一个用于磨粒图像快速分类的轻量化CNN模型  

A Lightweight CNN Model for Fast Classification of Wear Particle Images

作　　者：刘信良 陈国宁 苏化 王静秋[1] 王晓雷[1] LIU Xinliang;CHEN Guoning;SU Hua;WANG Jingqiu;WANG Xiaolei(National Key Laboratory of Helicopter Aeromechanics,Nanjing University of Aeronautics and Astronautics,Nanjing Jiangsu 210016,China)

机构地区：[1]南京航空航天大学,直升机动力学全国重点实验室,江苏南京210016

出　　处：《润滑与密封》2024年第10期102-107,共6页Lubrication Engineering

基　　金：直升机传动技术重点实验室基金项目(HTL-A-21G03)。

摘　　要：针对磨粒分析CNN模型存在参数多、运算速度慢、难以实际应用等问题,开展磨粒图像分类CNN模型的轻量化研究。通过分析模型各层的参数量、运算量和剪枝敏感度,确定卷积层4和卷积层5为滤波器剪枝的目标;对卷积层4和5所有滤波器重要性进行计算并排序,以75%的剪枝率去除重要性低的滤波器并重新训练,获得轻量化模型。实验结果表明,轻量化后的模型在保证准确率几乎不降低的情况下实现了磨粒图像的快速分类,其理论参数量和内存占用量均减少50%以上,运算速度提高20%以上。研究结果为CNN模型在便携式、移动式铁谱分析设备上的应用提供参考。Aiming at the disadvantages of CNN model for wear particle analysis,such as too many parameters,slow operation speed,and difficulty in practical application,research on lightweight of the CNN model for wear particle image classification was carried out.The parameters,computation and pruning sensitivity of each layer in the CNN model were analyzed,which determined the convolution layer 4 and 5 as the filter pruning targets.The importance of all filters in convolution layer 4 and 5 was calculated,and the filters were sorted by their importance.A lightweight model was obtained after removing the filters with low importance at 75%pruning rate and retraining.The experimental results show that after lightweight processing,the number of theoretical parameters and amount of memory usage are reduced by more than 50%,and the operation speed is increased by more than 20%while the accuracy barely drops.This study provides an idea for the application of CNN model on the portable and mobile ferrography equipment.

关 键 词：铁谱分析 卷积神经网络 磨粒图像分类 轻量化 

分 类 号：TH117.1[机械工程—机械设计及理论] V263.6[航空宇航科学与技术—航空宇航制造工程]