Wind power generation system fault diagnosis method and system based on multi-source data fusion

Abstract

The invention relates to the field of wind power generation system fault diagnosis, in particular to a wind power generation system fault diagnosis method and system based on multi-source data fusion. The method comprises the following steps: acquiring an original vibration signal in a wind power generation transmission system, and preprocessing the original vibration signal to obtain a two-dimensional frequency diagram of an original domain data sample; constructing a residual network based on separation attention mechanism optimization as a classifier, and obtaining a migration diagnosis teacher model for a target domain through a conditional domain adversarial migration method; a mask information entropy knowledge distillation strategy is adopted, an information entropy graph output by the teacher model is calculated and multiplied by a target area mask, and refined knowledge is migrated to a lightweight student model; and inputting the to-be-tested data sample of the target domain into the lightweight student model to obtain a diagnosis result. The invention aims to provide a wind power generation system fault diagnosis method and system based on multi-source data fusion, and solves the problems that fault data of a wind power generation transmission system is scarce and model calculation resources are limited.

Description

Technical Field

[0001] This invention relates to the field of fault diagnosis of wind power generation systems, and in particular to a method and system for fault diagnosis of wind power generation systems based on multi-source data fusion. Background Technology

[0002] With the deepening implementation of the "dual-carbon" strategy, the penetration rate of wind power in the power system has significantly increased, becoming a core support for energy transformation. As the core energy conversion link of wind turbine units, the wind turbine transmission system plays a crucial role in efficiently transferring the mechanical energy captured by wind energy to the generator. The reliability of this system directly affects the unit's output characteristics and the grid frequency stability. A failure in the transmission system can not only lead to unplanned outages of individual units but may also trigger regional active power fluctuations. However, because the transmission system operates under complex conditions of high torque and variable speed for extended periods, its mechanical and electrical coupling components are prone to progressive damage. Existing monitoring methods largely rely on mechanical vibration analysis, making it difficult to capture early changes in electrical characteristics, resulting in delayed fault warnings and threatening the safe and stable operation of the power grid.

[0003] Existing fault diagnosis methods face multiple limitations. Traditional physical models or signal processing methods heavily rely on expert experience to design features, resulting in poor transferability and difficulty in adapting to the complex characteristics of different gearbox models. While introducing machine learning methods has improved diagnostic capabilities, it is prone to overfitting under small sample conditions, exhibits poor generalization performance, and fails to effectively address the problem of data distribution differences across operating conditions or equipment.

[0004] In addition, existing deep learning models often have a large number of parameters and high computational complexity in pursuit of high accuracy, making them difficult to adapt to the limited computing resources of edge devices and hindering their large-scale application in distributed real-time monitoring of wind farms.

[0005] Therefore, how to design a fault diagnosis method and system for wind power generation systems based on multi-source data fusion that can solve the above-mentioned technical problems is a technical problem that needs to be solved. Summary of the Invention

[0006] To address the aforementioned issues, the present invention aims to provide a method and system for fault diagnosis of wind power generation systems based on multi-source data fusion. This method solves the problems of scarce fault data and limited model computation resources in wind power transmission systems, significantly improving the real-time performance and accuracy of edge-side diagnosis. It also effectively mines the document data characteristics of the State Grid Corporation of China, realizes the automatic generation of the State Grid document directory hierarchy, and improves the daily office efficiency of State Grid users.

[0007] In a first aspect, the present invention provides a method for fault diagnosis of wind power generation systems based on multi-source data fusion:

[0008] See Figure 1-4As shown, the scheme includes:

[0009] Step S1: Obtain the original vibration signal in the wind power transmission system and preprocess it to obtain a two-dimensional frequency diagram of the original domain data sample;

[0010] Step S2: Construct a residual network model optimized based on the separation attention mechanism as a classifier, adopt conditional domain adversarial transfer learning, and use original domain data samples and target domain data samples to train the residual network model to obtain a transfer diagnostic teacher model for the target domain.

[0011] Step S4: Compress the transfer diagnostic teacher model into a lightweight student model by masking information entropy knowledge distillation;

[0012] Step S5: Input the target domain test data sample into the lightweight student model to obtain the diagnostic results.

