A Novel Quality Monitoring Approach Based on Multigranularity Spatiotemporal Attentive Representation Learning During Climbing GTAW
Yuxiang Hong, Mingxuan Yang, Ruiling Yuan, Dong Du, Baohua Chang
Abstract
Reliable welding quality monitoring (WQM) is a long-standing challenge for climbing gas tungsten arc welding (GTAW) due to the inherent instability and complexity of the weld pool during upward welding, especially for the fabrication of large-scale structural components with medium-thick and thick aluminum plates. This article presents a novel WQM approach based on multigranularity spatiotemporal attentive representation learning, aiming to accurately characterize molten pool state and detect welding defects in real time. A passive vision sensing system is constructed to monitor the climbing GTAW process. A long-term dynamic information-enhanced multigranularity spatiotemporal attentive representation learning network is proposed. The network adopts a feature-level image fusion strategy and multigranularity attention mechanism to simultaneously aggregate discriminative information at different semantic levels on the temporal and spatial dimensions from a global view, while utilizing a bilateral branch structure to alleviate class imbalance in the data. Moreover, long-term dynamic information is mined from the molten pool time series images through motion edge history images. Experimental results show that the proposed approach has a remarkable classification performance and robustness compared with the typical comparison models even with class imbalance and noisy training data. This approach offers a promising new solution for WQM and is expected to be utilized to provide real-time feedback in a closed-loop quality control system.