DVAE + Particle Filter framework monitors arc welding without labeled data

A research team led by Yue Cao at the University of Kentucky has developed a novel monitoring framework for arc welding processes that addresses key limitations of existing deep learning approaches. The method, detailed in a paper published in the Journal of Manufacturing Processes, combines a Dynamic Variational Autoencoder (DVAE) with a Particle Filter (PF) to perform robust, real-time state estimation from weld pool images. The DVAE uses a CNN-based encoder to extract compact latent features and an LSTM-based transition model to capture the temporal evolution under control inputs. The Particle Filter then propagates both the latent states and LSTM hidden states, effectively suppressing sensor noise from arc radiation and specular reflections while preserving process history.

The framework’s key strength is its generalizability – it was validated on both Gas Tungsten Arc Welding (GTAW) and Gas Metal Arc Welding (GMAW) without any process-specific tuning or retraining. This is a significant improvement over traditional deep learning methods that require large labeled datasets and extensive fine-tuning for each new application. By relying on unsupervised learning and Bayesian filtering, the approach can scale across different welding processes and operating conditions. The architecture is specifically designed for welding’s slow, inertial dynamics, making it particularly suited for industrial continuous fabrication environments where real-time monitoring is critical for detecting disturbances that impair weld quality.