Wind turbine generator transmission chain online fault diagnosis method

Abstract

The invention belongs to the technical field of electromechanical equipment fault diagnosis, and discloses a wind turbine generator transmission chain online fault diagnosis method, which specifically comprises the following steps: S1, adopting an SCADA (Supervisory Control and Data Acquisition) system to collect environment, electrical and vibration multi-source monitoring parameters of a wind turbine generator in real time and pre-process the parameters, thereby effectively improving the accuracy and availability of data; s2, through the design of a multi-source feature association fusion module, complementarity of environment, electrical and vibration parameters is fully utilized, and the accuracy and reliability of fault diagnosis are improved; s3, a teacher-student network framework is constructed, online fault model construction and real-time diagnosis are achieved, the real-time performance and the intelligent level of fault diagnosis are improved, due to the application of a class balance sampling mechanism and a positive and negative combined learning strategy, the influence of continuous state label offset is eliminated, pseudo label noise interference in online samples is effectively reduced, and the real-time performance of fault diagnosis is improved. And the adaptive diagnosis capability of the model in a working condition environment dynamic change scene is improved.

Description

Technical Field

[0001] The invention belongs to the technical field of electromechanical equipment fault diagnosis, and in particular relates to an online fault diagnosis method for a transmission chain of a wind turbine generator set. Background Art

[0002] The wind turbine drive chain is a key component in wind power equipment with high failure rate, long downtime and high maintenance cost. Real-time and accurate health status monitoring is an important means to ensure the safe and efficient operation of wind power equipment. Wind power multi-source monitoring data fusion and intelligent diagnosis and prediction are important research directions in the field of wind power equipment health management. However, due to the complexity of wind power multi-source data and the dynamic and changeable operating environment, current research still faces many challenges and limitations in terms of information fusion adaptability and model generalization.

[0003] At present, there are many monitoring methods for wind turbine transmission systems, mainly vibration, oil sampling, speed, current, temperature analysis, etc. Domestic and foreign scholars have carried out a lot of exploration around mechanism modeling, signal processing, feature extraction, identification and prediction, and have achieved rich research results. With the rapid development of information theory and artificial intelligence, wind turbine status monitoring is developing towards multi-source information fusion diagnosis and intelligent identification and prediction, and the detection accuracy and reliability are constantly improving. However, due to the complex and dynamic changes in the operating conditions of wind turbines, it is difficult to collect effective fault data. Different methods have their limitations in application in engineering scenarios. There are generally problems such as limited information fusion, poor generalization, lack of real-time and adaptability. Therefore, it is urgent to carry out research on the health status monitoring of wind turbine transmission chains under complex dynamic environments, provide a scientific basis for fault warning and preventive maintenance, reduce unnecessary downtime and maintenance costs, and improve the operating efficiency and economic benefits of wind power equipment. Summary of the invention

[0004] The object of the present invention is to provide an online fault diagnosis method for a wind turbine transmission chain to solve the problems raised in the above background technology.

[0005] In order to achieve the above object, the present invention provides the following technical solution: a method for online fault diagnosis of a wind turbine transmission chain, the specific steps of which are as follows:

[0006] S1: Use the SCADA system to collect the environmental monitoring parameters, electrical monitoring parameters and vibration monitoring parameters of the wind turbine in real time, and perform data preprocessing on the collected monitoring data of the current time step;

[0007] S2: Design a multi-source feature association fusion module, input the online test samples into the multi-source feature association fusion module, calculate the cross-modal correlation of multi-source information, combine the complementary features of environmental, electrical, and vibration parameters, and calculate the fusion weight according to the feature importance to achieve weighted fusion of multi-source information, thereby obtaining the multi-source fusion feature Et ;

[0008] S3: Build an online fault model based on the teacher-student network framework, input the fusion features into the constructed teacher-student network model, and use the teacher network to output the online diagnosis prediction results The teacher model and student models The network structure is the same, both are improved ResNet18 networks. The improved ResNet18 networks are pre-trained using historical monitoring data, and the pre-trained parameters are used to initialize the teacher-student network.

