激光多普勒测振技术无损检测果冻橙粒化病  

Non-destruction detection of jelly orange granulation disease using laser Doppler vibrometry

作　　者：刘智 赖庆荣 张天禹 李斌 宋云峰[3] 陈楠 LIU Zhi;LAI Qing-rong;ZHANG Tian-yu;LI Bin;SONG Yun-feng;CHEN Nan(School of Mechatronics&Vehicle Engineering,East China Jiaotong University,Nanchang 330000,China;National and Local Joint Engineering Research Center of Fruit Intelligent Photoelectric Detection Technology and Equipment,East China Jiaotong University,Nanchang 330000,China;Ningbo Sunny Instrument Co.,Ltd,Yuyao 315400,China)

机构地区：[1]华东交通大学机电与车辆工程学院,江西南昌330000 [2]华东交通大学水果智能光电检测技术与装备国家地方联合工程研究中心,江西南昌330000 [3]宁波舜宇仪器有限公司,浙江余姚315400

出　　处：《中国光学(中英文)》2025年第1期160-169,共10页Chinese Optics

基　　金：国家重点研发计划(No.2022YFD2001804,No.2023YFD2001301);国家自然科学基金(No.12304447)。

摘　　要：粒化是柑橘类水果常见的一种内部病害,患有该病害的水果外部特征并不明显,难以从外观上直接识别出来。本文使用微型激光多普勒测振仪(micro-LDV)和共振喇叭搭建了一套声学振动实验装置,将其用于采集“爱媛38号”果冻橙的振动响应信号。然后,将一维的振动响应信号转换为振动多域图像,并构建了一个Resnet-Transformer(ResT)网络,用于提取振动多域图像中的深层特征,以识别果冻橙粒化病。本文中,使用振动多域图像分别训练ResT、Resnet50和Vision Transformer(ViT)模型,并将ResT的性能与Resnet50和ViT进行比较。最后,使用振动多域图像纹理特征或振动频谱特征训练偏最小二乘判别分析(PLS-DA)和支持向量机(SVM)模型,并与ResT模型进行性能对比。结果表明,使用振动多域图像训练的ResT模型可以精准识别果冻橙粒化病并且检测准确率为98.61%,模型的F1为0.986、精确率为0.986、召回率为0.986。由上述结果可知,提出的方法可在简单、快速、低成本的前提下准确识别粒化果冻橙。Granulation is a common internal disease of citrus fruits,and it is difficult to identify the fruits with this disease from their external features.In this study,an acoustic vibration experimental setup was constructed using a micro-laser Doppler vibrometer(micro-LDV)and a resonance speaker.This was used to collect vibration response signals of‘Aiyuan 38’jelly orange.The one-dimensional vibration response signals were converted into vibration multi-domain images,and a Resnet-Transformer network(ResT)was constructed to extract deeper features from the vibration multi-domain images for identifying granulation disease in jelly oranges.In this paper,the ResT,Resnet50,and Vision Transformer(ViT)models were trained using vibration multi-domain images,and their performances were compared.Then,partial least squares discriminant analysis(PLS-DA)and support vector machine(SVM)models were trained using vibration multi-domain image texture features or vibration spectrum features,and the performance was compared with the ResT model.The results show that the ResT model trained using vibration multi-domain images can achieve accurate identification of jelly orange granulation disease with detection accuracy of 98.61%,model F1 of 0.986,precision of 0.986,and recall of 0.986.The proposed method can accurately identify granulated jelly oranges with simplicity,fast speed,and low cost.

关 键 词：激光多普勒测振 声学振动 柑橘粒化病 无损检测 振动多域图像 

分 类 号：TN247[电子电信—物理电子学]