Digital Twin-Based Improved YOLOv8 Algorithm for Micro-Defect Detection of Labyrinth Drip Emitters in High-Speed...

In water-scarce regions such as Xinjiang, China, agricultural development is constrained not only by limited water resources but also by a strong reliance on water-saving irrigation technologies. Drip irrigation is a key measure for improving irrigation efficiency and promoting the sustainable development of water-saving agriculture. However, defects arising during the manufacture of labyrinth Drip emitters-the core components of drip irrigation systems-can undermine system reliability, leading to channel blockage and non-uniform irrigation. To tackle this issue, a defect detection approach is developed by integrating Digital Twin technology with an enhanced YOLOv8 model for online inspection of labyrinth Drip emitters on drip irrigation tape production lines. In parallel, a self-built dataset covering six defect categories is established. Supported by the DT framework, the standard YOLOv8 network is refined to strengthen its capability in identifying complex micro-defects. Specifically, DySnakeConv is introduced to better represent the curved and slender characteristics of labyrinth channels; DySample is incorporated to improve the reconstruction and representation of fine-grained details; an Efficient Multi-Scale Attention module is adopted to capture richer contextual information while suppressing background noise; and Inner-SIoU is applied to optimize the bounding-box regression process. Experimental results show that the model achieves 89.6% precision, 90.9% recall, and 93.9% mAP50. Compared with the baseline YOLOv8, precision, recall, and mAP50 are improved by 7.3, 3.9, and 3.3 percentage points, respectively. Under the same training conditions, the proposed model outperforms YOLOv10 and YOLOv11 in accuracy-related metrics. Specifically, compared with YOLOv11, precision, recall, and mAP50 are improved by 4.8, 5.0, and 2.6 percentage points, respectively; compared with YOLOv10, they are improved by 10.0, 7.7, and 7.3 percentage points, respectively. Meanwhile, the model maintains a lightweight size of 3.7 M parameters and a real-time inference speed of 150.2 FPS, demonstrating a favorable accuracy-efficiency trade-off. By extending manufacturing-level quality control to agricultural applications, the approach helps ensure uniform irrigation and improve water-use efficiency, providing practical technical support for precision agriculture in arid regions.