[0009] S4: collect current monitoring data of the wind turbine in real time and input it into the online diagnosis model, and output the online diagnosis results of the wind turbine transmission chain;

[0010] S5: Use class-balanced sampling mechanism to output results based on the current online output of the teacher model Select reliability samples from the current samples and store them in the memory base according to the "first in, first out" principle to form a balanced label distribution sample set and eliminate the impact of continuous state label offset;

[0011] S6: Design a positive-negative joint learning strategy, combining positive contrast learning of reliability samples in the memory bank and negative learning optimization of current input samples, apply positive-negative joint learning and teacher-student model knowledge alignment loss function, guide the student model to update online adaptively, and reduce the interference of pseudo-label noise in online samples;

[0012] S7: The framework implements the knowledge alignment of the teacher-student model, which can be achieved through the teacher model Instruct and correct student models The update direction is used to alleviate the covariate offset problem under dynamic working conditions. At the same time, the student model parameters are updated to the teacher model through the exponential sliding average method to achieve online continuous learning and continuously accumulate fault diagnosis knowledge.

[0013] S8: Repeating steps S4, S5, S6 and S7 can realize real-time online state identification and model adaptive update of the wind turbine transmission chain.

[0014] Preferably, the multi-source feature association fusion module in step S2 includes two sub-modules, wherein the sensor feature reshaping module based on convolutional neural network uses a convolutional layer with a flexible filter structure and a resampled pooling layer to map the original multi-source data to a low-dimensional space of the same size, unify the heterogeneous feature input dimensions, and retain the uniqueness of each sensor information and enhance the feature expression capability; the multi-source feature association fusion module applies a cross-modal attention mechanism to autonomously analyze the correlation between multi-source sensor features, obtain complementary information and discard redundant information according to the degree of correlation, and finally, according to the requirements of subsequent diagnostic tasks, calculate the attention weight of each related complementary information source, and improve the perception performance by dynamically combining multi-source related complementary features through weights.

[0015] Preferably, in order to adapt to one-dimensional input data, all filters of the improved ResNet18 network described in step S3 are replaced by one-dimensional convolution kernels, the bottom fully connected layer is replaced by a bottleneck layer containing 128 neurons, and each convolution layer uses batch normalization and dropout operations, and its classifier consists of a fully connected layer and a subsequent softmax operation.

[0016] Preferably, the model pre-training method in step S3 is standard supervised learning training, the loss function adopts the cross entropy function, and the learning rate is 1×10 -3 , the sampling stochastic gradient descent algorithm is used to tune the model parameters.

[0017] Preferably, the class-balanced sampling mechanism in step S5 initializes the memory bank, and based on the current prediction results of the teacher model, uses an adaptive threshold to select reliability samples for each state class and store them in the memory bank, thereby ensuring that the model learns from a uniform class distribution and avoiding local overfitting or model collapse. Specifically, a combination of confidence and entropy of the teacher model prediction results is used as a benchmark metric, and the adaptive threshold is defined as its mean for the current batch of samples:

[0018]

[0019] in, is the target sample, C is the number of fault categories, is the teacher model output, Output the cth element of the probability vector for the teacher model, i.e., the teacher model predicts the sample is the probability of type c failure, For sample The difference between the prediction confidence and entropy, B is the number of batch samples, w t is the global threshold;

[0020] Global threshold w tIt reflects the overall learning status of the model for the current test data. In order to consider the adaptation differences between classes, a class threshold is introduced for each state class to ensure that the fault class with low certainty is not ignored;

[0021]

[0022] Among them, l is the indicator function, which is used to judge Predict whether the result belongs to the cth health class, is the threshold of category c;

[0023] By modulating the threshold of each class through the global threshold, the final adaptive threshold of each class is as follows;

[0024]

[0025] in, is the class threshold set, Adaptive threshold for each class;

[0026] The adaptive threshold method is used to divide each type of sample in the current batch into reliable and unreliable groups. The partitioning rules are as follows:

[0027]

[0028] in, is the c-th sample set, is the reliable sample set in class c, is the unreliable sample set in class c.