[0013] Furthermore, the preprocessing is as follows: each column of vibration signal in the original vibration signal is extracted using the pandas library, the sequence is segmented by using a window of length 1024 and an overlap rate of 0.5, and each small segment of signal is processed by continuous wavelet transformation to generate the time-frequency distribution of the one-dimensional time series, thereby obtaining the two-dimensional frequency domain map of various states in the collected dataset.

[0014] Furthermore, the residual model network includes:

[0015] Grouped convolution module, used to reduce the number of parameters;

[0016] The channel separation attention module is used to control fine-grained segmentation of features.

[0017] The process of reducing the number of parameters in the grouped convolution module includes:

[0018] In grouped convolution, the input channels are divided into... Groups, each group has 10 channels. Each group uses an independent convolution kernel. The complete result obtained after concatenation is:

[0019]

[0020] Its total number of parameters is:

[0021]

[0022] The number of parameters can be reduced to that of standard convolution by using grouped convolution. ;

[0023] The channel separation attention module further subdivides each convolutional group into... There are several sub-paths, and the number of channels in each sub-path further becomes... At the same time, an attention mechanism that can span convolutional groups is introduced, enabling the model to actively perform weighted fusion of features across convolutional groups.

[0024] Furthermore, the specific process of actively weighting and fusing features from the convolutional groups is as follows:

[0025] First, the sub-path features of each convolutional group are analyzed. Global pooling is performed to obtain context information, and the formula is as follows:

[0026]

[0027] in, This indicates a global pooling operation to obtain the average value calculated along the spatial dimension;

[0028] Then, the global context The input is fed into a fully connected layer to generate attention weights for each sub-path, using the following formula:

[0029]

[0030] in, Indicates a fully connected layer. Used for normalizing weights, Indicates the first In the nth convolutional group Attention weights for each subpath;

[0031] Then, attention weights are used. Weighted fusion of sub-path features from each convolutional group:

[0032]

[0033] in, It is the first The output features of each convolutional group;

[0034] Finally, component feature fusion is performed, combining the output features of all convolutional groups. By stitching along the channel dimension, the final fused feature is obtained:

[0035]

[0036] Meanwhile, the output of the separate attention module is added to the input features through a residual connection, and the calculation formula is as follows:

[0037]

[0038] Furthermore, in step S2, the residual network model is trained using the original domain data samples and the target domain data samples to obtain the transfer diagnostic teacher model for the target domain, specifically as follows:

[0039] By learning the loss function through the conditional domain adversarial transfer method, and by leveraging the joint optimization of the feature extractor, task classifier and domain discriminator, the global feature distribution and conditional distribution of the source domain and the target domain are aligned, class-specific information is preserved, and the weight parameters of the source domain fault diagnosis teacher model are transferred to the new residual network model to obtain the target domain fault diagnosis teacher model.

[0040] Using the data samples generated by the actual operation of the wind turbine gearbox, the target domain fault diagnosis model is trained. The parameters of the target domain fault diagnosis model are adjusted by the loss function to obtain the transfer diagnosis model.

[0041] Furthermore, the specific process of learning the loss function using the conditional domain adversarial transfer learning method, and optimizing it jointly with the feature extractor, task classifier, and domain discriminator, is as follows:

[0042] The formula for the conditional domain adversarial migration method is as follows:

[0043]

[0044] in, For mission losses, For domain confrontation losses, For regularization terms;

[0045] The task classifier optimizes classification performance on the source domain data, and its loss is cross-entropy:

[0046]

[0047] The domain discriminator is based on features and classifier output The joint representation, used to distinguish data between the source and target domains, uses a binary cross-entropy loss function, as shown in the following formula:

[0048]

[0049] Conditional adversarial networks introduce entropy minimization regularization. The specific formula for calculating the entropy of the target domain classifier output is as follows:

[0050]

[0051] Among them, entropy Defined as:

[0052]

[0053] Furthermore, in step S4, the specific formula for compressing the transfer diagnosis teacher model through masked information entropy knowledge distillation is as follows:

[0054]

[0055] in, It is text detection loss. It is the information entropy transfer loss. It is a pixel-by-pixel classification loss. yes The weight, yes The weight.

