data of the object to be tested.

[0012] As a further solution of the present invention: the causal relationship inference model includes a graph convolutional network layer, a multi-layer perceptron and a visual graph neural network module arranged in sequence along the data transmission direction; based on the graph convolutional network layer, the multi-layer perceptron and the visual graph neural network module, the process of training the causal relationship inference model in combination with the relationship graph is as follows:

[0013] S51. Count all superpixels in all relationship graphs and store them in the same overall graph G middle, ;

[0014] in, Display the overall picture G The adjacency matrix represents the connection relationship between all superpixels, and each element in the adjacency matrix is ​​the cosine similarity between two superpixels; Display the overall picture G The feature matrix formed by combining the feature vectors of each superpixel in;

[0015] S52, the overall map G Input to the graph convolutional network layer for processing to obtain node representation ;

[0016] in, H Display the overall picture G Corresponding node representation; GCN represents a graph convolutional network, Represents the use of graph convolutional networks to map G to process;

[0017] S53. Based on the soft mask mechanism, the feature weights corresponding to the causal features and false features in the node representation are obtained through the multi-layer perceptron; the feature weights are expressed as follows:

[0018] ;

[0019] In the formula, Representation of node representation H The feature weights of the causal features in Representation of node representation HThe feature weights of the spurious features in ; represents a multilayer perceptron; Represented by multi-layer perceptron Node characterization H Perform weight estimation; Represents the normalization function softmax , Indicates the use of normalization function softmax right Perform normalization processing;

[0020] S54. Separate node representations based on the acquired feature weights H causal and spurious features in ;

[0021] ;

[0022] ;

[0023] In the formula, Representation of node representation H causal features in Representation of node representation H False features in Represents multiplication operation;

[0024] S55. Use the causal feature to train the visual graph neural network module, and predict the fault type corresponding to the causal feature through the visual graph neural network, and use the model loss function to calculate the prediction accuracy; the model loss function is specifically expressed as follows:

[0025] ;

[0026] In the formula, Represents the model loss function; Represents the loss function of the visual graph neural network module; represents the independence loss function, express The weight coefficient of represents the backdoor adjustment loss function, express The weight coefficient of

[0027] S56. If the model meets the set requirements, the training is stopped and the current causal relationship inference model is output, which is the optimal model; otherwise, the weight parameters in the graph convolutional network layer, multi-layer perceptron and visual graph neural network module are adjusted by negative feedback, and the training is repeated according to the contents of steps S52 to S55 until the optimal model is obtained.

[0028] As a further solution of the present invention: the loss function calculation formula of the visual graph neural network module is as follows:

[0029] ;

[0030] In the formula, N Display the overall picture G The total number of superpixels; Representing causal characteristics The n The feature vector of superpixels; Representation using visual graph neural network predict Type of fault; express The type of failure predicted; Indicates n The fault type corresponding to each superpixel; Indicates the use of normalization function softmax right Perform normalization processing; express and The cross entropy loss between .

[0031] As a further solution of the present invention: the independence loss function is specifically expressed as follows:

[0032] ;

[0033] In the formula, Represents false features The n The feature vector of superpixels; express and The Hilbert-Schmidt independence criterion between .

[0034] As a further solution of the present invention: the backdoor adjustment loss function is specifically expressed as follows:

[0035] ;

[0036] In the formula, Represents false features The m The feature vector of superpixels; Represents the use of graph convolutional networks predict Type of fault; express The type of failure predicted; Indicates the use of normalization function softmax right Perform normalization processing; express and The cross entropy loss between .

[0037] As a further solution of the present invention: there are two types of setting requirements, and the optimal model can be obtained by satisfying either of the two requirements; one is: the number of training times reaches the set value; the other is: the model accuracy reaches the set accuracy.

[0038] As a further solution of the present invention: the specific content of step S1 is as follows:

[0039] S11, obtaining fault data of the object to be tested, where the fault data of the object to be tested includes causal characteristics and their corresponding fault types;

[0040] S12, using fast continuous wavelet transform to process the fault data of each object to be tested, and obtain processing results;

[0041] S13. Performing a visualization operation on the processing result to convert it into a corresponding time-frequency image.

[0042] As a further solution of the present invention: a simple linear iterative clustering algorithm is used to perform superpixel segmentation on the time-frequency image to obtain a corresponding number of superpixels.

[0043] As a further solution of the present invention: a convolutional neural network is used to extract the feature vector of each superpixel from the current time-frequency image, and the feature vector It is expressed as: ;

[0044] in, Represents the grayscale mean in the current superpixel; Represents the mean value of the red channel, the green channel, and the blue channel in the current superpixel; Represents the texture mean in the current superpixel; Represents the area and perimeter of the current superpixel.

