Proof-of-concept study demonstrates real-time defect segmentation and optimized edge messaging for steel manufacturing, highlighting trade-offs between model accuracy and latency.
Key Points
To evaluate trade-offs between accuracy and computational footprint for steel defect segmentation models and optimize an edge-oriented IIoT messaging architecture.
Benchmarked a compact U-Net (0.49 M parameters), an ImageNet-pretrained DeepLabV3+, and an Otsu baseline on a defect-stratified test split of 1,886 Severstal images using Dice, IoU, precision, recall, and F1 metrics.
Evaluated a three-layer IIoT communication pipeline on a Raspberry Pi broker across 158,500 MQTT messages using a factorial experimental design.
DeepLabV3+ achieved a Dice score of 0.677 at 46.3 ms per image with 37 times more parameters than the compact U-Net, which achieved a Dice score of 0.416 at 38.6 ms per image, while Otsu reached 0.060.
Configuring the broker transport layer rather than the publisher achieved a 17-fold reduction in end-to-end latency and sustained 1,920 messages per second without message loss.