Penetration state recognition based on the double-sound-sources characteristic of VPPAW and hidden Markov Model

Abstract

A series of work is conducted to study the relationship between the welding sound and penetration states in VPPAW (variable polarity plasma arc welding), which are No-keyhole, Keyhole and Cutting Modes. The spatial distribution of sound pressure is measured and the optimal acquisition angle is determined as 150°. The mechanism, based on the special double-sound-sources characteristic of VPPAW, is proposed and demonstrated by a series of phenomena about sound of VPPAW. The sound source structure and sound generating mechanism are quite similar to those of speech. Five features, which are mean amplitude, mean energy, logarithm of energy, standard deviation and kurtosis, are extracted to depict the characteristics of each frame. Half of these feature vectors is employed to train a HMM (Hidden Markov Model) of 5 states, and the rest acts as the test samples. HMM is applied in VPPAW for the first time, and the recognition rate of prediction model, based on HMM, about acoustic signal and penetration states could reach up to 93.23%. The work shows that the acoustic signal of VPPAW is similar to speech to some extent, which establishes the theoretical basis for the application of speech process technology in VPPAW, and it is also practical, especially for HMM.