Real-time defect detection is critical for guaranteeing the integrity of Gas Metal Arc Welding (GMAW) joints. Conventional non-destructive testing (NDT) delivers high-quality diagnostics only after the weld has finished, precluding immediate corrective action. We present the Multi-modal BiFPN Gate Network (MBGN), a deep-learning framework that fuses two complementary data streams: (i) time-synchronized welding current and voltage waveforms processed by a one-dimensional CNN, and (ii) high-speed molten-pool imagery processed by a ResNet backbone. The two modalities are merged through a Bidirectional Feature Pyramid Network (BiFPN), and a lightweight Gate module adaptively recalibrates cross-modal interactions before a final classifier. Experiments on an industrially collected multimodal dataset demonstrate that MBGN attains 0.748 accuracy and 0.719 F1-score, outperforming state-of-the-art baselines by 8-12 % in F1. These results validate the efficacy of multimodal fusion for in-process defect detection and pave the way for autonomous real-time quality control in industrial welding.