[0056] Furthermore, the formula for calculating the pixel-by-pixel classification loss is as follows:

[0057]

[0058] in, and The partitioned graphs representing the teacher network and student network are located on the following sides. The The probability values ​​of each category, This represents the total number of categories.

[0059] Secondly, this invention provides an application system for fault diagnosis of wind power generation systems based on multi-source data fusion:

[0060] See Figure 5 As shown, the scheme includes:

[0061] The acquisition module is used to acquire the raw vibration signals of various types of gears;

[0062] The training module is used to train the residual network model to obtain the transfer diagnostic teacher model for the target domain; at the same time, by adopting the mask information entropy knowledge distillation strategy, the refined knowledge obtained from the transfer diagnostic teacher model is transferred to the lightweight student model.

[0063] The diagnostic module is used to deploy the distilled lightweight student model to the wind turbine edge computing device, and input the target domain test data sample into the fault diagnosis model to obtain the diagnostic results.

[0064] Furthermore, it includes a storage medium and a processor; the storage medium is used to store instructions; the processor is used to operate according to the instructions to perform the method according to any one of claims 1 to 8.

[0065] The present invention has the following beneficial effects:

[0066] This invention constructs a lightweight residual network classifier based on a separate attention mechanism, combined with conditional domain adversarial transfer learning and mask information entropy knowledge distillation techniques. This significantly improves the accuracy of gearbox fault diagnosis in wind power transmission systems while effectively overcoming the dual bottlenecks of insufficient model generalization ability under small sample conditions and limited computing power of edge devices. Compared to traditional methods relying on manual feature design, this invention utilizes an improved residual network to adaptively extract frequency domain features of vibration signals, achieving superior diagnostic accuracy compared to traditional CNN and LSTM models. Addressing the scarcity of fault samples, a conditional domain adversarial transfer learning mechanism is introduced to align cross-domain feature distributions, significantly reducing reliance on labeled data while maintaining diagnostic accuracy across operating conditions. Furthermore, a knowledge distillation method based on mask information entropy knowledge transfer transfers the soft-label entropy knowledge of the teacher model to a lightweight student model, adapting to the real-time monitoring needs of edge devices in wind farms. Ultimately, this resolves the contradiction between computational resource constraints and real-time diagnostic requirements in large-scale wind turbine deployments. Attached Figure Description

[0067] Figure 1 This is a flowchart of the wind power generation system fault diagnosis method based on multi-source data fusion provided in Example 1;

[0068] Figure 2 This is a diagram of the residual block structure in the residual network of Example 1;

[0069] Figure 3 This is a structural diagram of the convolutional layer of the separate attention mechanism module in Example 1;

[0070] Figure 4 This is a structural diagram of the residual network model optimized by the separation attention mechanism provided in Example 1;

[0071] Figure 5 This is a schematic diagram of the wind power generation system fault diagnosis module with multi-source data fusion in Example 2. Detailed Implementation

[0072] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments:

[0073] Example 1

[0074] like Figures 1 to 4 As shown, a fault diagnosis method for wind power generation systems based on multi-source data fusion includes:

[0075] Step S1: Obtain the original vibration signal from the wind power transmission system; use continuous wavelet transform to generate the time-frequency distribution of the one-dimensional time series from the original vibration signal, and obtain the two-dimensional frequency domain map of the original domain data sample. The method includes:

[0076] The pandas library was used to extract each vibration signal. The sequence was segmented using a window of length 1024 with an overlap rate of 0.5. Continuous wavelet transform was used to process each small segment of the signal to obtain two-dimensional frequency domain plots of various states in the collected dataset.

[0077] Step S2: Construct a residual network optimized based on the separation attention mechanism as the backbone classifier. It includes a grouped convolution module and a channel separation attention module. The number of parameters is reduced by grouped convolution, and the cross-group feature interaction capability is enhanced by channel attention, thereby improving the feature discrimination in high-noise environments. Conditional domain adversarial transfer learning is adopted. The residual network model is trained using original domain data samples and target domain data samples to obtain a teacher model for the target domain.

