Research on micro-arc oxidation discharge spark detection technology based on MO-DETR

Abstract During microarc oxidation (MAO), the unique characteristics of the discharge spark have a profound effect on the morphology, structure and surface properties of the oxide film. However, due to the high-pressure environment and the complexity of the spark characteristics, it is difficult to accurately observe and analyze these features by conventional detection methods. To this end, we propose a novel target detection model, MO-DETR, which combines the improved CSPDarknet53 as a backbone and a new lightweight encoder, DEEDP (DASi-enhanced feature diffusion pyramid), to systematically process and fuse multiscale features, reduce model complexity and increase processing speed. MO-DETR uses DETR’s multilayer decoder architecture to achieve real-time detection in multiple scenarios and without retraining. On our homegrown MAO dataset MO-Dataset, MO-DETR achieves an average precision of 92.1% and 135.9 FPS on 1080ti GPUs, which outperforms the Yolo series detectors overall. Compared to RT-DETR, MO-DETR achieves a 1.2% improvement in accuracy (from 90.9% to 92.1%), an 82% reduction in parameter volume (from 31 987 850 to 5 728 600), a 91.6% reduction in arithmetic requirements (from 103.4 GFLOPs to 11.8 GFLOPs), and a 71% speedup on the same task (from 39.6 FPS to 135.9 FPS). These improvements significantly increase performance and reduce hardware requirements, making it more practical for real-world applications.