An essential tremor (ET) is the oscillations of a part of body. The ET patients have difficulty performing daily lives. We have been developing exoskeltal robot for tremor patients to suppress tremor. In this paper, we focus on signal processing method to extract voluntary movement signal from sEMG of ET patients in which voluntary movement signal and tremor signal are mixed. We proposed a method to extract Average Rectified Value (ARV) of sEMG among a periodic time of tremor signal. As evaluation, we compared ARV, Short-Time Fourier Transform (STFT), and Low-Pass Filter (LPF) about 2 types of misrecognition of Time Delay Neural Network (TDNN). Those are misrecognition occurred during action and non-action. Through this evaluation, it was confirmed that proposed method can decline both 2 types of misrecognition rate compared to STFT and LPF.