[0078] Step S3: After transferring the knowledge learned from the source domain to the target domain, a teacher network for knowledge distillation is obtained. A knowledge distillation method based on masked information entropy knowledge transfer is used to distill the student model, enabling it to learn the classification of the teacher model on the wind turbine gearbox dataset. This reduces the number of parameters in the final model and improves its performance under computational resource constraints.

[0079] Step S4: Deploy the distilled lightweight student model to the wind turbine edge computing device, input the target domain test data sample into the lightweight student model to obtain diagnostic results. The fault types are Health, Root Crack, Missing Tooth, Root Crack, and Surface Wear.

[0080] In grouped convolution, the concepts of cardinality and grouping are introduced to divide the input channels into... Groups, each group has 10 channels. Each group uses an independent convolution kernel. The complete result obtained after concatenation is:

[0081]

[0082] Its total number of parameters is:

[0083]

[0084] It is evident that grouped convolution can reduce the number of parameters to that of standard convolution. .

[0085] The mathematical results of grouped convolution are completely consistent with those of regular convolution, which allows the accuracy to remain unchanged while significantly reducing the number of parameters.

[0086] Furthermore, the residual network optimized by the separate attention mechanism introduces an attention mechanism capable of spanning convolutional groups. While retaining the innovations of cardinality and grouping, each convolutional group is further subdivided into... Each sub-path is used to control fine-grained feature segmentation, and the number of channels in each sub-path is further increased to... Introducing a separate attention mechanism enables the model to actively perform weighted feature fusion on these convolutional groups. The steps include:

[0087] First, the sub-path features of each convolutional group are analyzed. Global pooling is performed to obtain context information, and the formula is as follows:

[0088]

[0089] in, This represents a global pooling operation, the purpose of which is to calculate the average value along the spatial dimension;

[0090] Then, the global context The input is fed into a fully connected layer to generate attention weights for each sub-path:

[0091]

[0092] in, Indicates a fully connected layer. Used for normalizing weights, Indicates the first In the nth convolutional group Attention weights for each subpath;

[0093] Then, attention weights are used. Weighted fusion of sub-path features from each convolutional group:

[0094]

[0095] in, It is the first The output features of each convolutional group;

[0096] Finally, component feature fusion is performed, combining the output features of all convolutional groups. By stitching along the channel dimension, the final fused feature is obtained:

[0097]

[0098] Meanwhile, the output of the separate attention module is added to the input features through a residual connection, and the calculation formula is as follows:

[0099]

[0100] Furthermore, a method for obtaining a transfer diagnostic teacher model for the target domain by training a residual network model optimized by a separation attention mechanism using source domain data samples and target domain data samples includes:

[0101] The source domain data sample set ImageNet-1k is a large-scale image dataset. The source domain data sample is used to train and debug a residual network model optimized by the separation attention mechanism to obtain the best source domain fault diagnosis teacher model. By using the conditional domain adversarial transfer method, with the joint optimization of feature extractor, task classifier and domain discriminator, the global feature distribution and conditional distribution of source domain and target domain are aligned, the category-specific information is preserved, and the weight parameters of source domain fault diagnosis teacher model are transferred to the new residual network model to obtain target domain fault diagnosis teacher model.

[0102] The original vibration signals corresponding to the target domain data samples are data generated in the actual operation of the wind turbine gearbox. The target domain fault diagnosis teacher model is trained using the target domain data samples. The parameters of the target domain fault diagnosis teacher model are adjusted through the conditional domain adversarial transfer learning mechanism loss function, and finally the transfer diagnosis teacher model is obtained.

[0103] The output of the joint representation feature extractor in the conditional domain adversarial transfer learning mechanism and the output of the task classifier Through outer product operation This is achieved. Based on this, the total loss of the conditional domain adversarial transfer learning mechanism includes the task loss. Domain confrontation loss Regularization term :

[0104]

[0105] The task classifier optimizes classification performance on the source domain data, and its loss is cross-entropy:

[0106]

[0107] Domain discriminators based on features and classifier output The joint representation, used to distinguish data between the source and target domains, uses a loss function of binary cross-entropy:

[0108]

[0109] To enhance the confidence of the target domain prediction, the conditional adversarial network introduces entropy minimization regularization to calculate the entropy of the target domain classifier output:

[0110]

[0111] Among them, entropy Defined as:

[0112]

[0113] Furthermore, according to the definition of Shannon entropy, the formula for calculating the entropy value corresponding to each pixel in the input image is as follows:

[0114]

[0115] in, Indicates that the segmentation graph is located in The information entropy value of each pixel is respectively used and This represents the information entropy of the teacher network and the student network. The segmentation map output by the detection head is represented by... and This represents the output of the teacher network and the student network.

