高危粉末炸药包装流水线的封装质量实时质检

Problem

The client’s production line packages hazardous powder explosives; any sealing-quality defect must be caught within the line’s cycle time, before the next piece enters the dangerous zone. The line has four nasty conditions: (i) poor and changing illumination, (ii) the sealing seam is small in frame, (iii) the seam edge is frequently motion-blurred, and (iv) the conveyor has irregular vibration that breaks naive temporal assumptions. Standard detectors trained on clean public data simply do not survive here.

Approach

• Robust detector. Multi-scale feature extractor + boundary-focused loss, so the model locks onto the seam edge instead of bulk shape; multi-frame fusion to average out the vibration; and per-frame histogram normalisation to neutralise lighting drift. These four together push accuracy from “unusable” to ≥ 99% on a held-out on-line test set.
• Rule gate in hot path. A buffered Cython-based rule layer runs in front of the network to reject trivially-good frames, so the expensive model only looks at ambiguous cases. The network itself is a detection–segmentation dual-task model, and the backbone uses structural reparameterisation so the deployment graph is a simple stack of conv layers even though training uses multi-branch blocks.
• Deployment. Exported via TensorRT to the customer’s on-site GPU. Inference throughput went from 4 FPS → 15 FPS (≈ 3.75×) on an RTX 4060, comfortably inside cycle time.
• System. Designed and shipped a fixed-protocol online-monitoring API so the existing PLC / SCADA stack can ingest alarms without changes. The pipeline ran continuously for 30 days in production with sustained ≥ 99 % accuracy.

Why it matters

This is not a benchmark leaderboard number — it is a defect detector deployed in an actual explosives-handling environment, where a false negative has real-world consequences. The engineering bias of the project (robustness at the hardware edge, strict latency budget, mutual-trust interface with legacy automation) is closer to shipping production ML than to publishing it.