Abstract: Aiming at on-line monitoring of welding quality, the time and frequency characteristics are analyzed of the arc sound signals in short circuit gas metal arc welding (GMAW). The concept of tone channel and its equivalent electrical model are suggested. The radical basis function (RBF) neural networks are applied for on-line pattern recognition of the gas-lack in welding, in which the input vectors of samples are constructed by the linear prediction coding (LPC) coefficients of sound signals. The research indicates that, the arc sound presents a ringing series of that occurs at the end of short circuit transfer or moment of arc re-ignition, and its energy mostly distributes in frequency band below 10 kHz. The arc tone channel is a time dependent distributed parameters system, of which the transmission properties depend upon many physical and geometrical factors of the arc and surroundings, and is excited by the sound source that generates from the change of arc energy, so that results in the sound. The tone channel can be equated with a time dependent coefficients digital filter, and can be represented parametrically with the LPC model of arc sound. The arc sound and its parametric model are valuable in on-line monitoring and controlling of welding quality.

Key words: Arc sound, Gas metal arc welding (GMAW), Linear prediction coding (LPC) model, Quality monitoring, Radical basis function (RBF) neural network, Time and frequency analysis