[0116] Furthermore, the formula for calculating the loss function of training the student network is as follows:

[0117]

[0118] in, It is text detection loss. It is the information entropy transfer loss. It is a pixel-by-pixel classification loss. yes The weight, yes The weights, pixel-by-pixel classification loss The calculation formula is:

[0119]

[0120] in, and The partitioned graphs representing the teacher network and student network are located on the following sides. The The probability values ​​of each category, This represents the total number of categories.

[0121] Example 2

[0122] This embodiment provides a fault diagnosis system for wind power generation systems based on multi-source data fusion. The diagnosis system described in this embodiment can be applied to the diagnosis method described in Embodiment 1, including:

[0123] The acquisition module is used to acquire the original vibration signals of various types of gears; and to perform continuous wavelet transform on the original vibration signals to obtain a two-dimensional frequency domain diagram of the data samples.

[0124] The training module is used to construct a residual network model optimized by the separation attention mechanism. The residual network model is trained using source domain data samples and target domain data samples through the conditional domain adversarial transfer method to obtain a transfer diagnostic teacher model for the target domain. The masked information entropy knowledge distillation strategy is used to transfer the refined knowledge obtained from the teacher model to the lightweight student model.

[0125] The diagnostic module is used to deploy the distilled lightweight student model to the wind turbine edge computing device, and input the target domain test data sample into the fault diagnosis model to obtain the diagnostic results.

[0126] Furthermore, it includes a storage medium and a processor; the storage medium is used to store instructions; the processor is used to operate according to the instructions to execute the method described in Embodiment 1.

[0127] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0128] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0129] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0130] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0131] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A fault diagnosis method for wind power generation systems based on multi-source data fusion, characterized in that, Includes the following steps: Step S1: Obtain the original vibration signal in the wind power transmission system and preprocess it to obtain a two-dimensional frequency diagram of the original domain data sample; Step S2: Construct a residual network model optimized based on the separation attention mechanism as a classifier, adopt conditional domain adversarial transfer learning, and use original domain data samples and target domain data samples to train the residual network model to obtain a transfer diagnostic teacher model for the target domain. Step S3: Compress the transfer diagnostic teacher model into a lightweight student model by masking information entropy knowledge distillation; Step S4: Input the target domain test data sample into the lightweight student model to obtain the diagnostic results.

2. The method for fault diagnosis of wind power generation system based on multi-source data fusion according to claim 1, characterized in that, The preprocessing is as follows: each column of vibration signal in the original vibration signal is extracted using the pandas library, the sequence is segmented by using a window of length 1024 and an overlap rate of 0.5, and each small segment of signal is processed by continuous wavelet transformation to generate the time-frequency distribution of the one-dimensional time series, thus obtaining the two-dimensional frequency domain map of various states in the collected dataset.

3. The method for fault diagnosis of wind power generation system based on multi-source data fusion according to claim 1, characterized in that, The residual model network includes: Grouped convolution module, used to reduce the number of parameters; The channel separation attention module is used to control fine-grained feature segmentation; The process of reducing the number of parameters in the grouped convolution module includes: In grouped convolution, the input channels are divided into... Groups, each group has 10 channels. Each group uses an independent convolution kernel. The complete result obtained after concatenation is: ； Its total number of parameters is: ； The number of parameters can be reduced to that of standard convolution by using grouped convolution. ; The channel separation attention module further subdivides each convolutional group into... There are several sub-paths, and the number of channels in each sub-path further becomes... At the same time, an attention mechanism that can span convolutional groups is introduced, enabling the model to actively perform weighted fusion of features across convolutional groups.

