Remaining useful life prediction for the harmonic reducer of industrial robots via in-situ current signal and lightweight multiscale attention deep networks

Reducer degradation in robot joints causes excessive vibrations, affecting product quality. Remaining useful life (RUL) prediction of reducers using in-situ signals can avoid robot disassembly and reduces production downtime. However, in-situ signals are more complex than experimental data due to transient robot operations and industrial noise. To address this challenge, an in-situ RUL prediction method via lightweight Multiscale Attention Deep Network (MSADN) and current signal is proposed. First, the full life cycle of harmonic reducer in-situ signals is collected to build a dataset. Subsequently, the MSADN model is employed for RUL prediction. Within MSADN, a multiscale feature extraction (MSFE) module is designed to capture multiscale information from in-situ signals, while a downsampling filter layer (DFL) is incorporated to expand the receptive field. Finally, a novel evaluation metric, Epoch Toleration Accuracy (ETA), alongside other standard evaluation indicators, is introduced to assess RUL prediction performance. Experimental studies on industrial robot datasets and rolling bearing datasets demonstrate the effectiveness and superiority of the proposed MSADN, and two ablation studies validate the necessity of each MSADN component.