This thesis investigates real-time defect identification in welding processes using thermal frames and audio signals. The objective is to distinguish normal from defective welds and to support diagnosis across twelve weld-quality classes. The study uses the Intel Robotic Welding Multimodal Dataset and compares three fusion approaches under a grouped, session-disjoint evaluation protocol: a handcrafted-feature Hierarchical Ensemble, a Backbone Fusion model using pretrained visual embeddings and audio features, and WeldFusionNet, an end-to-end neural model combining audio and thermal-video encoders. The experiments include synchronized preprocessing, offline classification, recording-level replay, and latency profiling. WeldFusionNet achieves the best binary defect-screening performance on the cross- session test partition, with binary F1 of 0.9895, ROC-AUC of 0.9986, and accuracy of 0.9833.