A Novel Multimodal Approach for Gas Metal Arc Welding Quality Control

Quality control in Gas Metal Arc Welding (GMAW) poses challenges due to the complex nature of its underlying processes, making it difficult to identify defects during the welding operations. Conventional post-weld inspections, including non-destructive testing (NDT) methods are inefficient and restrict real-time process optimization. This study introduces a novel deep learning-based multimodal approach for real-time GMAW quality monitoring. The proposed solution integrates electrical data, including sensor readings such as voltage (V), current (I) and gas flow rate (GFR) along with high-resolution welding images captured by an advanced welding camera. To enhance the representation for the vision encoder model, timeseries sensor data is converted into spectrograms. The EfficientNetB2 pretrained model stands out as a balanced choice as a vision encoder, achieving high accuracy of 97.33% with a relatively efficient design in terms of both parameters and inference speed. In contrast, ResNet50 achieves the highest accuracy at 98.67%, but it comes with increased complexity and longer inference times. Moreover, comparative analysis underscores the effectiveness of our multimodal approach in achieving superior results over unimodal approaches. Furthermore, the proposed approach extends its applicability to Gas Tungsten Arc Welding (GTAW), providing a comprehensive solution for monitoring and enhancing quality control across various welding methods.