Frequency informed convolutional autoencoder for in situ anomaly detection in wire arc additive manufacturing

In the context of Industry 4.0, the importance of anomaly detection is growing, particularly in Additive Manufacturing, as itallows for the detection and localization of defects, thereby reducing waste and costs. However when normal and anomalysignals have similar shapes in time this task is particularly challenging. Despite that, the frequency content of time seriessignals often holds valuable information that, when integrated into the learning process, can greatly improve the recognitionof hidden patterns in the data and enhance feature separability. In this study, we propose an unsupervised anomaly detectiontechnique for Wire Arc Additive Manufacturing (WAAM) based on deep learning, namely 1D-Convolutional AutoEncoder.By integrating frequency-regularization terms based on wavelet analysis of defect-free welding signals during the trainingphase, the results demonstrated a significant 54.8% improvement in anomaly detection performance compared to similarmethods. This improvement enables the effective use of unsupervised learning for anomaly detection in WAAM, minimizingthe need for labeled data and making it suitable for industrial applications, even when dealing with unbalanced datasets.