用于白酒识别的电子鼻数据分析与参数优化

Data Analysis and Parameters Optimization of Electronic Nose Systems for Chinese Liquors Recognition

对电子鼻中数据分析(包括数据预处理、特征生成、特征降维和分类识别)问题进行研究.首先提出了将消除工频干扰、小波阈值去噪和数据归一化联用的电子鼻数据预处理新方法;然后从相对电导变化率及其微分、积分、曲率和平均数不等式的角度出发生成110维的初始特征空间;接着采用8种特征选择算法综合降维至41维,再利用核熵成分分析提取12维新特征;最后分别采用Softmax回归和改进的BP神经网络进行分类识别.在数据分析的基础上考察实验参数——气路流量和水浴蒸发温度对白酒识别的影响,同时结合主成分分析和线性判别分析得出较佳气路流量为200,sccm、水浴蒸发温度为70,℃;并应用此参数对11种浓香型白酒进行识别,经温湿度补偿后BP神经网络的识别准确率达91.36%.

The data analysis(including data preprocessing,feature generation,feature reduction and classification) and parameters optimization of electronic nose were investigated. Firstly,a data pre-processing method,in the se-quence of eliminating power-frequency interference,wavelet threshold filtering,and then normalization,was pro-posed to obtain the relative variation ratio of conductance. Secondly,One 110-dimensional feature space was gener-ated from the relative change rate of conductivity,its derivative,integration,curvature and inequality of mean. Thirdly,the 110 features were reduced to 12 new features. A new method of feature reduction which combined fea-ture selection with feature extraction was proposed. The feature selection method used 8 feature selection algorithms based on information theory and the dimension of the feature space was reduced to 41. Then kernel entropy component analysis was employed to extract the 12 new features. Finally,classification of Chinese liquors was performed using the Softmax regression and the improved back propagation artificial neural network(BP-ANN). The effects of gas flow and evaporation temperature of water bath were studied and the optimum parameters(200,sccm gas flow,70,℃evaporation temperature)were acquired for later detection of 11 kinds of strong-flavor Chinese liquors. Equipped with the temperature and humidity compensations,the BP-ANN result shows that the accuracy rate was 91.36%.