Deep Learning Based Kalman Filter for Variable-Frequency Disturbance Elimination in Force Sensing

This paper proposes a new approach for force sensation with the elimination of variable-frequency disturbances using deep learning based Kalman filter. The disturbance observer is employed for force estimation. To cancel the effect of undesired variable-frequency components superimposed in force information, the deep learning based Kalman filter is designed as an integration of the deep learning based frequency estimation and the periodic component elimination Kalman filter. The Kalman filter is designed to eliminate the periodic component by estimating the signal with periodic component, the first derivative and the second derivative of that signal. The deep learning based frequency estimation is constructed by the long short-term memory deep neural network to estimate the frequency of the periodic disturbance during force sensing operation. Hence, the frequency variation of the periodic component is detectable. This estimated frequency is a parameter which determines the performance of the Kalman filter in periodic component elimination. Therefore, the deep learning based Kalman filter is capable of excluding the undesired components with variable frequencies. The effectiveness of the proposed method is verified by numerical simulation results.