基于分形维数和HHT的蚜虫刺吸电位波形机器识别

昆虫刺吸电位(electricalpenetrationgraph,EPG)技术在研究刺吸式昆虫取食行为、昆虫与植物的关系、昆虫传毒机制、作物抗虫机制等方面得到了广泛的应用,然而EPG信号的识别和分析一直是靠人工进行,亟需波形自动识别系统以提高分析效率。刺吸式昆虫取食植物时,产生的EPG波形跟昆虫和植物的种类有关,不同类别的刺吸式昆虫EPG波形差别很大,即使是同种类型的EPG波形其幅值和频率间也会有差异,这给EPG波形的机器识别带来了困难。该文以蚜虫的EPG信号为研究对象,对np波、pd波、E1波、E2波、G波、C波和F波的特征提取和分类识别进行了研究,提出了一种基于分形维数、希尔伯特-黄变换(hilbert-huang transform,HHT)和决策树的EPG波形识别方法。首先对EPG仪器采集得到的信号进行去噪预处理,分别提取分形维数和HHT共10维特征,组成不同维数的特征向量进入决策树分类器进行对比试验。试验结果表明,可采用分形盒维数、Hurst指数、前2层的谱质心和加权频率融合的6维特征向量获得较高的识别率。在EPG波形的机器识别中采用6维特征向量输入的决策树进行分类,通过对4组不同样本进行测试,得到了92.14%、89.29%、95%和89.29%的识别率,平均识别率为91.43%。研究结果表明该文提出的基于分形维数和HHT的特征提取方法以及构建的决策树分类器具有一定的可行性,可为研发EPG信号自动识别分析系统提供理论参考。

Machine identification of electrical penetration graphic waveforms of aphid based on fractal dimension and Hilbert-Huang transform

Insect electrical penetration graph (EPG) technology has been widely applied in researching the feeding behavior of piercing-sucking insects, the relationship between insects and plants, insect transmission mechanism and crop resistance mechanism. However, the identification and analysis of EPG signals have been carried out manually, it is urgent to develop the automatic identification system of EPG waveforms to improve the efficiency. EPG waveforms produced by piercing-sucking insects are related to the insects and plant species, and the EPG waveforms of different types of piercing sucking insects vary greatly, and even the same type of EPG waveform has different amplitude and frequency, which brings difficulties to machine recognition of EPG waveform. EPG waveform is a time series, and its irregularity can be described by fractal theory, fractal theory can reveal the similarity of local part with the whole of the EPG waveform in a certain aspect, the fractal dimension (FD) of the EPG waveform can reflect the characteristic change and the complexity of the geometric shape. EPG waveform belongs to the bioelectrical signal and is nonlinear and non-stationary in nature. Hilbert-Huang transform (HHT) is a powerful tool for analyzing time-varying non-stationary signals, it decomposes the nonlinear signal into several single-mode signals, and adaptively selects the transforming substrate according to the signal itself, so that the bioelectrical signal can be decomposed in essence. In this paper, the EPG signals of aphid were taken as the research object, the feature extraction and classification of np, pd, E1, E2, G, C and F waveform were studied. An EPG waveform recognition method based on fractal dimension, HHT and decision tree was proposed. Firstly, the signals collected by the EPG instrument were denoised and preprocessed, then the features of fractal dimension and HHT were extracted respectively, and the different dimensions vectors were put into the decision tree classifier for comparative experiments, decision tree was used as a classifier, which was generated by C4.5 algorithm. In the process of constructing decision tree, there were 2 main steps: one was to select attribute by information gain ratio, and the other was to complete classification by post-pruning method. In machine recognition of EPG waveform, six-dimensional feature vectors were used as input signals, and 4 groups of samples were tested. The experimental results showed that the six-dimensional feature vectors with fractal box dimension, hurst exponent, spectral centroid and weighted frequency of the first 2 layers had the highest recognition rate. After 10 steps of pruning, the decision tree completed classification, and the recognition rates of the 4 tested groups were 92.14%, 89.29%, 95% and 89.29% respectively. By analyzing the confusion matrix of the 4 groups of test data, it could be seen that the np, E1 and G waveform could be accurately identified, the recognition rate of E2 and C waveform was low, which was prone to misjudgment, this was because that there was no obvious difference between the extracted characteristic values (such as box dimension, spectral centroid of the first 2 layers and weighted frequency of the second layer), C waveform was the most complex of all waveforms, which usually containing A, B waveform and some unrecognizable waveform, and was easy to be confused with other waveforms. The same test samples used for machine recognition were adopted in manual classification. The experimental results showed that the average recognition rate of artificial recognition was 99.11%, the average recognition rate for machine recognition was 91.43%, which was lower than the artificial recognition by 7.68 percent point, average time of the machine recognition was 18.22 s, which was only about 1/46 of that of artificial recognition 839.13 s. The proposed feature extraction method based on fractal dimension and HHT and the constructed decision tree classifier were feasible, which provided a theoretical reference for the research and development of EPG signals automatic identification and analysis system. This research can shorten the analysis time of EPG signal, accelerate the progress of scientific research, and promote the efficient use and intelligent development of EPG.