[0045] As a further solution of the present invention: the relationship diagram corresponding to the current time-frequency image It is expressed as: ;in, represents a node, i.e., a superpixel; Represents the edge, that is, the connection relationship between superpixels; Indicates hidden state.

[0046] Compared with the prior art, the present invention has the following beneficial effects:

[0047] 1. This application can efficiently process and analyze the fault data of the object to be tested, convert complex signals into intuitive time-frequency images through fast continuous wavelet transform, and then refine image features using superpixel segmentation technology. It not only improves the accuracy of feature extraction, but also realizes deep mining and accurate prediction of the cause-effect relationship of the fault of the object to be tested by constructing relationship graphs and graph convolution network models, providing strong support for real-time monitoring and diagnosis of faults of the object to be tested, and helping to improve the maintenance efficiency and operation safety of mechanical equipment.

[0048] 2. By introducing graph convolutional network layers, multi-layer perceptrons, and visual graph neural network modules, a complete model architecture is formed, which can fully utilize the advantages of graph convolutional networks in processing graph structure data, combined with the feature weight allocation capability of multi-layer perceptrons and the fault type prediction capability of visual graph neural networks, to achieve accurate decoupling prediction of the cause-effect relationship of faults of the object to be tested. In addition, through detailed training steps and loss function design, the effectiveness and stability of model training are ensured, and the prediction accuracy and generalization ability of the model are improved.

[0049] 3. The loss function calculation formula of the visual graph neural network module uses cross entropy loss to measure the difference between the predicted fault type and the actual fault type, thereby guiding the optimization of model parameters. This loss function design is not only concise and clear, but also can effectively reflect the accuracy of model prediction, which helps to improve the training efficiency and prediction performance of the model.

[0050] 4. The independence loss function uses the Hilbert-Schmidt independence criterion to measure the independence between false features, thereby ensuring that the model can effectively separate causal features from false features during the decoupling process. This independence constraint helps improve the decoupling ability of the model, allowing the model to maintain a high prediction accuracy in complex fault scenarios.

[0051] 5. The backdoor adjustment loss function further constrains the model's learning process by calculating the cross entropy loss between false features and predicted fault types, which helps reduce the overfitting phenomenon of the model during training and improves the model's generalization ability and prediction stability.

[0052] 6. Two setting requirements are used to determine whether the model has reached the optimal state, that is, the number of training times reaches the set value or the model accuracy reaches the set accuracy. This flexible stop condition design not only ensures the adequacy of model training, but also avoids unnecessary waste of computing resources. At the same time, this setting also helps to improve the controllability and stability of model training.

[0053] 7. Using a simple linear iterative clustering algorithm to perform superpixel segmentation on the time-frequency image can quickly obtain the corresponding number of superpixels, providing strong support for subsequent feature extraction and relationship graph construction.

[0054] 8. Use convolutional neural networks to extract superpixel feature vectors from time-frequency images, and by comprehensively considering the grayscale, color channel, texture, area, perimeter and other information of superpixels, this feature extraction method can more comprehensively reflect the characteristics of superpixels and provide a more accurate data basis for subsequent causal relationship deduction.

[0055] 9. The relationship diagram intuitively shows the correlation between each superpixel in the time-frequency image, providing a clear data structure for subsequent graph convolutional network processing and causal relationship derivation. At the same time, the introduction of hidden states also provides the model with more information sources and learning capabilities, which helps to improve the model's prediction performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 It is a schematic diagram of the decoupling process of the present invention.

[0057] Figure 2 is the original time-frequency image in the embodiment of the present invention.

[0058] Figure 3 Schematic diagram of segmentation of an original time-frequency image into 20 superpixels in an embodiment of the present invention.

[0059] Figure 4 Schematic diagram of segmentation of an original time-frequency image into 50 superpixels in an embodiment of the present invention. DETAILED DESCRIPTION

[0060] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0061] See also Figure 1 The object to be tested in the embodiment of the present invention is a bearing, which mainly includes the following contents:

[0062] 1. Obtain time-frequency images.

[0063] In bearing fault diagnosis, the generation of time-frequency images is crucial to capturing the frequency characteristics and time changes of signals. Traditional short-time Fourier transform (STFT) and wavelet transform often face the problem of uneven time-frequency distribution when processing signals, which may lead to reduced efficiency and accuracy of feature extraction. To solve this problem, this embodiment uses the fast continuous wavelet transform (FCWT) method for signal preprocessing to generate a more representative time-frequency image. The specific steps are as follows:

[0064] S11. Acquire bearing fault data, where the bearing data includes causal characteristics and their corresponding fault types.

