Answers are grounded in this paper and cite their sources.
In thermal power plants, accurate detection of torsion faults in tubular belt conveyors is crucial for operational safety. Traditional methods lack efficiency and accuracy, while existing deep learning models struggle with real-time processing demands. This paper introduces an enhanced Dynamic-UNet architecture, incorporating depthwise separable convolutions, a quality-aware gating module, and a simplified decoder structure. Evaluated on a dedicated industrial dataset, our model achieves a 97.89% mean Intersection over Union (mIoU) and a 32.28 Frames Per Second (FPS) inference speed on an NVIDIA RTX 3090, which is 2.72 percentage points higher in accuracy and 2.2× faster in inference than the original UNet, and outperforms mainstream semantic segmentation models. Real-world deployment on a Huawei Atlas DK A2 edge platform with a dynamic Region of Interest (ROI) mechanism further delivers an effective inference speed of 33.46 FPS, fulfilling industrial real-time requirements.
No takes yet. Share an insight, caveat, or question.
Gao et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: