Contact Feature Recognition Based on MFCC of Force Signals

Abstract

Contact-based tasks such as assembly and grinding often require the information on contact states. This letter therefore proposes a recognition method based on the Mel Frequency Cepstrum Coefficient (MFCC) of force signals. It demonstrates that the combination of MFCCs and time delayed neural networks is effective for learning features of contact events. This method is able to recognize instantaneous responses that do not generate repetitive waveforms. As a result, the recognition rate of the click response during a pen cap closing task increased from 75% to 96% following the proposed method. It is confirmed that this method is applicable not only to the data obtained by the robot s highly reproducible motions but also to the data whose parameters are scattered due to human interference.