基于改进YOLOv4的扫描电镜磨粒图像智能识别

磨损颗粒分析是设备磨损故障诊断和预测的有效手段，为了提高磨粒检测的自动化和智能化程度，提出1种基于改进YOLOv4的目标检测算法，并应用于航空发动机扫描电镜磨粒图像识别. 首先，新算法采用VoVNetv2-39替换YOLOv4原主干网络CSPDarknet53，并引入BiFPN特征金字塔结构与新主干相连，同时调整模型中所有3×3标准卷积为深度可分离卷积，以加强多层次特征融合，构造轻量级网络；其次，利用迁移学习解决扫描电镜磨粒图像数量较少的问题，并通过冻结训练加速模型训练过程；最后，应用实际发动机扫描电镜磨粒图像验证，结果表明：新算法相较于原YOLOv4网络，在保证精度的前提下，网络参数量大幅降低，识别速度提升51.1%，满足实际扫描电镜磨粒图像快速、简洁和高精度的检测需求，具备潜在的工程应用价值.

Wear particle analysis is an effective approach for equipment wear fault diagnosis and prognosis. To enhance the automation and intelligence level of abrasive particle detection, this paper proposes an improved YOLOv4-based object detection algorithm and applies it to the recognition of scanning electron microscope (SEM) images of abrasive particles from aero-engines. Firstly, the new algorithm replaces the original backbone network CSPDarknet53 of YOLOv4 with VoVNetv2-39, introduces the BiFPN feature pyramid structure connected to the new backbone, and converts all 3×3 standard convolutions in the model to depthwise separable convolutions to strengthen multi-level feature fusion and construct a lightweight network. Secondly, transfer learning is utilized to solve the problem of insufficient quantity of SEM abrasive particle images, and frozen training is adopted to accelerate the model training process. Finally, validation is carried out using actual SEM abrasive particle images from aero-engines. The results demonstrate that compared with the original YOLOv4 network, the proposed algorithm greatly reduces the number of model parameters while maintaining detection accuracy, with a 51.1% increase in recognition speed. It meets the requirements of fast, efficient and high-precision detection for actual SEM abrasive particle images, and has potential engineering application value.

搜集汇总

背景与挑战

背景概述

该数据集聚焦于改进YOLOv4算法在扫描电镜磨粒图像识别中的应用，旨在提升航空发动机磨损颗粒检测的自动化和智能化水平。通过优化网络结构和训练方法，实现了轻量级网络设计，显著提高了识别速度，具有工程应用价值。