卷积神经网络在昆虫刺吸电位波形识别中的应用研究

Application of convolutional neural network in recognition insect EPG waveforms

昆虫刺吸电位(Electrical Penetration Graph,EPG)波形一直以来靠人工识别,不仅耗时费力,且主观性强。针对这一问题,文中提出一种利用深度学习中的卷积神经网络对其进行自动识别的方法。实验中首先对获取的EPG波形进行去噪、分帧等预处理;然后进入一维卷积神经网络进行训练,通过对卷积层数、卷积核大小、学习率、迭代次数等参数进行对比选择,确定两个卷积层和池化层的网络结构,得到了97.5%的平均识别率。这是深度学习在EPG波形识别方面做的初次尝试,相比于传统的机器学习方法,具有更高的识别性能。实验结果表明,文中提出的基于一维卷积神经网络的EPG波形识别方法切实可行。

The electrical penetration graph(EPG)waveform of insects has always been manually recognized,which is time⁃consuming and subjective.An automatic recognition method based on convolution neural network in deep learning is proposed to solve above problems.In the experiments,the obtained EPG waveform is preprocessed by de⁃noising and framing,and then trained by means of the one⁃dimensional convolution neural network.By comparing and selecting the convolution layer,convolution kernel size,learning rate,iteration times and other parameters,the network structure of two convolution layer and pooling layer are determined to acquire the average recognition rate of 97.5%.This is the first attempt of deep learning in EPG waveform recognition.In comparison with the traditional machine learning method,this method has higher recognition performance.The experimental results show that the EPG waveform recognition method based on one⁃dimensional convolutional neural network proposed in this paper is feasible.