Research On Domain Adaptation Fault Diagnosis Of Wind Power Machinery Transmission System Based On Optimal Transport Theory And Deep Learning

The wind power transmission system has a pivotal position in the wind power system.The quality of the wind power transmission system directly affects the efficiency and quality of power generation and the safety of the power grid.The transmission system（bearings,gearbox）of the wind power unit is due to the variable working environment（wind speed,Temperature,humidity,etc.）make it prone to various faults,so it is particularly important to seek a stable and reliable fault diagnosis method to ensure its safe operation.With the advent of the big data era and the vigorous development of deep learning technology,fault diagnosis technology based on deep learning has received extensive attention from researchers.This article focuses on the fault diagnosis of wind power transmission system under different working conditions.The main contents are as follows:In view of the difference in the distribution of the training data set and the target test data caused by the changes in the load and environmental factors during the operation of the wind power transmission system,the generalization of the model is weakened,a multi-layer adaptation network based on the optimal transport（MAOT）is established,and the stack Auto-encoding is used to extract the abstract information in the data and calculates the difference between the source domain（training set）and the target domain（test set）based on the optimal transport theory,based on this,reduces the difference between the source domain features and the target domain features This effectively solves the problem that the characteristics of the source domain and the target domain do not obey the same distribution,which leads to the degradation of the performance of the diagnosis algorithm.Experiments show that the proposed method can effectively improve the diagnostic performance of the diagnostic algorithm under different working conditions and the selection of the self-encoding structure and its hyper-parameters.Aiming at the two main shortcomings of the existing fault diagnosis methods based on domain adaptation: 1)It is difficult to accurately measure and estimate the difference between the source domain and the target domain;2)Only the difference in the feature space is considered,and the difference in label space is not considered.An optimal transport based deep domain adaptation is proposed.the auto-encoder network extracts the discriminative features from the raw data,and then searches the transport plan between the source domain and the target domain based on a predefined cost matrix and minimizes the joint distribution difference of the feature and label space based on the optimal transport theory to extract domain invariant representation features,thereby further improving the diagnostic performance of the model when cross-domain.Experimental analysis verifies the best choice of model hyperparameters,and intuitively explains the proposed model in the feature space.transferability.Experimental results show that the method is superior to existing machine learning and domain adaptive fault diagnosis methods in terms of classification accuracy and generalization ability.The existed domain adaptation fault diagnosis methods only work under the assumption of closed set,and open set is more common in the actual wind power operating environment,which means that unknown faults will appear in the testing processing.Aiming at this problem,a sparse auto-encoder based adversarial open set domain adaptation（SAOSDA）model is proposed.Extracting the sparse features by constructing sparse auto-encoder to reduce the redundancy in the data.Utilizing the adversarial learning between the feature generator（sparse auto-encoder）and the label predictor to align the known target features with the known source features to compensate for the domain shift and separate the known target features from the unknown target features.These features are used by the improved Softmax based label predictor to classify target samples and recognize unknown target samples precisely.The experimental results on the real bearing data set prove the superiority of this method and can greatly improve the ability of recognize unknown faults in the target sample.