Abstract
As clean and low-carbon energy, wind energy has attracted the attention of many countries. The main bearing in the transmission system of large-scale wind turbines (WTs) is the most important part. The research on the condition monitoring of the main bearing has received more attention from many scholars and the wind industry, and it has become a hot research topic. The existing research on the condition monitoring of the main bearing has the following drawbacks: (1) the existing research assigns the same weight to each condition parameter variable, and the model extracts features indiscriminately; (2) different historical time points of the condition parameter variable are given the same weight, and the influence degree of different historical time points on the current value is not considered; and (3) the existing literature does not consider the operating characteristics of WTs. Different operating conditions have different control strategies, which also determine which condition parameters are artificially controlled. Therefore, to solve the problems above, this paper proposes a novel method for condition monitoring of WT main bearings by applying the dual attention mechanism and Bi-LSTM, named Dual Attention-Based Bi-LSTM (DA-Bi-LSTM). Specifically, two attention calculation modules are designed to extract the important features of different input parameters and the important features of input parameter time series, respectively. Then, the two extracted features are fused, and the Bi-LSTM building block is utilized to perform pre-and post-feature extraction of the fused information. Finally, the extracted features are applied to reconstruct the input data. Extensive experiments verify the performance of the proposed method. Compared with the Bi-LSMT model without adding an attention module, the proposed model achieves 19.78%, 2.17%, and 18.92% improvement in MAE, MAPE, and RMSE, respectively. Compared with the Bi-LSTM model which only considers a single attention mechanism, the proposed model achieves the largest improvement in MAE and RMSE by 28.84% and 30.37%. Furthermore, the proposed model has better stability and better interpretability of the monitoring process.