基于轻量级RepVIT的农机具工况识别方法研究  

Lightweight RepVIT-based Working Condition Recognition Method for Agricultural Implements

作　　者：安麒麟 汪凤珠[1,2] 刘阳春 邓学[1,2] 周利明 赵博 伟利国[1,2] AN Qilin;WANG Fengzhu;LIU Yangchun;DENG Xue;ZHOU Liming;ZHAO Bo;WEI Liguo(Chinese Academy of Agricultural Mechanization Sciences Group Co.,Ltd.,Beijing 100083,China;State Key Laboratory of Agricultural Equipment Technology,Beijing 100083,China)

机构地区：[1]中国农业机械化科学研究院集团有限公司,北京100083 [2]农业装备技术全国重点实验室,北京100083

出　　处：《农业机械学报》2025年第2期187-194,205,共9页Transactions of the Chinese Society for Agricultural Machinery

基　　金：国家重点研发计划青年科学家项目(2022YFD2000300)。

摘　　要：为解决田间复杂环境下拖拉机机载农机具监测困难、模型参数量过大等问题,提出了一种基于轻量化RepVIT的农机具识别模型TMAInet。利用自主开发的农机服务平台“农业机械化精准作业平台暨希望田野”收集了6种工作状态的农机具数据集,并通过Copy-paste等数据增强方法将训练集扩增至6627幅。基于RepVIT网络模型框架,设计了一种卷积前馈模块(CFF)以提升不同尺度细粒度特征提取能力,引入了注意力机制ECA以优化模型参数结构并简化特征提取模块。通过Pre-training+Fine-tuning(PF)迁移学习方法对模型进行了训练,并在Jetson nano边缘设备上进行了部署。实验结果表明,通过PF迁移学习方法,TMAInet模型的识别准确率、F1分数和召回率分别达到99.13%、98.53%和98.78%,相较于原始的RepVIT模型分别提升1.86、3.04、1.95个百分点,在边缘设备端保持帧速率73 f/s的同时参数量降低至7.3×10^(6)。TMAInet能够在实际应用中准确、高效监测农机具常见类别,为无人化智慧农场的发展提供技术参考。Aiming to address the problems of difficulty in monitoring tractor-mounted agricultural implements in complex field environments and the excessive amount of model parameters,a lightweight RepViT-based agricultural implements recognition model,tractor-mounted agricultural implements net(TMAInet),was proposed.Firstly,the self-developed agricultural machinery service platform'Agricultural Mechanisation Precision Operation Platform'was used to collect the datasets of agricultural implements in six working states,and the training set was expanded to 6627 frames by data enhancement methods such as copy-paste.Secondly,based on the RepVIT network model framework,a convolutional feed-forward module(CFF)was designed to improve the ability of fine-grained feature extraction at different scales,and an attention mechanism,ECA,was introduced to optimize the model parameter structure and simplify the feature extraction module.Finally,the model was trained by pre-training+fine-tuning(PF)migration learning method and deployed on Jetson nano edge devices.The experimental results showed that the recognition accuracy,F1 score and recall of the TMAInet model reached 99.13%,98.53 and 98.78%,respectively,by the PF migration learning method.Compared with the original RepVIT model,the recognition accuracy,F1 score and recall were improved by 1.86 percentage points,3.04 percentage points and 1.95 percentage points,respectively,and the number of parameters was reduced to 7.3×10^(6) while maintaining 73 f/s at the edge device side.TMAInet was able to accurately and efficiently monitor the common categories of agricultural implements in practical applications,and it can provide a technical reference for the development of unmanned smart farms.

关 键 词：拖拉机农机具 工况识别 无人化智慧农场 RepVIT 迁移学习 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]