Anomaly detection in friction stir welding using variational autoencoder with probe temperature

In friction stir welding (FSW), real-time monitoring of process states is essential. Conventional studies on anomaly detection in FSW have mainly relied on supervised models that require both normal and anomalous data. However, inducing anomalous data in actual production is costly and often impractical, limiting the industrial applicability of these models. This study proposes an unsupervised anomaly detection model based on a Variational Autoencoder (VAE) trained solely on normal data, with a particular focus on internal probe temperature. To evaluate its effectiveness, butt-joint welding experiments of aluminum alloy plates were conducted using a wireless sensor holder capable of measuring probe temperature, shoulder temperature, and bending force. Systematic experiments were conducted to evaluate sensor selection, model performance, and the effects of input dimensionality relative to process thermal diffusion scales, and robustness across welding distances. The results showed that probe temperature was the most responsive signal for identifying localized transient thermal drops near the defect formation zone, unaffected by the shoulder’s inherent thermal smoothing. Furthermore, the proposed VAE model achieved superior and stable anomaly detection compared to deterministic models, capturing dynamic defect progression before macro-defect formation without inflating plunge-stage thresholds. These findings demonstrate that the proposed VAE framework enables robust inline anomaly detection in FSW without labor-intensive data labeling, providing practical process insights and design guidelines for manufacturing environments.