Degradation-Aware Remaining Useful Life Prediction of Industrial Robot via Multiscale Temporal Memory Transformer Framework

Remaining useful life (RUL) prediction is of great importance to ensure stable operation of industrial robots (IRs). Deep learning-based methods have been proven effective in the RUL prediction tasks of IR. However, they are not effective in perceiving the state variation from a health state to a degradation state of IR and fail to reveal multi-term patterns of IR for RUL prediction. To address these challenges, a multiscale temporal memory Transformer framework is proposed to implement RUL prediction combined with state change identification. This proposed framework comprises a memory autoencoder Transformer network and a multiscale temporal Transformer network. The former Transformer network captures variation hidden in temporal information to detect the state change point, while the latter Transformer network is adopted to mine multi-term temporal dependencies for RUL prediction once state change point is identified. A self-built IR platform is constructed to validate our proposed method. Compared with the other advanced methods, the prediction results show that our method can locate the state change point in advance and achieve high-precision RUL prediction for IRs.