[0065] Obtain bearing fault data from open source datasets (such as the CWRU bearing dataset) or from actual industrial environments. Bearing fault data should include causal features (such as vibration signals) and their corresponding fault types (such as inner race fault, outer race fault, etc.). Ensure that the bearing fault data is in the format of a time series signal, usually stored in .csv, .mat, or .txt file formats. Use a programming language (such as Python) to import bearing fault data and use libraries such as Pandas and NumPy to process bearing fault data. Extract time series data such as vibration signals as causal features.

[0066] S12. Use fast continuous wavelet transform to process the fault data of each bearing and obtain the processing result.

[0067] Use PyWavelets or the FCWT function in the SciPy library to process the vibration signal.

[0068] S13. Performing a visualization operation on the processing result to convert it into a corresponding time-frequency image.

[0069] Use Matplotlib or other libraries to visualize the FCWT results as a time-frequency image. The horizontal axis of the time-frequency image represents time, the vertical axis represents frequency (or scale), and the color represents the amplitude of the wavelet coefficient.

[0070] 2. Obtain superpixels.

[0071] The Simple Linear Iterative Clustering (SLIC) algorithm is a commonly used superpixel segmentation algorithm. It is based on the k-means clustering algorithm, but uses a more efficient initialization method and iteration strategy, and can generate high-quality superpixel segmentation results in a shorter time. This embodiment uses the simple linear iterative clustering algorithm to perform superpixel segmentation on the time-frequency image to obtain a corresponding number of superpixels.

[0072] like Figure 2 The original time-frequency image shown in the SLIC algorithm sets different numbers of partitions. For example, the original time-frequency image is a 4×4 pixel grayscale image, which is represented as follows:

[0073] ;

[0074] After processing by the SLIC algorithm, the superpixel label corresponding to the original time-frequency image can be obtained, which is specifically expressed as follows:

[0075] ;

[0076] Based on the above segmentation method, in actual use, the number of segmentations can be adjusted in real time according to the final prediction accuracy. Figure 3 As shown in , the original time-frequency image is divided into 20 super pixels. Figure 4 As shown in Figure 1, the original time-frequency image is segmented into 50 superpixels. By adjusting the number of segmentations, the prediction accuracy of the model is further improved.

[0077] 3. Calculate cosine similarity.

[0078] The feature vectors of each superpixel in the same time-frequency image are obtained. The specific process is as follows:

[0079] The convolutional neural network is used to extract the feature vector of each superpixel from the current time-frequency image, and the feature vector It is expressed as: . The corresponding feature is the grayscale feature, which represents the grayscale mean in the current superpixel. The corresponding color feature represents the mean value of the red channel, the mean value of the green channel, and the mean value of the blue channel in the current superpixel. The corresponding one is the texture feature, which represents the texture mean in the current superpixel. The corresponding is the shape feature, which indicates the area and perimeter of the current superpixel. The feature extraction is performed on the above 4×4 pixel grayscale image, and the extraction results are shown in Table 1.

[0080] Table 1 Feature extraction results

[0081] ;

[0082] Table 1 represents the averaging operation, represents the red channel, represents the green channel, Represents the blue channel.

[0083] The cosine similarity between the superpixels is calculated through the feature vector, as follows:

[0084] Select the first three superpixels in Table 1, calculate the cosine similarity, and obtain the corresponding adjacency matrix. The calculation results are shown in Table 2.

[0085] Table 2 Cosine similarity

[0086] ;

[0087] 4. Build a relationship diagram.

[0088] Based on the cosine similarities in the same time-frequency image, a relationship graph is constructed to represent the connection relationship between superpixels in the time-frequency image. It is expressed as: ;in, represents a node, i.e., a superpixel; Represents the edge, that is, the connection relationship between superpixels; Indicates hidden state.

[0089] 5. Obtain the optimal model.

[0090] Based on the graph convolutional network, a causal relationship inference model is constructed, and each relationship graph is used to train the causal relationship inference model to obtain a causal relationship inference model that meets the set requirements, and the model is recorded as the optimal model. The details are as follows:

[0091] The causal relationship inference model includes a graph convolutional network layer, a multi-layer perceptron, and a visual graph neural network module arranged in sequence along the data transmission direction; based on the graph convolutional network layer, the multi-layer perceptron, and the visual graph neural network module, the process of training the causal relationship inference model in combination with the relationship graph is as follows:

[0092] S51. Count all superpixels in all relationship graphs and store them in the same overall graph G middle, .

[0093] S52, the overall map G Input to the graph convolutional network layer for processing to obtain node representation .

