摘要： 目前多数白酒分类方法需要进行特征选取,但特征选取算法会增加计算复杂度,限制特征数量,而且选取结果的好坏直接影响识别效果。为此,提出应用压缩感知理论对白酒香型进行分类的方法。通过压缩感知对白酒飞行时间质谱进行整体分析,运用训练数据构造冗余字典作为稀疏基,选择高斯随机矩阵作为测量矩阵,通过求解最小l1 范数得到反映白酒香型特征的稀疏表示,进而根据K 近邻法(KNN)实现对白酒香型的分类识别。将4 种不同重构算法分别结合最小冗余误差和KNN 进行香型分类,实验结果表明,将压缩感知用于白酒香型分类是可行的,能避免特征选取的问题,其中采用稀疏度自适应匹配追踪算法求解l1 范数,并根据KNN 进行分类的稳定性较好,准确率达到91. 45% 。

关键词: 压缩感知, 飞行时间质谱, 稀疏表示, 白酒香型, K 近邻法, 最小冗余误差

Abstract: Most present liquor classification methods need feature selection,but the feature selection algorithm will increase the computational complexity and limit the number of the characteristics. The selection result directly affects the recognition results. Therefore,this paper applies the Compressive Sensing(CS) theory into holistic analysis for Time-offlight Mass Spectrometry(TOFMS) of liquor. Using the training data to form the over complete dictionary and taking it as a sparse matrix,the Gaussian random matrix builds the measurement matrix. By calculating the minimum l1 norm solution,it obtains the sparse representation of the liquor aroma,then realizes liquor aroma recognition based on the KNearest Neighbor(KNN) algorithm. Combining four reconstruction algorithms with minimum residual error and KNN classify liquor aroma,experimental results show that it is feasible to use CS for classification of liquor aroma,and it can avoid the problem of feature selection. Using Sparsity Adaptive Matching Pursuit ( SAMP ) to solve l1 norm and recognition with KNN has a accuracy rate about 91. 45% and better stability.

Key words: Compressive Sensing (CS), Time-of-flight Mass Spectrometry (TOFMS), sparse representation, liquor aroma, K-Nearest Neighbor(KNN) algorithm, minimum residual error

中图分类号: 

• TP18