Flexspline wear prediction and remaining life assessment for harmonic drive based on machine learning

This paper proposes a RUL prediction method based on long short-term memory network and geometric mean optimization particle filterFirstly, the vibration signal is decomposed and reconstructed based on local mean decomposition and Pearson correlation coefficient, and the multi-dimensional time-frequency domain features are extractedThe long short-term memory network is used to perform deep learning on the temporal relationship between features, and a comprehensive degradation index with high monotonicity and correlation is constructed. Secondly, the physical mechanism of the flexwheel crack growth is fully considered, and according to its growth characteristics in the first, middle and last three stages, the multi-stage degradation state equations based on Paris model and Foreman model are established respectivelyFurthermore, in order to overcome the weight degradation problem of particle filter, a particle resampling strategy based on geometric mean optimization is proposed by introducing double fitness index to improve the diversity of particles and the continuity of state estimationFinally, the harmonic reducer accelerated life test was used to verify the proposed method, and the proposed method was compared with a variety of mainstream prediction algorithmsThe results show that the proposed LSTM-GMOPF method has the root mean square error decreased by 7.3% compared with the single root mean square indexCompared with the single Paris degradation model, the root mean square error of the proposed LSTM-GMOPF method is reduced by 4.6%Compared with deep neural network, bidirectional long short-term memory network and other methods, the determination coefficient of the proposed method is as high as 0.994, and the root mean square error and mean absolute error are as low as 0.018 and 0.010, respectivelyThe prediction accuracy is significantly improved, which proves the effectiveness and superiority of the proposed method in the wear prediction and remaining life assessment of harmonic reducer flexible wheelIt provides a reliable theoretical basis and technical approach for the predictive maintenance of industrial robots.