Evaluation of welding skill using probability density distributions and neural network analysis

Manual Metal Arc Welding (MMAW) is learned best by practice and the current procedure of assessing this learning is by inspection and/or testing of the weld. This is an indirect, expensive and time consuming method as the assessment can be made only after completion of weld and its subsequent inspection or testing. A possible alternative to this is the acquisition of electrical signals at a very high speed while welding is in progress and their subsequent analysis. Skill of the welder largely depends on ability of the welder in maintaining constant arc gap which, in turn results in steady state arc voltage. Hence, if voltage during welding can be acquired at a sufficiently high rate of acquisition, then this data can be analysed to assess welders' skill. Accordingly, data was acquired from trainee welders and from an experienced welder at a sampling rate of 100,000 samples/s and subsequently subjected to statistical and neural network analyses. Comparison of probability Density Distributions (PDDs) generated from these data and the neural network analysis revealed improvement in the learning of the welders with progress of training. These procedures were also employed independently to assess the skill of a large number of trainee welders at the end of their training. Ranking based on this procedure matched fairly well with that produced independently from visual examination of the weld.