Anomaly Detection in Square Butt Joint by Friction Stir Welding Using Variational Autoencoder

This study proposes an anomaly detection method for the friction stir welding process using a variational autoencoder, which is a representative deep generative model in machine learning, with two time-series data: temperature near a tool probe tip and a tool shoulder tip and bending force on a tool. We mention a square butt welding process of a pair of aluminum alloy plates. Through preliminary welding experiments, normal and anomalous data are collected to construct a VAE network used for anomaly detection in square butt joining. The effectiveness of the proposed method is demonstrated through validation experiments by comparing the proposed VAE method with an autoencoder.