欢迎访问林业科学,今天是

林业科学 ›› 2020, Vol. 56 ›› Issue (3): 100-108.doi: 10.11707/j.1001-7488.20200311

• 论文与研究报告 • 上一篇    下一篇

基于轻量级神经网络的2种害虫钻蛀振动识别方法

孙钰1,3,脱小倩1,蒋琦2,张海燕1,*,陈志泊1,宗世祥2,骆有庆2   

  1. 1. 北京林业大学信息学院 北京 100083
    2. 北京林业大学林学院 北京 100083
    3. 北京航空航天大学网络空间安全学院 北京 100191
  • 收稿日期:2019-06-04 出版日期:2020-03-25 发布日期:2020-04-08
  • 通讯作者: 张海燕
  • 基金资助:
    北京市科技计划"影响北京生态安全的重大钻蛀性害虫防控技术研究与示范"(Z171100001417005)

Drilling Vibration Identification Technique of Two Pest Based on Lightweight Neural Networks

Yu Sun1,3,Xiaoqian Tuo1,Qi Jiang2,Haiyan Zhang1,*,Zhibo Chen1,Shixiang Zong2,Youqing Luo2   

  1. 1. School of Information Science and Technology, Beijing Forestry University Beijing 100083
    2. School of Forestry, Beijing Forestry University Beijing 100083
    3. School of Cyber Science and Technology, Beihang University Beijing 100191
  • Received:2019-06-04 Online:2020-03-25 Published:2020-04-08
  • Contact: Haiyan Zhang

摘要:

目的: 设计轻量级神经网络,使用声音识别技术构建钻蛀振动识别模型,自动识别双条杉天牛和臭椿沟眶象幼虫蛀干取食振动,为提高钻蛀性害虫的早期预警能力提供技术支撑。方法: 在接双条杉天牛、臭椿沟眶象幼虫木段中嵌入AED-2010L便携式声音探测仪的SP-1L压电式传感器探头,使用录音笔以音频格式记录钻蛀振动信号。双条杉天牛钻蛀振动、臭椿沟眶象钻蛀振动和无钻蛀振动3种声音信号经端点检测、时间规整操作后,计算对数梅尔声谱作为卷积神经网络学习和识别的数据集。由于钻蛀性害虫取食振动脉冲持续时间短,数据量远小于图像,为避免模型出现过拟合,设计轻量级卷积神经网络InsectFrames,网络包含4层3×3卷积,全连接层前接全局平均池化进一步降低网络参数量。使用不同的中间层特征维度和降维方法,实现4种网络变体结构InsectFrames_1—4。结果: 基于轻量级卷积神经网络的钻蛀振动识别方法可有效监测早期虫害的发生,较准确地识别害虫种类。利用InsectFrames_1—4模型,对双条杉天牛钻蛀振动、臭椿沟眶象钻蛀振动和无钻蛀振动3种信号进行识别,在测试集上的平均识别准确率均达90%以上,CPU上平均识别时间为0.1~1.3 s。InsectFrames_2模型识别准确率达95.83%,较广泛用于虫声识别的高斯混合模型提高34.2%,较传统重量级神经网络ResNet18提高6.94%,时间效率提高171.1倍。结论: 将神经网络和声音识别技术用于幼虫钻蛀振动的自动化侦听,具有高效、简单、成本低等优势,可提升林业钻蛀性害虫的早期预警能力。

关键词: 钻蛀性害虫, 神经网络, 钻蛀振动, 声音识别

Abstract:

Objective: In order to early warn stealthy wood boring pest, lightweight neural networks were designed and implemented in the present study, and the acoustics recognition technique was used to automatically identify larvae boring vibrations of Semanotus bifasciatus (Coleoptera:Cerambycidae) and Eucryptorrhynchus brandti (Coleoptera:Curculionidae). Method: The SP-1L piezoelectric vibration probe of the AED-2010L sound pick-up was embeded into the woods sections with larvae of S.bifasciatus and E.brandti, the boring vibrations were recorded by the sound recorder. The sounds of S.bifasciatus, E.brandti and no insect were preprocessed by endpoint detection algorithm and time warping algorithm to calculate the log mel-spectrograms. The log mel-spectrograms were then fed into the convolutional neural networks (CNNs). The boring vibrations were short impulses, thus their size was far smaller than that of images. In this paper the InsectFrames of a light-weight CNN were designed in order to prevent over-fitting of the mode. The insectFrames consisted of four convolutional layers using 3×3 kernels. Global average pooling was located before the connected layer, which can reduce parameters of network.The present study proposed four different CNNs which named InsectFrames_1-4 based on different feature interlayer dimensions and dimensionality reduction method. Result: The vibration identification method for identifying the two insect boring vibrations based on neural networks was able to efficently monitor the occurrence of forest boring insects early, and accurately identify the pest species.The sounds of S.bifasciatus, E.brandti and no insect were recognized using InsectFrames_1-4.The average accuracy on the test set reached more than 90%, and the average recognition time of CPU was about 0.1-1.3 s. Among them, model InsectFrame-2 was the best and its accuracy was 95.83%, and the accuracy was 34.2% higher compared to the Gaussian mixture model which is widely used for insect sound recognition.The accuracy increased by 6.94% compared to the heavy-weight ResNet18 model designed for image classification, and the time efficiency increased by more than 170 times. Conclusion: The proposed method applies neural network and sound identification technique to automatic monitoring of larvae boring vibrations, which can improve the early warning of forestry boring insects, and has the advantages of high efficiency, simplicity and low cost.

Key words: wood boring insect, neural network, boring vibration, sound recognition

中图分类号: