[2607.16013]
Yue Cao, Hai Lin, YuMing Zhang
Arc welding processes are essential for continuous fabrication but prone to disturbances that impair weld quality, making real-time monitoring critical yet difficult due to complex visual patterns and nonlinear, time-varying dynamics. Deep learning shows promise but faces scalability limits because of its dependence on large labeled datasets and application-specific tuning. We explore whether a unified approach can characterize major arc welding processes across applications and improve scalability through consistent state monitoring. This paper introduces a robust and generalizable monitoring framework for arc welding. It combines unsupervised deep latent representation learning, which extracts compact features from weld pool images, with Bayesian filtering to handle persistent and fluctuating disturbances such as arc radiation and specular reflections. Specifically, a Dynamic Variational Autoencoder (DVAE), consisting of a CNN-based encoder-decoder and an LSTM-based transition model, jointly learns latent representations and their evolution under control inputs. For robust real-time inference, a specialized Particle Filter (PF) propagates the latent and LSTM hidden states, preserving process history while suppressing sensor noise. This design is well suited to welding's slow and inertial dynamics. Validation on GTAW and GMAW without process-specific tuning demonstrates the framework's generalizability and robustness.