KJCKorea
Journal Central

This study compares modality-specific specialized models distributed across a heterogeneous edge cluster with single-device omnimodal deployment under identical tasks. A cluster of Jetson AGX Orin, Raspberry Pi 5+Hailo-8, and Orange Pi 5 Plus runs MobileViT-S, Whisper-tiny, and DistilBERT via gRPC, benchmarked against LLaVA-1.5-7B FP16, its INT4 quantization, and Qwen2-VL-2B FP16. The distributed configuration is 80–288× faster on Vision, 3.7–25× faster on Text, and up to 806× more energy-efficient, with Text F1 within 2.8 percentage points across all four configurations. Effect decomposition attributes 3.93× to quantization and 1.75× to model size reduction, with an additional 63.5× from distribution itself, confirming independence from model size and quantization. The 17–195× cost-normalized advantage is retained across five wired-to-wireless network scenarios; distributed clusters thus suit concurrent multi-modal serving, while omnimodal deployment suits cross-modal reasoning.

This study developed a real-time deformation measurement system using a triangulation-based laser displacement sensor to evaluate the structural safety of Falling Object Protective Structures (FOPS) on agricultural tractors. The system was applied during drop impact tests compliant with OECD Code 10, where a 45 kg steel sphere was dropped from 3.06 m, generating 1365.8 J of impact energy. Two FOPS configurations were tested: Case 1 with a steel roof and Case 2 with a frame-only structure. The sensor system successfully captured maximum elastic deflections of 45.4 mm at 0.06 s (Case 1) and 132.3 mm at 0.14 s (Case 2), both within the safety clearance zone. Measurement reliability was validated using gage blocks, yielding RMSE < 0.53 mm and R² > 0.99. These results confirm the system’s capability for accurate, high-speed deformation monitoring under dynamic loading, offering a robust tool for future FOPS certification and analysis.

This paper proposes a vision-based framework for worker safety-state recognition and risk-aware decision support using helmet detection. The proposed method classifies worker head regions into helmet and no-helmet states using YOLOv8-based object detection models and uses the detection results as input to a downstream decision module. Experiments on a public safety helmet dataset showed that YOLOv8s achieved the most balanced performance, with an mAP@50 of 0.933 and a no-helmet recall of 0.867, while robustness analysis showed that noise and occlusion caused the largest performance degradation. In the simulation-based downstream decision experiment, the proposed method reduced both safety violations and unsafe exposure time to zero under the tested simulation settings, while achieving a shorter completion time than the distance-only strategy. These results indicate that helmet-state recognition can provide useful contextual information for simulation-based industrial safety monitoring and risk-aware decision support.