Building a multi-signal based defect prediction system for a friction stir welding process

Abstract

Abstract This paper presents a study undertaken to build a multi-signal based defect prediction system for a friction stir welding process, as a step toward developing an online monitoring system. Compared with previous studies, this study is new and more complete in the following senses: (i) the experiments cover wider range of conditions than before, yielding sufficient number of good and defective welds (including both “cold” and “hot”); (ii) multiple signals are used simultaneously; and (v) it is the first study on friction stir welding process monitoring to apply ant colony optimization (ACO) based feature selection to include only discriminatory features in the model prediction. In friction stir (FS) welding of Aluminum Alloy 2219-T87, signals were acquired. Next signals were segmented into discrete windows to imitate data availability in an on-line monitoring system, and finally features were extracted using discrete wavelet transform (DWT). Extracted features were then pooled together and selected using ACO in the process of building model for testing. Predicting the quality of the weld is considered to be a 3-class classification problem that is evaluated using the wrapper method with KNN as the evaluation model. The leave-one-out test results show that feature extraction and selection on signals obtained in friction-stir-welding (FSW) can be a powerful method of defect detection when used with a machine learning algorithm for classification. The predictive power of models trained with selected features was able to yield 98.8487% classification accuracy. The work completed in this study provides a solid foundation for the creation of an on-line sensing system for monitoring the quality of a friction stirred welded specimen.