Investigation Of Transfer Learning For Cross-turbine Diagnosis Of Wind Turbine Faults

With the gradual implementation of the "carbon peak" and "carbon neutrality" strategies,more and more wind turbines are installed and connected to the grid for power generation.Due to the complex and harsh installation and operation environment,various components of wind turbines are prone to malfunctions,and it is urgent to develop available fault diagnosis methods for wind turbines.At present,wind turbines are equipped with Supervisory Control and Data Acquisition(SCADA)to collect multivariable time series data when they leave the site.This data has the characteristics of sufficient,comprehensive information,and multivariable coupling,making it an excellent object for conducting fault diagnosis research in the wind power field.Due to geographical location,installation environment,seasonal changes,and other reasons,there is heterogeneity in data distribution among wind turbines,which leads to a lack of applicability of traditional models for newly built turbines.Therefore,it is urgent to study highly generalized models.In this paper,we consider the multivariate variable connection characteristics of SCADA samples,and comprehensively consider the spatiotemporal correlation,data imbalance and distribution heterogeneity,and aiming at the problem that the fault features contained in the wind turbine SCADA data are difficult to extract effectively and the distribution difference is too large,which results in the extremely poor generalization ability of the fault diagnosis model,so we proposes a single-unit fault diagnosis model based on the deep multi-scale residual attention convolution network(DMRACNN).What’s more,the cross-turbines fault diagnosis model based on global and local domain adaptation and that based on ensemble transfer learning solve the problem of feature extraction and distribution difference of SCADA data,and provide new technology for fault diagnosis of different wind turbines in the same wind farm.The proposed fault diagnosis model was analyzed through multiple comparative experiments to verify its generalization ability on real wind turbine SCADA datasets.The main proffer of the task is summarized as follows:(1)Research on fault diagnosis method of single turbine based on deep multi-scale residual attention convolution neural network(DMRACNN).Guided by Inception-Res Net,a deep multi-scale residual attention convolution neural network based on self-built basic convolution layer is proposed in view of the characteristics of SCADA sample variables with multiple variables and complex correlations.The innovation lies in the addition of residual module on the basis of multi-scale,the addition of receptive field and the retention of the universal information of deep features,which verifies the effectiveness of the model in single turbine diagnosis and direct application across turbines.(2)Research on fault diagnosis method of cross-wind turbines deep transfer based on domain adaptation.Aiming at the problem of weak generalization ability of the model caused by the large difference in the data distribution of different turbines,a fault diagnosis framework for cross-turbine deep migration based on global and local domain adaptation is proposed.The main idea of this framework is to map the feature instances extracted from the basic network into the feature space by combining a small amount of unmarked target wind turbine data based on the domain adaptation method,Minimize the distance measurement between source data and target data.The method is verified in the cross-transfer scenario of two real wind turbines,and the effectiveness and reliability of the feature transfer method are proved.(3)Research on cross-turbine fault diagnosis algorithm based on ensemble transfer learning.Aiming at the problem that the model is difficult to train and has poor performance due to the imbalance of data samples,a cross-wind turbine fault diagnosis model based on ensemble transfer learning is proposed directly from the original unbalanced data.From the perspective of instance transfer,the model combines transfer learning and ensemble learning,and uses the idea of weighted weights to highlight the samples in the source domain that are conducive to the correct classification of the target domain,and weighted average of multiple weak classifiers to effectively solve the problem of data imbalance and increase the generalization ability of the model.The model is verified on balanced data and unbalanced data,and its ability to deal with distribution differences and sample imbalance at the same time is proved.