The analysis of airborne acoustics of S.A.W. using neural networks

The analysis of acoustic emissions for machine health monitoring has made rapid advances in the last five years due to a revival of interest in the application of Artificial Neural Networks (ANNs). Complex signal analysis, which has often thwarted conventional statistical methods and expert systems, is now more possible with the introduction of 'neural' based computing methods. Acoustic emissions from welding processes are well documented. In particular, it has been established that a manual welder is capable of making intrinsic decisions concerning electrode position based on process noise. The analysis of time / amplitude signals and Fast Fourier Transforms (I-I-1s), within salient frequency bandwidths of the weld acoustic, has yielded erratic, unpredictable and noise polluted data. Extracting a meaningful interpretation from this data is computationally intensive when utilising standard statistical methods and leads to data explosions, especially when an 'on-line' corrective control signal is required. An Artificial Neural Network is 'trained' on examples from acquired data and performs a robust signal recognition task rather than relying on a programmed set of data samples as in the case of an expert system. This technique enables the network to generalise and, as a consequence, allows the input data to be erratic, erroneous and even incomplete. This research defines the development of a hybrid system, utilising high speed date capture and 141-1' computation for the signal pre-processing and a 'self organising' network paradigm to establish weld stability and real time corrective control of the process parameters. The paper describes a successful application of a Neural Network hybrid system to determine weld stability in submerged arc welding (S.A.W) through the interpretation of airborne acoustics.