Penetration feature extraction and modeling of arc sound signal in GTAW based on wavelet analysis and hidden Markov model

Abstract

In this paper, a prediction model between arc sound signal and penetration states in Gas Tungsten Arc Welding (GTAW) is proposed, based on multi-scale analysis of wavelet transform and applying the theory of hidden markov model (HMM) for its good dynamic time sequence modeling ability, using hidden markov model toolkit (HTK) software. The region of interest (ROI) of arc sound signal is firstly extracted by means of the wavelet transform, then Mel frequency cepstral coefficients (MFCC) of arc sound signal are extracted and analyzed. Furthermore, training and recognition are implemented on the prediction model to get the best setting model. The experimental results demonstrated that the recognition rate of ‘wavelet analysis+HMM’ prediction model of arc sound signal and penetration states could reach more than 90%, which has higher recognition rate and adaptive capacity during dynamic robot GTAW welding process. Consequently, this prediction model has its advantage in dynamic process modeling of arc sound signal.