基于AE-BiRNN的片烟含水率高光谱检测

AE-BiRNN-based hyperspectral detection of moisture content in tobacco strips

为解决高光谱数据维度高、预测精度容易受降维方法影响等问题,构建了一种嵌入点积注意力机制的双向循环神经网络AE-BiRNN(Attention mechanism Embed Bidirectional Recurrent Neural Networks)模型用于片烟含水率预测。AE-BiRNN通过双向门控制单元BiGRU(Bidirectional Gate Recurrent Unit)关联片烟含水率光谱序列前后波段的光谱信息,使用基于点积结构的注意力机制根据相似度信息对波段分配不同的权重参数,起到特征波长选择的作用。利用155个尺寸为640 px×640 px的片烟高光谱图像数据开展训练与验证,结果表明:(1)AE-BiRNN在无特征波长选择条件下同样能够获得较好预测准确率,点积注意力机制的引入显著提高了模型收敛速度;(2)AE-BiRNN对验证集回归预测的决定系数Rv2为0.9404,均方根误差RMSEV为0.01471,明显优于PLSR(Partial Least Squares Regression)、RF(Random Forest)、卷积神经网络LeNet和MLP(Multilayer Perceptron)方法。该研究可为实时监测片烟含水率分布、快速调控松散回潮机工艺参数提供支持。

To address the issues including high dimensionality of hyperspectral data and the influence of dimensionality reduction on prediction accuracy,an Attention mechanism Embed Bidirectional Recurrent Neural Networks(AE-BiRNN)model embedded with a dot product attention mechanism was developed to predict moisture content in tobacco strips.The AE-BiRNN model associates the spectral information of the front and back bands of the spectral sequence of moisture content in tobacco strips through the Bidirectional Gate Recurrent Unit(BiGRU),uses the dot product structure-based attention mechanism to assign different weight parameters to the bands according to the similarity information,and plays the role of characteristic wavelength selection.Training and verification were performed on data from 155 hyperspectral tobacco strip images 640 px×640 px in size.The results showed that:1)The AE-BiRNN model could achieve satisfactory prediction accuracy in the absence of characteristic wavelength selection,and the introduction of the dot product attention mechanism significantly increased the convergence speed of the model.2)The determination coefficient R^(2) v of AE-BiRNN for the regression prediction of the verification set was 0.9404 with the root mean square error RMSEV of 0.01471,which were obviously better than those of Partial Least Squares Regression(PLSR)and Random Forest(RF),Convolutional Neural Network LeNet and Multilayer Perceptron(MLP)methods.The method supports real-time monitoring of moisture content in tobacco strips and timely regulation of ordering cylinder parameters.