[0029] Preferably, the positive-negative joint learning strategy in step S6 adopts two different training strategies to utilize reliable and unreliable samples to ensure that the model is effectively optimized in the adaptation phase. Specifically, the memory library samples have more reliable pseudo labels generated by the teacher model, so these samples help the model benefit from positive learning. Contrastive learning is applied to these reliability samples so that features from the same class are aggregated and features from different classes are separated. First, the student model feature extractor is used to obtain embedded features. Then the cosine similarity between the two features is calculated, and the loss function of forward learning is expressed as:

[0030]

[0031] in, is the cosine similarity between two eigenvectors, E is the mean sign, is the embedding feature of the c-th reliable sample, L pst is the positive learning loss;

[0032] In order to make full use of the data stream samples of the current time step, complementary labels are generated for training, and negative learning is performed on reliable samples and unreliable samples. is randomly selected from categories other than those predicted by the teacher model. The loss function of negative learning is expressed as:

[0033]

[0034] in, Complementary tags The cth element in X t is the target sample set, The teacher model outputs the cth element of the probability vector, L neg is the negative learning loss.

[0035] Preferably, the teacher-student model knowledge alignment in step S7 further enhances the consistency of teacher-student model predictions through a symmetric cross entropy loss function, and the formula of the symmetric cross entropy loss function is:

[0036]

[0037] Among them, M is the class-balanced memory sample set, and are the prediction probabilities of the cth class of the teacher model and the student model, respectively, alg is the symmetric cross entropy loss.

[0038] Preferably, the student model parameters in step S7 are calculated by the total loss function L total The stochastic gradient descent algorithm is applied to optimize the update, with a learning rate of 1×10 -2 , the total loss function L total The formula is:

[0039] L total =L alg +λ 1 L pst +λ 2 L neg

[0040] Among them, λ 1 and λ 2 is the trade-off parameter of the corresponding loss term, L total The total training loss for the model.

[0041] Preferably, the teacher model parameters in step S7 are updated by exponential sliding average of the student model parameters, the smoothing factor β is set to 0.99, and the exponential sliding average formula is:

[0042] θ′ t+1 =βθ′ t+(1-β)θ t+1 Among them, the smoothing factor β is set to 0.99, θ′ t are the parameters of the teacher model at test time step t, θ t+1 are the parameters of the student model at test time step t+1, and the real-time online diagnosis results of each test time step t are predicted by the teacher model.

[0043] The beneficial effects of the present invention are as follows:

[0044] Multi-source monitoring data is collected in real time through the SCADA system and preprocessed, which effectively improves the accuracy and availability of the data; the design of the multi-source feature association fusion module makes full use of the complementarity of environmental, electrical and vibration parameters to improve the accuracy and reliability of fault diagnosis; the construction of the teacher-student network framework realizes the construction and real-time diagnosis of online fault models, and improves the real-time and intelligent level of fault diagnosis; the application of the positive and negative joint learning strategy effectively reduces the interference of pseudo-label noise in online samples and improves the adaptive update capability of the model; in addition, by implementing the knowledge alignment of the teacher-student model, the problem of covariate offset under dynamic conditions is alleviated, online continuous learning is realized, fault diagnosis knowledge is continuously accumulated, and the generalization ability and long-term operation stability of the model are improved; it provides an efficient, accurate and reliable solution for the fault diagnosis of wind turbine transmission chains. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 It is a schematic diagram of the process of the present invention;

[0046] Figure 2 This is a schematic diagram of the structure of the multi-source feature association fusion module of the present invention;

[0047] Figure 3 This is the improved ResNet18 network structure diagram of the present invention;

[0048] Figure 4 Schematic diagram of the online fault diagnosis model based on the teacher-student framework of the present invention. DETAILED DESCRIPTION

[0049] 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.

[0050] like Figures 1 to 4 As shown, the embodiment of the present invention provides a method for online fault diagnosis of a wind turbine transmission chain, and the specific steps are as follows:

[0051] S1: Use the SCADA system to collect the environmental monitoring parameters, electrical monitoring parameters and vibration monitoring parameters of the wind turbine in real time, and perform data preprocessing on the collected monitoring data of the current time step;

[0052] S2: Design a multi-source feature association fusion module, input the online test samples into the multi-source feature association fusion module, calculate the cross-modal correlation of multi-source information, combine the complementary features of environmental, electrical, and vibration parameters, and calculate the fusion weight according to the feature importance to achieve weighted fusion of multi-source information, thereby obtaining the multi-source fusion feature E t ;

[0053] S3: Build an online fault model based on the teacher-student network framework, input the fusion features into the constructed teacher-student network model, and use the teacher network to output the online diagnosis prediction results The teacher model and student models The network structure is the same, both are improved ResNet18 networks. The improved ResNet18 networks are pre-trained using historical monitoring data, and the pre-trained parameters are used to initialize the teacher-student network.

