Implementation and evaluation of an echo state network for a quality inspection system for laser welding

Abstract

Laser welding is a widespread process in series production due to its high welding accuracy and speed. However, performing quality controls manually slows down the manufacturing process and causes high personnel costs combined with potential unreliability. In this paper, various Echo State Network (ESN) architectures for seam classification serving the facilitation and acceleration of the manual postproduction quality inspection are presented. In order to improve the performance of the ESN, a data standardization method is proposed, which applies accumulated information from all seams of the training set to a single seam in the test set. Considering that the automotive industry requires a high level of classification reliability, classification probability histograms are introduced. These enable comparing the reliability of models and clearly outperform typical confusion matrix metrics. Based on the evaluation of different tests performed with these histograms, a combined ESN architecture is suggested. This architecture uses two types of readouts, i.e., Multilayer Perceptron and Support Vector Machine, and utilizes the strong sides of both. In this work, a real laser welding dataset from a series production is used to evaluate all proposed methods and architectures. The data represent the records of light emission values during the welding process.