🎓 Lesson 21
D5
Edge-AI Deployment for Real-Time Anomaly Detection in ROC Feeds
Edge-AI deployment means putting smart, lightweight AI models directly on cameras or sensors at the mine site—so they can instantly spot unusual events (like equipment failure or safety violations) in ROC video feeds without waiting for cloud processing.
🎯 Learning Objectives
- ✓ Analyze latency–accuracy trade-offs when quantizing a YOLOv8n model for edge inference on ROC-grade IP cameras
- ✓ Design an edge-AI deployment pipeline—including model pruning, INT8 quantization, and TensorRT optimization—for a given ROC camera specification (e.g., 4K@30fps, 2GB RAM)
- ✓ Explain how anomaly detection thresholds impact false alarm rates in ROC safety-critical contexts using precision-recall curves
- ✓ Apply ISO/IEC 23053:2022 guidelines to validate edge-AI model robustness against lighting, dust, and motion blur degradation in underground ROC feeds
📖 Why This Matters
In modern mines, Remote Operations Centers rely on dozens of real-time video feeds—from haul trucks to conveyor belts to personnel zones. Waiting seconds for cloud-based AI to flag a dropped roof bolt or unauthorized entry creates unacceptable safety and productivity risk. Edge-AI cuts detection latency from >2s to <150ms—enabling immediate operator intervention, automated shutdown triggers, and regulatory compliance with real-time situational awareness mandates like ICMM’s Digital Safety Framework.
📘 Core Principles
Edge-AI for ROC feeds rests on three interdependent layers: (1) Anomaly modeling—where unsupervised (e.g., autoencoder reconstruction error) or semi-supervised (e.g., few-shot adapted YOLO) approaches learn 'normal' operational signatures from historical ROC footage; (2) Edge compute constraints—requiring model compression (pruning, quantization), hardware-aware compilation (TensorRT, ONNX Runtime), and thermal/power budgeting for industrial enclosures; and (3) Operational integration—ensuring detections feed into ROC SCADA alarms, log audit trails per MSHA Part 46, and support explainability via saliency maps for incident review. Crucially, edge deployment shifts the failure mode from 'model inaccuracy' to 'system resilience'—demanding fail-safe fallbacks (e.g., rule-based heuristics) when AI confidence drops below threshold.
📐 End-to-End Inference Latency Budget
Total latency must stay ≤200 ms for actionable ROC alerts. This budget is partitioned across capture, preprocessing, inference, and postprocessing. The inference component dominates—and is governed by model FLOPs and hardware throughput. Optimized latency prediction enables hardware selection and model sizing before field deployment.
Edge Inference Latency Estimate
L_infer = (FLOPs_per_frame / Throughput_TOPS) × k_overheadEstimates worst-case inference time (ms) for a quantized model on target hardware, where k_overhead accounts for memory bandwidth, kernel launch, and driver latency.
Variables:
| Symbol | Name | Unit | Description |
|---|---|---|---|
| L_infer | Inference latency | ms | Time taken solely for model execution on edge device |
| FLOPs_per_frame | Floating-point operations per frame | FLOPs | Computational load of one inference pass |
| Throughput_TOPS | Hardware inference throughput | TOPS | Trillion operations per second (FP16 or INT8) measured on target platform |
| k_overhead | Empirical overhead factor | dimensionless | Multiplier (typically 6–12) capturing non-ideal execution conditions |
Typical Ranges:
Jetson AGX Orin (INT8): 6–9
Intel Core i5 + OpenVINO: 8–12
💡 Worked Example
Problem: A ROC site deploys an anomaly detector on an NVIDIA Jetson AGX Orin (32GB) running at 15W TDP. The model requires 1.2 GFLOPs per frame. Measured FP16 throughput is 105 TOPS. Camera resolution is 1920×1080@25fps. Calculate worst-case inference latency and verify against ROC safety threshold.
1.
Step 1: Convert model FLOPs to same unit: 1.2 GFLOPs = 1.2 × 10⁹ FLOPs
2.
Step 2: Compute theoretical latency = FLOPs / (Throughput × 10¹²) = (1.2 × 10⁹) / (105 × 10¹²) = 0.0114 ms — but this ignores memory bandwidth and kernel launch overhead
3.
Step 3: Apply empirical overhead factor (industry-observed): multiply by 8.5 → 0.0114 × 8.5 ≈ 0.097 ms. Add preprocessing (2.1 ms) and postprocessing (1.3 ms) → total inference stage = 3.5 ms
4.
Step 4: Add camera capture (40 ms @25fps), network jitter (≤5 ms), and SCADA dispatch (≤10 ms) → total system latency = 58.5 ms
Answer:
The result is 58.5 ms, which falls well within the safe ROC threshold of ≤200 ms and leaves margin for future model upgrades.
🏗️ Real-World Application
At Rio Tinto’s Gudai-Darri ROC (Pilbara, WA), edge-AI anomaly detectors were deployed on 42 Hikvision DS-2CD7XXG-ZY cameras monitoring autonomous haul truck loading zones. Models were trained on 18 months of annotated feed (including 3,200+ instances of bucket misalignment, tire damage, and proximity breaches). Using TensorRT-optimized YOLOv8s with INT8 quantization, inference ran at 87 FPS on Jetson Orin modules mounted inside NEMA-4X enclosures. System reduced false positives by 63% vs. cloud-only baseline and achieved 99.2% recall on near-miss proximity events—triggering automatic speed reduction in adjacent trucks within 112 ms (validated per ISO 13849-1 PLd). Integration included MSHA-compliant audit logging and weekly model drift monitoring via embedded SHAP analysis.