Problem
Bar & Rod Mill (BRM) plant floors run continuous video feeds across loopers, rolling stands and material-handling zones. Equipment anomalies, near-miss events and safety violations need to be caught early — but a human watching every camera, every shift, does not scale.
Why it matters
Sending every frame to the cloud for analysis adds latency and bandwidth cost that industrial networks often can't absorb. Safety alerts also lose their value if they arrive late. The goal of this project is to push detection to the edge — close to the camera and the machine — so that anomalies and unsafe conditions are flagged in near real time.
Dataset / input source
The system is designed around industrial video feeds from the Bar & Rod Mill environment, with emphasis on Stand-16 looper monitoring. The data includes plant-floor video frames affected by dust, illumination variation, motion blur, vibration, occlusion, and camera-angle limitations.
Architecture
The pipeline starts with a production-shaped YOLO detector trained on plant-floor footage, optimised for the target edge device using the OpenVINO inference runtime. A lightweight event layer sits after detection — it smooths short-lived false positives, applies zone rules, maps detections to equipment regions, and emits timestamped events with confidence scores and frame evidence for auditability.
The deployment direction is edge inference using OpenVINO / Jetson-class hardware for low-latency plant-floor monitoring.
Tools used
Target deployment hardware is Jetson / industrial edge-server class compute, with OpenVINO as the inference runtime for low-latency, on-site processing.
Result
Current status: M.Tech research prototype and deployment-focused industrial case study. The system is being developed for real-time anomaly and safety-event detection under industrial visual degradation.
Limitations
Performance can be affected by dust, glare, occlusion, lighting change, camera vibration, and limited annotated plant-floor data.
Future work
A natural extension is feeding detected events into a multi-modal LLM/RAG layer that fuses vision, sensor telemetry and logs into plain-language plant-intelligence alerts for operators — the focus of a related project on this site.
Additional planned work includes expanding coverage to more cameras and zones, on-device model retraining as new data is collected, and tighter integration with plant maintenance workflows.
Links
Code, thesis writeup and demo materials will be linked here as they are published. In the meantime, see github.com/Nitin-Mane for related repositories.