[0094] S53. Based on the soft mask mechanism, the feature weights corresponding to the causal features and false features in the node representation are obtained through the multi-layer perceptron; the feature weights are expressed as follows:

[0095] ;

[0096] S54. Separate node representations based on the acquired feature weights H causal and spurious features in ;

[0097] ;

[0098] ;

[0099] S55. Use the causal feature to train the visual graph neural network module, and predict the fault type corresponding to the causal feature through the visual graph neural network, and use the model loss function to calculate the prediction accuracy; the model loss function is specifically expressed as follows:

[0100] ;

[0101] ;

[0102] ;

[0103] ;

[0104] S56. If the model meets the set requirements, that is, the number of training times reaches the set value or the model accuracy reaches the set accuracy, the training is stopped and the current causal relationship inference model is output. The causal relationship inference model is the optimal model; otherwise, the weight parameters in the graph convolutional network layer, the multilayer perceptron and the visual graph neural network module are adjusted by negative feedback, and the training is repeated according to the contents of steps S52 to S55 until the optimal model is obtained.

[0105] 6. Obtain the fault type corresponding to the real-time bearing fault data.

[0106] According to the above content, the real-time bearing fault data is processed to obtain a corresponding relationship diagram; the relationship diagram is input into the optimal model to predict the fault type corresponding to the real-time bearing fault data.

[0107] 7. Model evaluation.

[0108] In order to verify the performance of the present invention in a real industrial system, the present invention is compared with the existing baseline methods model.

[0109] The training set of laboratory bearing failure data was obtained, and the training set was randomly horizontally flipped, randomly rotated by ±10 degrees, and image normalized to enhance the generalization ability of the model. The validation set was only normalized in the same way.

[0110] To provide stronger evidence, accuracy, recall, F1 Score and confusion matrix are used as indicators. The complete model is trained 100 times, the learning rate is set to 0.001, and the weight of the loss term is and The SLIC algorithm is used for superpixel segmentation, and the number of superpixels is set to 100.

[0111] The experimental results were compared with the results of the Baseline Methods model. The average accuracy of this application was 95.25% and 88.16%, respectively, while the average accuracy of the Baseline Methods model was 87.65% and 70.82%, respectively. The accuracy of this application was 7.6% and 17.34% higher than that of the Baseline Methods model, respectively. This shows that this application can effectively improve the accuracy of bearing fault diagnosis.

[0112] 8. Three-phase flow facility data.

[0113] Since the Three-phase Flow Facility (TFF) of Cranfield University and its detailed data set have far-reaching application value in the field of fault diagnosis technology, they are applied to the method of this application and the Baseline Methods model. The experimental results show that the accuracy of this application is 9.41% higher than that of the Baseline Methods model, which once again proves that this application can effectively improve the accuracy of fault diagnosis.

[0114] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A causal decoupling representation method based on non-uniform superpixel graph neural network data, characterized in that: The decoupling steps include: S1. Obtaining fault data of the object to be tested including causal features and their corresponding fault types, and processing each fault data of the object to be tested using a fast continuous wavelet transform to generate a corresponding time-frequency image; S2, performing super-pixel segmentation on each time-frequency image to obtain a plurality of super-pixels; S3, obtaining feature vectors of each superpixel in the same time-frequency image, and calculating the cosine similarity between the superpixels through the feature vectors; S4, constructing a relationship graph representing the connection relationship between each superpixel in the time-frequency image based on each cosine similarity in the same time-frequency image; S5. Based on the graph convolutional network, a causal relationship inference model is constructed, and each relationship graph is used to train the causal relationship inference model to obtain a causal relationship inference model that meets the set requirements, and the model is recorded as the optimal model; S6. According to the process from step S1 to step S4, the real-time fault data of the object to be tested is processed to obtain a corresponding relationship diagram; the relationship diagram is input into the optimal model to predict the fault type corresponding to the real-time fault data of the object to be tested; The causal relationship inference model includes a graph convolutional network layer, a multi-layer perceptron, and a visual graph neural network module arranged in sequence along the data transmission direction; based on the graph convolutional network layer, the multi-layer perceptron, and the visual graph neural network module, the process of training the causal relationship inference model in combination with the relationship graph is as follows: S51. Count all superpixels in all relationship graphs and store them in the same overall graph G middle; S52, the overall map G Input to the graph convolutional network layer for processing to obtain node representation ; S53, based on the soft mask mechanism, the feature weights corresponding to the causal features and the false features in the node representation are obtained through the multi-layer perceptron; S54. Separate node representations based on the acquired feature weights H causal and spurious features in ; S55, using the causal feature to train the visual graph neural network module, and predicting the fault type corresponding to the causal feature through the visual graph neural network, and calculating the prediction accuracy using the model loss function; S56. If the model meets the set requirements, the training is stopped and the current causal relationship inference model is output, which is the optimal model; otherwise, the weight parameters in the graph convolutional network layer, multi-layer perceptron and visual graph neural network module are adjusted by negative feedback, and the training is repeated according to the contents of steps S52 to S55 until the optimal model is obtained.