4. The method for fault diagnosis of wind power generation system based on multi-source data fusion according to claim 3, characterized in that, The specific process of actively weighting and fusing features from convolutional groups is as follows: First, the sub-path features of each convolutional group are analyzed. Global pooling is performed to obtain context information, and the formula is as follows: ； in, This indicates a global pooling operation to obtain the average value calculated along the spatial dimension; Then, the global context The input is fed into the fully connected layer to generate attention weights for each sub-path, using the following formula: ； in, Indicates a fully connected layer. Used for normalizing weights, Indicates the first In the nth convolutional group Attention weights for each subpath; Then, attention weights are used. Weighted fusion of sub-path features from each convolutional group: ； in, It is the first The output features of each convolutional group; Finally, component feature fusion is performed, combining the output features of all convolutional groups. By stitching along the channel dimension, the final fused feature is obtained: ； Meanwhile, the output of the separate attention module is added to the input features through a residual connection, and the calculation formula is as follows: 。 5. The method for fault diagnosis of wind power generation system based on multi-source data fusion according to claim 1, characterized in that, In step S2, the residual network model is trained using the original domain data samples and the target domain data samples to obtain the transfer diagnostic teacher model for the target domain. Specifically: By learning the loss function through the conditional domain adversarial transfer method, and by leveraging the joint optimization of the feature extractor, task classifier and domain discriminator, the global feature distribution and conditional distribution of the source domain and the target domain are aligned, class-specific information is preserved, and the weight parameters of the source domain fault diagnosis teacher model are transferred to the new residual network model to obtain the target domain fault diagnosis teacher model. Using the data samples generated by the actual operation of the wind turbine gearbox, the target domain fault diagnosis model is trained. The parameters of the target domain fault diagnosis model are adjusted by the loss function to obtain the transfer diagnosis model.

6. The method for fault diagnosis of wind power generation system based on multi-source data fusion according to claim 5, characterized in that, The specific process of learning the loss function using the conditional domain adversarial transfer learning method, and optimizing it jointly with the feature extractor, task classifier, and domain discriminator, is as follows: The formula for the conditional domain adversarial migration method is as follows: ； in, For mission losses, For domain confrontation losses, For regularization terms; The task classifier optimizes classification performance on the source domain data, and its loss is cross-entropy: ； The domain discriminator is based on features and classifier output The joint representation, used to distinguish data between the source and target domains, uses a binary cross-entropy loss function, as shown in the following formula: ； Conditional adversarial networks introduce entropy minimization regularization. The specific formula for calculating the entropy of the target domain classifier output is as follows: ； Among them, entropy Defined as: 。 7. A method for fault diagnosis of a wind power generation system based on multi-source data fusion as described in claim 1, characterized in that, In step S3, the specific formula for compressing the transfer diagnosis teacher model through masked information entropy knowledge distillation is as follows: ； in, It is text detection loss. It is the information entropy transfer loss. It is a pixel-by-pixel classification loss. yes The weight, yes The weight.

8. The method for fault diagnosis of wind power generation system based on multi-source data fusion according to claim 7, characterized in that, The formula for calculating the pixel-by-pixel classification loss is as follows: ； in, and The partitioned graphs representing the teacher network and student network are located on the following sides. The The probability values ​​of each category, This represents the total number of categories.

9. An application system for fault diagnosis of wind power generation systems based on multi-source data fusion, characterized in that, include: The acquisition module is used to acquire the raw vibration signals of various types of gears; The training module is used to train the residual network model to obtain the transfer diagnostic teacher model for the target domain; at the same time, by adopting the mask information entropy knowledge distillation strategy, the refined knowledge obtained from the transfer diagnostic teacher model is transferred to the lightweight student model. The diagnostic module is used to deploy the distilled lightweight student model to the wind turbine edge computing device, and input the target domain test data sample into the fault diagnosis model to obtain the diagnostic results.

10. The application system for fault diagnosis of wind power generation system based on multi-source data fusion according to claim 8, characterized in that, The method includes a storage medium and a processor; the storage medium is used to store instructions; the processor is used to operate according to the instructions to perform the method according to any one of claims 1 to 8.

Images