[0054] S4: collect current monitoring data of the wind turbine in real time and input it into the online diagnosis model, and output the online diagnosis results of the wind turbine transmission chain;

[0055] S5: Use class-balanced sampling mechanism to output results based on the current online output of the teacher model Select reliability samples from the current samples and store them in the memory base according to the "first in, first out" principle to form a balanced label distribution sample set and eliminate the impact of continuous state label offset;

[0056] S6: Design a positive-negative joint learning strategy, combine positive contrast learning of reliability samples in the memory bank with negative learning optimization of current input samples, guide the online update of the student model, reduce the interference of pseudo-label noise in online samples, apply positive-negative joint learning and teacher-student model knowledge alignment loss function, and adaptively update the student model parameters;

[0057] S7: The framework implements the knowledge alignment of the teacher-student model, which can be achieved through the teacher model Instruct and correct student models The update direction is used to alleviate the covariate offset problem under dynamic working conditions. At the same time, the student model parameters are updated to the teacher model through the exponential sliding average method to achieve online continuous learning and continuously accumulate fault diagnosis knowledge.

[0058] S8: Repeating steps S4, S5, S6 and S7 can realize real-time online state identification and model adaptive update of the wind turbine transmission chain.

[0059] The SCADA system is used to collect environmental, electrical and vibration monitoring parameters and perform data preprocessing. Data preprocessing includes removing outliers by setting upper and lower thresholds, denoising, and normalizing various sensor data. The data range is mapped to the [0,1] interval to eliminate the dimensional differences between different physical quantities, and the input samples are divided into X at equal time intervals. t ; Design a multi-source feature association fusion module to realize the dynamic weight combination of cross-modal multi-source related complementary features; construct an online fault model based on the teacher-student network framework, alleviate the continuous covariate shift through the knowledge alignment of the teacher-student model, and continuously accumulate diagnostic knowledge; through the class balanced sampling mechanism, the model learns from a uniform label distribution to eliminate the impact of continuous state label drift; design a positive and negative joint learning strategy to guide the online adaptive update and optimization of the model; input the online collected wind power real-time monitoring signal into the model to realize the online fault diagnosis of the wind turbine transmission chain and the adaptive update of the model, thereby improving the accuracy, real-time and adaptability of fault detection under the dynamic changes of wind turbine operating conditions.

[0060] Among them, the multi-source feature association fusion module in step S2 includes two sub-modules. The sensor feature reshaping module based on convolutional neural network uses a convolutional layer with a flexible filter structure and a resampling pooling layer to map the original multi-source data to a low-dimensional space of the same size, unify the heterogeneous feature input dimensions, and retain the uniqueness of each sensor information and enhance the feature expression capability; the multi-source feature association fusion module uses a cross-modal attention mechanism to autonomously analyze the correlation between multi-source sensor features, obtain complementary information and discard redundant information according to the degree of correlation, and finally, according to the requirements of subsequent diagnostic tasks, calculate the attention weight of each related complementary information source, and improve the perception performance by dynamically combining multi-source related complementary features through weights.

[0061] By setting up the sensor feature reconstruction module and the multi-source feature association fusion module, the synergy of the two sub-modules not only optimizes the processing capability of the overall system, but also reduces the computational complexity, improves the real-time performance and response speed of the system, and enables the multi-source feature association fusion module to process multi-source data more efficiently and meet the needs of practical applications.

[0062] In step S3, the improved ResNet18 network is adapted to one-dimensional input data. All filters are replaced by one-dimensional convolution kernels, the bottom fully connected layer is replaced by a bottleneck layer containing 128 neurons, and each convolution layer uses batch normalization and dropout operations. Its classifier consists of a fully connected layer and a subsequent softmax operation.