2. According to claim 1, a causal decoupling characterization method based on non-uniform superpixel graph neural network data is characterized in that: General Map G It is expressed as follows: ； , Display the overall picture G The adjacency matrix represents the connection relationship between all superpixels, and each element in the adjacency matrix is ​​the cosine similarity between two superpixels; Display the overall picture G The feature matrix formed by combining the feature vectors of each superpixel in; Node Characterization It is expressed as follows: ； in, GCN represents a graph convolutional network, Represents the use of graph convolutional networks to map G to process; The feature weights are expressed as follows: ； In the formula, Representation of node representation H The feature weights of the causal features in Representation of node representation H The feature weights of the spurious features in ; represents a multilayer perceptron; Represented by multi-layer perceptron Node characterization H Perform weight estimation; Represents the normalization function softmax , Indicates the use of normalization function softmax right Perform normalization processing; The causal and spurious features are represented as follows: ； ； In the formula, Representation of node representation H causal features in Representation of node representation H False features in Represents multiplication operation; The model loss function is specifically expressed as follows: ； In the formula, Represents the model loss function; Represents the loss function of the visual graph neural network module; represents the independence loss function, express The weight coefficient of represents the backdoor adjustment loss function, express The weight coefficient of .

3. According to claim 2, a method for characterizing causal decoupling of non-uniform superpixel graph neural network data is characterized in that: The loss function calculation formula of the visual graph neural network module is as follows: ； In the formula, N Display the overall picture G The total number of superpixels; Representing causal characteristics The n The feature vector of superpixels; Representation using visual graph neural network predict Type of fault; express The type of failure predicted; Indicates n The fault type corresponding to each superpixel; Indicates the use of normalization function softmax right Perform normalization processing; express and The cross entropy loss between .

4. According to claim 3, a method for characterizing causal decoupling of non-uniform superpixel graph neural network data is characterized in that: The independence loss function is specifically expressed as follows: ； In the formula, Represents false features The n The feature vector of superpixels; express and The Hilbert-Schmidt independence criterion between .

5. According to claim 4, a method for characterizing causal decoupling of non-uniform superpixel graph neural network data is characterized in that: The backdoor adjustment loss function is specifically expressed as follows: ； In the formula, Represents false features The m The feature vector of superpixels; Represents the use of graph convolutional networks predict Type of fault; express The type of failure predicted; Indicates the use of normalization function softmax right Perform normalization processing; express and The cross entropy loss between .

6. The method for characterizing causal decoupling of non-uniform superpixel graph neural network data according to claim 5, characterized in that: There are two types of setting requirements, and the optimal model can be obtained by meeting either of them; one is: the number of training times reaches the set value; the other is: the model accuracy reaches the set accuracy.

7. A method for characterizing causal decoupling of non-uniform superpixel graph neural network data according to any one of claims 1 to 6, characterized in that: The specific content of step S1 is as follows: S11, obtaining fault data of the object to be tested, where the fault data of the object to be tested includes causal characteristics and their corresponding fault types; S12, using fast continuous wavelet transform to process the fault data of each object to be tested, and obtain processing results; S13. Performing a visualization operation on the processing result to convert it into a corresponding time-frequency image.

8. The method for characterizing causal decoupling of non-uniform superpixel graph neural network data according to claim 7, characterized in that: The time-frequency image is segmented into superpixels using a simple linear iterative clustering algorithm to obtain a corresponding number of superpixels.

9. The method for characterizing causal decoupling of non-uniform superpixel graph neural network data according to claim 8, characterized in that: The convolutional neural network is used to extract the feature vector of each superpixel from the current time-frequency image, and the feature vector It is expressed as: ; in, Represents the grayscale mean in the current superpixel; Represents the mean value of the red channel, the green channel, and the blue channel in the current superpixel; Represents the texture mean in the current superpixel; Represents the area and perimeter of the current superpixel.

10. The method for characterizing causal decoupling of non-uniform superpixel graph neural network data according to claim 9, characterized in that: The relationship diagram corresponding to the current time-frequency image It is expressed as: ;in, represents a node, i.e., a superpixel; Represents the edge, that is, the connection relationship between superpixels; Indicates the hidden state.

Images