Explainable Models for Multivariate Time-series Defect Classification of Arc Stud Welding

Abstract

Arc Stud Welding (ASW) is widely used in many indus-tries such as automotive and shipbuilding and is employed in building and jointing large-scale structures. While defec-tive or imperfect welds rarely occur in production, even a single low-quality stud weld is the reason for scrapping the entire structure, financial loss, and wasting time. Preventive machine learning-based solutions can be leveraged to mini-mize the loss. However, these approaches only provide predictions rather than demonstrating insights for characterizing defects and root cause analysis. In this work, an investigation of defect detection and classification to diagnose the possi-ble leading causes of low-quality defects is proposed. More-over, an explainable model to describe network predictions is explored. Initially, a dataset of multi-variate time series of ASW utilizing measurement sensors in an experimental environment is generated. Next, a set of techniques to leverage synthetic measurements, reference, and residual signals, and generate a residue dataset, are proposed. Finally, the archi-tecture of classification models is optimized and by Bayesian black-box optimization methods to maximize their perfor-mance. Our best approach reaches an F1 score of 0.84 on the test set. Furthermore, an explainable model is employed to provide interpretations on per class feature attention of the model to extract sensor measurement contribution in detecting defects as well as its time attention.