[0063] By replacing filters with one-dimensional convolution kernels, the network is able to directly process one-dimensional input data.

[0064] The model pre-training method in step S3 is standard supervised learning training, the loss function is the cross entropy function, and the learning rate is 1×10 -3 , the sampling stochastic gradient descent algorithm is used to tune the model parameters.

[0065] By adopting the cross entropy loss function, the difference between the model prediction results and the true labels can be effectively measured. By minimizing the cross entropy loss, the model can gradually optimize its prediction ability and improve the accuracy of classification.

[0066] Among them, the class-balanced sampling mechanism in step S5 initializes the memory bank, and based on the current prediction results of the teacher model, uses an adaptive threshold to select reliability samples for each state class and store them in the memory bank, so as to ensure that the model learns from a uniform class distribution and avoids local overfitting or model collapse. Specifically, the combination of confidence and entropy of the teacher model prediction results is used as the benchmark metric, and the adaptive threshold is defined as its mean for the current batch of samples:

[0067]

[0068] in, is the target sample, C is the number of fault categories, is the teacher model output, Output the cth element of the probability vector for the teacher model, i.e., the teacher model predicts the sample is the probability of type c failure. For sample The difference between the prediction confidence and entropy, B is the number of batch samples, w t is the global threshold;

[0069] Global threshold w t It reflects the overall learning status of the model for the current test data. In order to consider the adaptation differences between classes, a class threshold is introduced for each state class to ensure that the fault class with low certainty is not ignored;

[0070]

[0071] Among them, l is the indicator function, which is used to judge Predict whether the result belongs to the cth health class, is the threshold of category c.

[0072] By modulating the threshold of each class through the global threshold, the final adaptive threshold of each class is as follows;

[0073]

[0074] in, is the class threshold set, Adaptive threshold for each class;

[0075] The adaptive threshold method is used to divide each type of sample in the current batch into reliable and unreliable groups. The partitioning rules are as follows:

[0076]

[0077] in, is the c-th sample set, is the reliable sample set in class c, is the unreliable sample set in category c;

[0078] The memory is updated by sampling the obtained reliable samples, storing class-balanced samples in a first-in-first-out manner. At each time step, the sampling algorithm is summarized as follows:

[0079]

[0080] Among them, the positive and negative joint learning strategy in step S6 adopts two different training strategies to utilize reliable and unreliable samples to ensure that the model is effectively optimized in the adaptation stage. Specifically, the memory library samples have more reliable pseudo labels generated by the teacher model, so these samples help the model benefit from positive learning. Contrastive learning is applied to these reliability samples to aggregate features from the same class and separate features from different classes. First, the student model feature extractor is used to obtain embedded features. Then the cosine similarity between the two features is calculated, and the loss function of forward learning is expressed as:

[0081]

[0082] in, is the cosine similarity between two eigenvectors, E is the mean sign, is the embedding feature of the c-th reliable sample, L pst is the positive learning loss.

[0083] In order to make full use of the data stream samples of the current time step, complementary labels are generated for training, and negative learning is performed on reliable samples and unreliable samples. is randomly selected from categories other than those predicted by the teacher model. The loss function of negative learning is expressed as:

[0084]

[0085] in, Complementary tags The cth element in X t is the target sample set, The teacher model outputs the cth element of the probability vector, L neg is the negative learning loss.

[0086] Among them, the teacher-student model knowledge alignment in step S7 further enhances the consistency of teacher-student model prediction through the symmetric cross entropy loss function. The formula of the symmetric cross entropy loss function is:

[0087]

[0088] Among them, M is the class-balanced memory sample set, and are the prediction probabilities of the cth class of the teacher model and the student model, respectively, alg is the symmetric cross entropy loss.

[0089] The symmetric cross entropy loss function can more effectively measure the difference between the predictions of the teacher model and the student model by dealing with inconsistent classification results. This difference measurement method is more comprehensive and accurate, which helps to improve the consistency of the predictions of the two.

[0090] Among them, in step S7, the student model parameters are calculated by the total loss function L total The stochastic gradient descent algorithm is applied to optimize the update, with a learning rate of 1×10 -2 , the total loss function L total The formula is:

[0091] L total =L alg +λ 1 L pst +λ 2 L neg

[0092] Among them, λ 1 and λ 2 is the trade-off parameter of the corresponding loss term, L total The total training loss for the model.

