J. J. Wang, J. Huissoon, University of Waterloo, Waterloo, ON, Canada
Arc acoustic information in GMA welding provides essential feedback for skilled welders and enables them to control the welding processes and the quality of the weld. However, it is difficult to determine what it is in the arc sound that provides the cues that these skilled welders have learned to react to for adjusting controllable weld process parameters. This is partly due to the difficulty in extracting a ‘clean’ arc sound in a typical welding (or even laboratory) environment, and the inability of most welders to be able to succinctly describe what it is that they listen for. In this paper, we examine two classification algorithms for analyzing the arc sound to establish a correlation with the controllable welding parameters. The two classification techniques are an artificial neural network with backpropagation (ANN-BP) and a naïve Bayesian classifier (NBC). The ANN-BP is able to tackle the complex input-output mapping given adequate data samples, while the NBC is well suited for classifying patterns with uncertainties.
The arc acoustic data were first grouped into a 15-feature data set, which consisted of the mean values, frequencies, and scatter of acoustic signals; the output voltage (one of the dominant parameters) was considered as the main output, and was categorized into three classes. The performance of each algorithm was evaluated by the Macro-averaged F-measures. The analytical results are significant, and the performance of the NBC was found to be better than that of the ANN-BP. The NBC also exhibited promising abilities for dealing with pattern uncertainties.
