State-Wise LSTM-GRU Method for Ball Screw Prediction

Ball screws play important roles in wide applications related to aircraft equipment and its manufacturing process. For these highly complex systems, their demands for reliability is high and strict. Thus, the precise evaluation and prediction of the health status of the ball screws is necessary. As the degradation trend often varies due to multiple independent or coupled factors, the degradation of the ball screw performance may go through a “decrease-increase-decrease” trend. In such cases, a single degradation model cannot accurately describe how the degradation develops. In this paper, we predict the health of the ball screw by dividing the degradation process into several different trends with the assumption that the trend changes with the status of the ball screw. To achieve this, a new prediction model is proposed based on a state-wise deep learning method of the long short-term memory network and gated recurrent unit (LSTM-GRU). First, the degradation trend of the ball screw is obtained by the combination of Kernel Principal Component Analysis (KPCA) and Mahalanobis distance (MD). Second, the method of the density-based spatial clustering of applications with noise (DBSCAN) is used to evaluate the health status of the ball screw. Then, the proposed method of the state-wise LSTM-GRU is validated on a life-cycle experiment of the ball screw. Compared with the standard long short-term memory network (LSTM) and the method of gated recurrent unit GRU), the results show that the proposed model of state-wise LSTM-GRU can predict the performance state of the ball screw more accurately and more robust, which indicates it is a promising method for the performance prediction of the ball screw.