[0093] Among them, in step S7, the teacher model parameters are updated by the exponential moving average method of the student model parameters, the smoothing factor β is set to 0.99, and the exponential moving average method formula is:

[0094] θ′ t+1 =βθ′ t +(1-β)θ t+1

[0095] Among them, the smoothing factor β is set to 0.99, θ′ t are the parameters of the teacher model at test time step t, θ t+1 are the parameters of the student model at test time step t+1, and the real-time online diagnosis results of each test time step t are predicted by the teacher model.

[0096] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device.

[0097] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for online fault diagnosis of a wind turbine transmission chain, characterized in that: The specific steps are as follows: S1: Use the SCADA system to collect the environmental monitoring parameters, electrical monitoring parameters and vibration monitoring parameters of the wind turbine in real time, and perform data preprocessing on the collected monitoring data of the current time step; S2: Design a multi-source feature association fusion module, input the online test samples into the multi-source feature association fusion module, calculate the cross-modal correlation of multi-source information, combine the complementary features of environmental, electrical, and vibration parameters, and calculate the fusion weight according to the feature importance to achieve weighted fusion of multi-source information, thereby obtaining the multi-source fusion feature E t ; S3: Build an online fault diagnosis model based on the teacher-student network framework, input the fusion features into the constructed teacher-student network model, and use the teacher network to output the online diagnosis prediction results The teacher model and student models The network structure is the same, both are improved ResNet18 networks. The improved ResNet18 networks are pre-trained using historical monitoring data, and the pre-trained parameters are used to initialize the teacher-student network. S4: collect current monitoring data of the wind turbine in real time and input it into the online diagnosis model, and output the online diagnosis results of the wind turbine transmission chain; S5: Use class-balanced sampling mechanism to output results based on the current online output of the teacher model Select reliability samples from the current samples and store them in the memory bank M on a "first in, first out" basis to form a balanced label distribution sample set and eliminate the impact of continuous state label offset; S6: Design a positive-negative joint learning strategy, combining positive contrast learning of reliability samples in the memory bank and negative learning optimization of current input samples, apply positive-negative joint learning and teacher-student model knowledge alignment loss function, guide the student model to update online adaptively, and reduce the interference of pseudo-label noise in online samples; S7: Implement knowledge alignment between teacher and student models, which can be achieved through the teacher model Instruct and correct student models The update direction can alleviate the covariate offset problem under dynamic working conditions. The model refers to updating the student model parameters to the teacher model through the exponential sliding average method at the same time, realizing online continuous learning and continuously accumulating fault diagnosis knowledge; S8: Repeating steps S4, S5, S6 and S7 can realize real-time online status identification and diagnosis of the wind turbine transmission chain and adaptive model update.

2. A method for online fault diagnosis of a wind turbine transmission chain according to claim 1, characterized in that: The multi-source feature association fusion module in step S2 includes two submodules, wherein the sensor feature reconstruction module based on convolutional neural network uses a convolutional layer with a flexible filter structure and a resampled pooling layer to map the original multi-source data to the same low-dimensional space, unify the heterogeneous feature input dimensions, and retain the uniqueness of each sensor information and enhance the feature expression capability; The multi-source feature association fusion module applies the cross-modal attention mechanism to autonomously analyze the correlation between multi-source sensor features, obtain complementary information and discard redundant information based on the degree of correlation, and finally, according to the requirements of subsequent diagnostic tasks, calculate the attention weight of each relevant complementary information source, and dynamically combine multi-source relevant complementary features through weights to improve perception performance.

3. The method for online fault diagnosis of a wind turbine transmission chain according to claim 1, characterized in that: In the improved ResNet18 network described in step S3, in order to adapt to one-dimensional input data, all filters are replaced by one-dimensional convolution kernels, the bottom fully connected layer is replaced by a bottleneck layer containing 128 neurons, and each convolution layer uses batch normalization and dropout operations. Its classifier consists of a fully connected layer and a subsequent softmax operation.

