The real time analysis of acoustic weld emissions using neural networks

Artificial Neural Networks (ANNs) are becoming an increasingly viable computing tool in control scenarios where human expertise is so often required. The development of software emulations and dedicated VLSI devices is proving successful in real world applications where complex signal analysis, pattern recognition and discrimination are important factors. An established observation is that a skilled welder is able to monitor a manual arc welding process by subconsciously changing the position of the electrode in response to an adverse change in audible process noise. Expert systems applied to the analysis of chaotic acoustic emissions have failed to establish any salient information due to the inabilities of conventional architectures in processing vast quantities of erratic data at real time speeds. This paper describes the application of a hybrid ANN system, utilising a combination of multiple ANN architectures and conventional techniques, to establish system parameter acoustic signatures for subsequent on line control.