Gas Sensor Drift Compensation Based On Deep Belief Network

Electronic nose is not only fit for monitoring gases in terms of various gases pollution coming out frequently today,but also has already been applied in diagnosis,industry,food safety and so on,due to the development of computer technique and pattern recognition.Gas sensor drift of electronic nose,which weakens the reliability and practicability of electronic nose,seriously restricts its appliance and generalization.On account of the complicated reasons causing gas sensor drift,such as chemical poisoning,aging and environmental changing,the drift characteristic is highly nonlinear and chaotic.So it's hard to make drift compensation for gas sensors.Deep belief network,which is a typical deep learning structure,owns the ability to fit nonlinear distribution and extract object's deep characteristics through nonlinear mappings by hidden layers.This thesis applies DBN on gas sensor drift compensation on advantages of DBN's superiority above.The specific work is followed:(1)For the highly nonlinear and chaotic of gas sensors' drift,it's hard to make drift compensation on characteristic level.This thesis verifies DBN could restrict gas drift on characteristic level,and figures out the reason is that DBN could strength the relativity between different features by decomposing and reconstructing the characteristic again and again.(2)The parameters design for DBN has been a tough work,because there is no theoretical guidance for parameters design.So this thesis analyzes its parameters in detail and figure out the optimal parameters through experiment,on gas sensor drift compensation.(3)Via summarizing the present adaptive correction methods for drift compensation,a DBN drift compensation method based on transfer learning is proposed.This method could make drift compensation on both characteristic and decision level,which is verified though experiments.