Electronic Nose System Based On Intrinsic Motivation Learning Mechanism

Electronic nose technology has been developed for decades and has been widely used in environmental monitoring,food safety and medical diagnosis.Like other human perceptions,smell is a process of Active Perception,which can be described by the Markov Decision Process(MDP),reinforcement learning is an important method to solve MDP problems.In recent years,reinforcement learning algorithms(such as DQN,A3C,and etc.)combined with deep learning had made great breakthroughs and received more and more attention.The traditional reinforcement learning algorithms rely on the external reward signals,but cannot be applied when external rewards are sparse or absent.While organisms can learn with sparse external reward signals or with no external reward signals.Based on bionics of biological learning,this paper proposes a reinforcement learning framework based on intrinsic motivation learning mechanism,which simulates that biology produces internal reward signals(such as curiosity and empowerment)in the learning process,and makes up for the defects of reinforcement learning through the interaction of internal reward signals and external reward signals.In view of the deficiency of the electronic nose at present,this paper applies the reinforcement learning framework based on intrinsic motivation learning mechanism to improve the performance of the electronic nose.The following are the main research:(1)The electronic nose cannot compare with the biological olfactory due to its electronic properties,quantity and other reasons,this paper tries to improve the inlet airway design,sensor placement and modulate sample speed dynamically,convert the traditional static classification algorithm to dynamic markov decision process in order to make full use of the dynamic characteristics of signal,so as to make up for the lack of number and specificity of sensors,and other aspects;(2)The on-policy classification learning using reinforcement learning algorithm combined with deep learning can effectively solve the problems such as feature extraction and state space dimension disaster and others.In order to solve the problem that reinforcement learning relies on external rewards,this paper proposes a reinforcement learning framework based on intrinsic motivation learning mechanism,through the combination of internal rewards and external rewards,exploration and exploitation problem in reinforcement learning can be effectively solved;(3)Two designs of intrinsic motivation signals are proposed,one is curiosity or novelty based on prediction error,and the other is maximum channel capacity based on information theory.The motivation signal based on the prediction error is to encourage to explore the state space of the prediction error of the internal model,while the maximum channel capacity based on information theory encourages the exploration to output the state space that action can influence observation(input),so that to give the Agent higher empowerment;(4)Based on the above research,designed the classification of yellow rice wine and VOC gas classification experiments to verify the proposed reinforcement learning framework of intrinsic motivation mechanism,the experimental results show that compared with traditional reinforcement learning,the frame presented in this paper does not need external reinforcement signals;compared with the traditional classification algorithm based on steady-state signals,it can make full use of the rich structure of dynamic signals and shorten the detection time while maintaining the accuracy of classification.