Remaining Useful Life Prediction of Ultrasonic Motors Based on Multiscale Feature Fusion and LSTM

To address lifetime assessment of ultrasonic motors in space applications such as low-Earth-orbit laser communication mechanisms, the paper proposed a remaining useful life (RUL) prediction method based on multiscale feature fusion and long short-term memory (LSTM). Lifetime monitoring data were preprocessed to construct multiscale degradation features. A state transition score (STS) was used to identify the first prediction time (FPT), which was combined with the end of life (EOL) to establish piecewise RUL labels. Spearman rank correlation analysis was then used for feature selection, followed by LSTM-based RUL prediction and nested leave-one-motor-out (Nested LOMO) validation. The proposed method achieved mean root mean square error (RMSE), mean absolute error (MAE), and Pearson correlation coefficient (r) values of 0.11, 0.071, and 0.96, respectively. Compared with the model using only time-domain statistical features under the same RUL labeling scheme, the RMSE decreased from 0.25 to 0.11, the MAE decreased from 0.15 to 0.071, and r increased from 0.77 to 0.96. These results demonstrate the effectiveness of multisource and multiscale degradation information for cross-motor RUL prediction under small-sample conditions.