基于复小波及动态神经网络的植物电信号研究

Study of plant electrical signal based on complex wavelet and dynamical neural network

[目的]针对植物电信号数量级小、易受干扰的问题,提出了双树复小波变换(DT-CWT)结合双变量收缩消噪及不带输入变量的非线性自回归神经网络(NAR)模型,旨在能将植物电信号用于研究温室内植物生长模型。[方法]在屏蔽环境下获取生长状况良好的鸟巢蕨植株的电信号。采用双树复小波变换将电信号进行分解,利用层间小波系数具有相关性的特点,将分解后的小波系数进行双变量收缩消噪。通过对植物电信号进行自相关分析,确定迟滞阶数。再通过NAR网络训练消噪信号。[结果]采用双树复小波消噪后的信号虚部树的高频分量明显减少。消噪后的植物电信号前序98个样本点的自相关系数均大于0.8,迟滞阶数98。采用本模型对消噪后的电信号进行预测时相关系数为0.973,均方误差(MSE)为0.593 mv^2。相比于软阈值消噪与硬阈值消噪,本模型的消噪方法信噪比(SNR)最大,MSE最小。对碧玉、白鹤芋2种植物应用本模型,决定系数分别为0.975和0.972,MSE分别为0.112 mv^2和4.459×10^(-2)mv^2。[结论]植物电信号2个相邻时刻间具有很强的关联性,消噪过程对虚部树的影响更大,双树复小波分解结合双变量收缩的消噪方法更大程度上保留了信号的原始信息。本模型具有可推广性。

[ Objectives ] In terms of the small order of magnitude and being susceptible to interference of plant electrical signal, a model was presented and this model integrated dual tree complex wavelet transform （DT-CWT）, bivariate shrinkage denoising and nonlinear autoregressive model（NAR）. In addition,the model aimed to apply plant electrical signal to the study of plant growth models in the greenhouse. [ Methods] Firstly,we observed the electrical signal of Asplenium nidus that was in good condition under controlled environ- ment. Next,we used the dual-tree complex wavelet transformation to decompose electrical signals and took advantage of the correlation of wavelet coefficient between layers. Then the decomposed wavelet coefficients were denoised using by bivariate shrinkage. Based on the analysis of autocorrelation of the plant electrical signal, we determined the order of hysteresis. At last, using NAR for denoising training,plant electrical signal was predicted. [ Results] The high frequency in denoised signal＇s imaginary tree dealt by DT-CWT exhibited significant reduction. After denoising, the autocorrelation coefficient from 98 points data were all greater than 0.8, and the order of hysteresis was 98. When the denoised signal was predicted using this model, the correlation coefficient was 0.973 and the mean square error（MSE）was 0.593 mv2. Compared with both soft and hard threshold denoising methods, our model reached a lowest MSE but highest SNR. Applying this module to Peperomia tetraphylla and Spathiphyllum kochii Engl. ＆ K. Krause, coefficients of determination were 0.975 and 0.972 respectively. Furthermore ,MSE were 0.112 my2 and 4.459× 10-1 mv2 respectively. [ Conclusions ] There is a strong correlation of plant electrical signal between two neighboring time. The processing of denoising has a stronger influence to imaginary tree. The model integrates DT-CWT and bivariate shrinkage to preserve the original information of the signal to a greater degree. This model has a great popularization among other plants.