A Weighted Representation Learning Based Feature Extraction Model for Multivariate Time Series Data in the Wind Turbine PHM System

Abstract

Feature extraction is essential for accurate knowledge-based prognostics and health management (PHM) of the wind turbine system. As a classic solution of knowledge-based PHM, representation learning faces the challenges of long sequence length, missing values, and insufficient labeling when dealing with multivariate time series wind power data. This paper proposes a weighted representation learning-based feature extraction model to address the above problems simultaneously. Firstly, we design a novel self-attentive convolutional autoencoder (SACAE) model which can automatically extract high-level compressed feature matrix representations from long-length wind power time series, thereby retaining more necessary information. Then a missing-weighted dependent wild bootstrapped maximum mean discrepancy (MDMMD) loss function is proposed to capture the temporal dependence, numerical similarity, and distribution similarity of samples and avoid the influence of these missing values. Finally, we present the wind turbine PHM framework to improve the utilization of unlabeled data by using them to train the MDMMD-SACAE model and then the limited labeled data can get more accurate feature representation through the trained MDMMD-SACAE model. Experimental studies on three PHM cases verify the effectiveness and robustness of the proposed approach in the wind turbine system.