| Multiple sensor monitoring utilizes the inherent advantages of individual sensors and combines them into an all-encompassing system. In this research, infrared (IR), ultraviolet (UV), audible sound (AS), and acoustic emission (AE) sensors were used for real-time monitoring of laser welding. Current process monitoring technology is limited by inadequate signal processing algorithms. Pattern classification using linear and quadratic discriminant algorithms was developed in combination with various methods of feature extraction including: class mean scatter, singular value decomposition, and Renyi information distribution. Optimal window sizes and dimensionality of features were selected for identification of porous and inadequately penetrated welds. Results showed that very good classification of 100% was possible for detecting porosity when using seven or more fused features from all sensors. Inadequate penetration was investigated in laboratory and industrial settings with good results of 100% classification.; Time frequency analysis was developed using reduced interference distributions (RID) and the Renyi information distribution. Promising results were shown with 100% classification for discriminant analysis using Renyi numbers and energy.; Adaptive process monitoring was also developed using an adaptive quadratic classifier. The adaptive system was shown to increase classification percentage by continuously training the system. Classification was increased from 90% to 93% from the rejection of outliers that normally corrupt the training set. The adaptive classifier's ability to change the decision boundary reflected in the improved classification from 50% to 100% during process drift. |
