Benchmarking one-class and binary classifiers for weld defect detection using acoustic emission signals

Abstract

The aim of this study is to apply diagnostic techniques to welds based on machine learning methods. This approach works exclusively with signal data, making it nondestructive and enabling easier and faster implementation. It detects anomalies immediately after the welding procedure by analyzing acoustic emissions, and it can also be extended to other inputs. The simplicity, completeness, and clarity of our method make it uniquely effective in detecting anomalies during the manufacturing process. Although previous research has explored the integration of acoustic emission testing with deep autoencoders, the full potential of real-time diagnostics using machine-learning-based binary and one-class classifiers, especially under dynamic welding conditions, remains investigated insufficiently. This study provides a comparative evaluation of supervised, semi-supervised, and unsupervised machine learning approaches. Supervised machine learning methods use binary classification, while semi-supervised and unsupervised methods use one-class classifiers. Binary classifiers such as Ridge Classifier and Extra Trees achieved over 95% accuracy and outperformed one-class classifiers with respect to Cohen’s Kappa and MCC. These classifiers and results are compared and discussed in detail in the discussion section.