Electronic Nose System And Its Application Based On DQN-LSTM

Electronic nose is a technology in realizing the classification of different gases by simulating biological olfaction.It is mainly composed of gas sensor array,data acquisition and processing unit,pattern recognition unit and so on.Compared with the analytical instruments such as chromatograph,the electronic nose has the advantages of simple structure,fast recognition speed and intuitive results.It has been widely used in many fields.However,compared with the biological olfactory system,there is still a big gap between the existing electronic nose,which are mainly reflected in:(1)affected by the structure of the sensors,size and other factors,the number of gas sensor array is far less than the biological olfaction;(2)biological olfaction usually adopts active perception(Active Perception),according to the recognition task to change the sample data(the so-called sniff)dynamicly with different recognition task,can complete the identification quickly.Most of the electronic nose technologies mainly use the traditional pattern recognition algorithm as the one-way and passive way,and the recognition speed and precision are limited.For this reason,this paper intends to solve the following problems:(1)how to increase the selectivity of sensors without increasing the number of sensors;(2)to explore the recognition algorithm based on biological active olfaction mechanism.This paper researches from the existing electronic nose technology at the present stage of the problem,combined with sniffing behavior,biological olfactory nasal mucosa chromatographic effect and reinforcement learning,recurrent neural network algorithm and so on,proposes a DQN-LSTM model to simulate biological olfaction.The main contributions of this paper includes:(1)design and achieve a electronic nose hardware device which can adjust the air speed dynamicly and a test platform;(2)study on the mapping relationship between inlet velocity of different kinds of gases and response values of sensors by experiment;(3)proposed a recursive LSTM neural network algorithm based DQN reinforcement learning algorithm and achieved the on-policy optimization of intake flow rate.The main work is as follows:1.Referring to the relevant information at domestic and abroad,the research status of electronic nose in recent years is described in detail,and expound the research background and significance of the topic.Finally,the shortcomings and solutions of the electronic nose are briefly introduced.2.Starting from the active mechanism of biological olfaction,this paper introduces the characteristics of using flow modulation in detail.The chromatographic effect and sniffing behavior of nasal mucosa can lay a solid theoretical foundation for the study of the following papers.3.On the basis of laboratory project,through the improvement of the hardware design,including re-plate,air cavity structure,improved flow rate selection unit,using STM32F407ZGT6 chip as the main control chip to realize data acquisition,electronic nose hardware system and software system based on velocity modulation.The validity of flow modulation is validated by the data of beer classifaction.4.Based on the velocity modulation and nasal mucosal chromatographic effect,this paper proposed an optimized DQN model and algorithm,which combined LSTM algorithm to realize active selection of electronic nose.The DQN algorithm can select the bestoutput action according to the current state and reward,and the LSTM algorithm can predict the classification according to the output action and the current state,and give the corresponding reward.5.Finally,using the experimental data of VOC gases and Chinese yellow wine to verify the algorithm,and compared with PCA and SVM and other commonly used offline and passive algorithms.Experimental results show that the algorithm can effectively reduce the cost of training and testing,improve the recognition speed of electronic nose,and achieve the purpose of active selection of electronic nose.