4. The method for online fault diagnosis of a wind turbine transmission chain according to claim 1, characterized in that: The model pre-training method in step S3 is standard supervised learning training, the loss function adopts the cross entropy loss function, and the learning rate is 1×10 -3 , the stochastic gradient descent algorithm is used to tune the model parameters.

5. The method for online fault diagnosis of a wind turbine transmission chain according to claim 1, characterized in that: The class-balanced sampling mechanism described in step S5 initializes the memory bank and selects reliability samples for each state class based on the current prediction results of the teacher model using an adaptive threshold and stores them in the memory bank, thereby ensuring that the model learns from a uniform class distribution and avoids local overfitting or model collapse. Specifically, a combination of confidence and entropy of the teacher model prediction results is used as a baseline metric, and the adaptive threshold is defined as its mean for the current batch of samples: in, is the target sample, C is the number of fault categories, is the teacher model output, Output the cth element of the probability vector for the teacher model, i.e., the teacher model predicts the sample is the probability of type c failure, For sample The difference between the prediction confidence and entropy, B is the number of batch samples, w t is the global threshold; Global threshold w t It reflects the overall learning status of the model for the current test data. In order to consider the adaptation differences between classes, a class threshold is introduced for each state class to ensure that the fault class with low certainty is not ignored; Among them, l is the indicator function, which is used to judge Predict whether the result belongs to the cth health class, is the threshold of category c; By modulating the threshold of each class through the global threshold, the final adaptive threshold of each class is as follows; in, is the class threshold set, Adaptive threshold for each class; The adaptive threshold method is used to divide each type of sample in the current batch into reliable and unreliable groups. The partitioning rules are as follows: in, is the c-th sample set, is the reliable sample set in class c, is the unreliable sample set in class c.

6. The method for online fault diagnosis of a wind turbine transmission chain according to claim 1, characterized in that: The positive-negative joint learning strategy described in step S6 uses two different training strategies to utilize reliable and unreliable samples to ensure that the model is effectively optimized during the adaptation phase. Specifically, the memory samples have more reliable pseudo labels generated by the teacher model, so these samples help the model benefit from positive learning. Contrastive learning is applied to these reliability samples, so that features from the same class are aggregated and features from different classes are separated. First, the student model feature extractor is used to obtain embedded features. Then the cosine similarity between the two features is calculated, and the loss function of forward learning is expressed as: in, is the cosine similarity between two eigenvectors, E is the mean sign, is the embedding feature of the c-th reliable sample, L pst is the positive learning loss; In order to make full use of the data stream samples of the current time step, complementary labels are generated for training, and negative learning is performed on reliable samples and unreliable samples. is randomly selected from categories other than those predicted by the teacher model. The loss function of negative learning is expressed as: in, Complementary tags The cth element in X t is the target sample set, The teacher model outputs the cth element of the probability vector, L neg is the negative learning loss.

7. The method for online fault diagnosis of a wind turbine transmission chain according to claim 1, characterized in that: The teacher-student model knowledge alignment described in step S7 further enhances the consistency of teacher-student model predictions through a symmetric cross entropy loss function. The formula of the symmetric cross entropy loss function is: Among them, M is the class-balanced memory sample set, and are the prediction probabilities of the cth class of the teacher model and the student model, respectively, alg is the symmetric cross entropy loss.

8. The method for online fault diagnosis of a wind turbine transmission chain according to claim 1, characterized in that: The student model parameters in step S7 are calculated by the total loss function L total The stochastic gradient descent algorithm is applied to optimize the update, with a learning rate of 1×10 -2 , the total loss function L total The formula is: L total =L alg +λ1L pst +λ2L neg Among them, λ1 and λ2 are the trade-off parameters of the corresponding loss terms, L total The total training loss for the model.

9. The method for online fault diagnosis of a wind turbine transmission chain according to claim 1, characterized in that: The teacher model parameters in step S7 are updated by the exponential moving average method of the student model parameters. The formula of the exponential moving average method is: θ′ t+1 =βθ′ t +(1-β)θ t+1 Among them, the smoothing factor β is set to 0.99, θ′ t are the parameters of the teacher model at test time step t, θ t+1 are the parameters of the student model at test time step t+1, and the real-time online diagnosis results of each test time step t are predicted by the teacher model.

Images