In-Process Laser Welding Monitoring by Fusing the Uncertain Signal Information of Multi-Photodiode Sensors

Abstract

Remote laser welding is an emerging joining technology to meet the increasing demand of corrosion resistance, fast, non-contacted and single sided joining for automotive body-in-white assemblies. However, the quality of laser welding has been a critical issue in the popularization of this technology. Traditionally, various stochastic detection methods have been developed for in-process weld defect detection by monitoring and classifying various weld signals. The main objective of this thesis is to develop an in-process welding monitoring system including; (i) a novel defect detection algorithm based on a multi-sensor fusion technique, (ii) a new optical sensor configuration to capture in-process weld signal, and (iii) an offline weld signal analysis/training module and an user interactive online detection module.

The three weld signals are monitored: weld pool temperature, plasma intensity, and back reflected laser intensity. Their nominal trends are identified by estimating a probability distribution function for the signals and appropriate thresholds are specified by the standard statistical analysis of the residuals at the confidence interval of 95%. We propose a probability assignment function, characterized by shape controllability with respect to the extracted thresholds. We can analyze the in-tolerance defect problems by the proposed probability assignment function that can deal with the decision uncertainty near the thresholds. The individual sensor information is utilized to identify the probability of the normal state. The probabilities are aggregated by using the combination rule of the Dempster-Shafer theory. The performance of the developed detection method is evaluated by the statistical comparison with conventional visual inspection results.