A Physics-Informed Neural Network Combined with Domain Generalization Method for Harmonic Reducers Rul Prediction

Harmonic reducers are an important component in industrial robots. Remaining useful life (RUL) prediction of harmonic reducers plays a significant role for the safety and reliability of industrial robots. However, due to the manufacturing deviations, various operational conditions and random degradation processes, the degradation behaviors of different harmonic reducers present significant unit-to-unit variability. While current data-driven methods show low capability in RUL prediction under serious unit-to-unit variability. To deal with this issue, this paper proposes a physicsinformed neural network (PINN) combined with domain generalization method for harmonic reducers RUL prediction. Specifically, the self-attention mechanism is introduced for feature extraction and the maximum mean discrepancy is computed based on the extracted features to quantify the deviation of data distributions among different units. Subsequently, a PINN serves to regularize the mapping from features to RUL. The proposed method was validated based on an accelerated degradation test of industrial robot harmonic reducers. The results demonstrate that the proposed method can provide more accurate RUL prediction result than other competitors.