Feasibility study of Hierarchical Temporal Memories applied to welding diagnostics

• To employ a new algorithm to classify welding defects.
• Analyze its performance against a well known processing scheme (ANN).
• Give some steps for tuning HTMs to particular problems.

Defect classification in on-line welding quality monitoring systems is an active area of research with a significant relevance to several industrial sectors where welding processes are extensively employed. Approaches based on some artificial intelligence implementations, like Artificial Neural Networks or Fuzzy Logic have been attempted, but their impact in real industrial scenarios is nowadays rather modest. In this paper a new approach based on Hierarchical Temporal Memories and the acquired plasma spectra is explored and analyzed by means of several arc-welding experimental tests. Results show the ability of the proposed solution to perform a suitable classification among several weld perturbations. The search for an optimal configuration of the algorithm and the usefulness of both spatial (spectral) and temporal identification of patterns will be also discussed, and the results will be compared with those provided by a solution based on feature selection and neural networks, exhibiting the better performance of the HTM model in terms of performance and handling of the input data.