Research On The Discrimination Between Authentic Panax Notoginseng And Its Fakes Based On Electronic Nose

Electronic nose,also known as artificial olfaction,is composed of a gas sensor array and signal process unit,along with pattern recognition algorithms.This system,designed by imitating the mechanism of human olfaction,is able to discriminate between different categories of mixed gas.In this research,we design and assemble an electronic nose system based on 16 metal oxide semiconductors to control the whole process automatically and save data.Compared with other gas analysis technologies,this system enjoys the advantages of fast response,low cost,easy operation and wide application.In this research,we select authentic panax notoginseng with different head counts and its 7 common fakes as experimental objects.We have two classification goals,one is to explore the multi-classification effect between authentic panax notoginseng(20 heads)and its 7 fakes,another is to distinct between panax notoginseng with different head counts.Using the electronic nose system established above,we design a scientific experimental process,complete the sample gas preparation and perform data collection and storage.Then we select appropriate data pretreatment methods including baseline correction and noise reduction filtering preprocessing on the response signal.We try a whole set of feature engineering methods,including feature generation,feature selection and pattern recognition.In the feature selection part,three feature selection strategies are designed,including recursive feature elimination,feature representation strategy based on multivariate analysis,and feature selection based on XGBoost model.Support vector machine(SVM)is used as the basic classifier to compare their feature selection effect.We also optimize the weighted voting rules where the prior knowledge of the basic classifiers is fully used,and different weight values are set according to the recognition accuracy of different categories of samples,thereby further improving the classification accuracy.Also,the deep learning method is used,including two networks:stacked anto-encoder networks(SAE)and deep belief networks(DBNs).Two schemes are designed for each network:one is to directly output to the classifier after feature reconstruction,another is to add a classifier to the deep network structure and introduce the BPNN back propagation algorithm to guide the fine-tuning of parameters.We not only compare the performances between manual and automatic feature extraction,but also explore the applicability of deep learning methods on electronic nose data.The results show that both feature engineering method and deep learning method can effectively improve the classification effect.In the classification of panax notoginseng and its fakes,manual feature selection combined with machine learning has achieved good classification results.The classification accuracy rate is 87.47%.It is further improved to 91.57%with optimized ensemble learning algorithm,which is higher than any deep learning result.In the classification of panax notoginseng with different head counts,due to the high similarity of the response signals,the accuracy is 73.68%with feature engineering and 74.65%with DBNs+BPNN+Softmax.The deep learning method performs well on complex feature extraction.For the current feature dimension,the number of samples is small,and as the sample size increases,better recognition results can be obtained.