Abstract
This paper presents EdgeBakeTwin, a prototype multimodal edge perception architecture for industrial imaging and sensing workloads in continuous bread manufacturing, evaluated in a hardware-in-the-loop simulation environment as a pre-deployment validation. Four complementary modalities—RGB visual stream, LWIR thermal array imaging, ambient temperature/humidity, and three-phase power telemetry—are fused on a Raspberry Pi 5 gateway coordinating 17 ESP32 IIoT nodes, with all latency-critical inference executing at the edge without cloud dependency. A quantized MobileNetV2 (INT8, 3.4 MB backbone footprint; 4.8 MB total including MLP fusion head) performs on-device defect classification across four quality categories at 11.2 ms mean latency (end-to-end, backbone + fusion head). The model was trained on 14,200 labeled images (2,800 real + 11,400 synthetic); on the 1,820-image evaluation set the model achieves weighted-average precision 0.975, recall 0.975, and F1-score 0.975. On the 420-image real-only subset the weighted F1 is 0.947, providing preliminary generalization evidence. Simulation-estimated system-level evaluation over 15 independent runs suggests potential reductions of 24.4% in production cycle time and 23% in energy consumption relative to a rule-based baseline; these figures are simulation-based estimates pending live-line deployment validation.