Abstract

Objective Dust is a typical hazardous byproduct in industrial production and fire scenarios， easily causing occupational health hazards， equipment wear， and even explosion risks. Accurate and efficient concentration detection is the core prerequisite for dust safety prevention and control. Traditional detection methods suffer from significant limitations in single-point measurement， high computational complexity of pixel-by-pixel concentration inversion， and poor real-time performance. Existing deep learning-based research mostly focuses on dust existence recognition and lacks systematic quantitative analysis of concentration. Meanwhile， restricted by the characteristics of small dust targets， low contrast， and blurred boundaries， they fail to meet the core requirements of on-site concentration safety early warning and risk quantitative assessment. Therefore， this paper carries out research on the optimization of dust concentration level detection algorithms.

Methods The YOLOv8.3.138 was adopted as the base detection model. First， a dust scattering image dataset was constructed using images captured by a charge-coupled device （CCD） industrial camera， and the images were divided into four risk levels according to relative concentration. After removing abnormal samples， the dataset was split into training， validation， and test sets at a ratio of 8：1：1. Meanwhile， a three-stage optimization strategy was proposed： a density-aware and edge-guided dynamic upsampling module was used to enhance fine-grained dust feature reconstruction； a concentration sensing feature enhancement （CAFE） module was employed to strengthen concentration-related feature responses； and a dynamic adaptive label assignment strategy （Dynamic ATSS） was adopted to optimize multi-scale target matching， thereby comprehensively improving the model detection performance.

Results and Discussion Validated by multiple sets of comparative experiments， the improved algorithm achieved a precision of 90.0%， a recall rate of 86.7%， and a mean average precision （mAP） of 89.5% at an intersection over union （IoU） threshold of 0.5 on the target test set. Ablation experiment results showed that the introduction of each of the three improved modules alone achieved stable performance improvement， and their joint use produced significant synergistic gains. Compared with the original YOLOv8 baseline model， the four core metrics were improved by 7.4%， 4.0%， 4.6%， and 7.9%， respectively. The comprehensive performance was significantly superior to mainstream object detection algorithms such as RetinaNet， EfficientDet， and YOLOv5. With only a small increase in the number of parameters and computational cost， the improved algorithm effectively solved the problems of false detection and missed detection of dust concentration levels， greatly improved the classification discrimination accuracy in critical concentration regions， and clarified two engineering application strategies for abnormal sample processing， balancing detection accuracy in normal scenes with response capability in abnormal scenes.

Conclusion The improved algorithm proposed in this study can realize real-time processing of dust scattering images and rapid classification of concentration levels. It accomplishes accurate mapping from continuous concentration values to discrete risk levels， avoids the drawback of excessive computational load of traditional inversion methods， and verifies the feasibility of converting continuous concentration measurement into interval-based concentration early warning. This work can provide an efficient technical solution for the quantitative assessment of industrial dust risks. Future work could further enhance its engineering applicability through model lightweighting and edge deployment.

Keywords： dust detection； dust concentration level detection； deep learning； feature enhancement

Get Citation: Liu Dandan， Guo Chiheng， Guo Shengjun， et al. Deep learning-based algorithm for dust concentration level detection［J］. China Powder Science and Technology， 2026， 32（4）： 1-10.

Received: 2026-01-10, Revised: 2026-05-11, Online: 2026-06-04.

DOI：10.13732/j.issn.1008-5548.2026.04.012

CLC No.：TP391.4； TD76； TB4

Type Code: A

Serial No.：1008-5548（2026）04-0001-10