Study On The Long-term Sensor Drift Of Air Monitoring Electronic Nose

Electronic nose (E-nose), rapidly developed since1990s, is an instrument whichcan analyze, recognize and detect complicated odors and most of the volatile gases.Electronic nose is also called artificial olfactory system. It can seize the odors’characteristics, and accurately identify different smells and its concentration based onthese characteristics. An E-nose is mainly composed of three parts, namely the sensorarray, signal preprocessing unit and pattern recognition unit. It has the characteristics ofrapid detection and real-time online lossless or non-invasive monitoring. Taking theapplication of E-nose in air quality monitoring as the research background, this thesisfocuses on the study of the drift problem for metal oxide semiconductor sensors in theE-nose. The long-term drift of gas sensors is mainly studied in this thesis. There are tworesearch directions on the study of sensors’ long-term drift. One is using statisticalknowledge to counteract drift according to its law or feature. The other is adaptive driftcompensation. From both points of view, relevant algorithms are proposed or utilizedrespectively to solve the problem of long-term drift.For the drift law prediction, a drift compensation algorithm based on predictionmodel is proposed in this thesis. The algorithm uses support vector regression (SVR)models to predict sensor responses without drift according to temperature, humidity andconcentration information. Then, the law of sensor response variation before and afterdrift is analyzed. On the basis of this, drift compensation algorithm models areestablished. In this thesis, data samples, of which the corresponding sensor workingtime is212to580days, are used to analyze drift law. Accordingly, two algorithmmodels namely SVR model and RWLS regression model are established and applied tothe drift compensation for data samples of212thto580thday. What’s more, PCA-CC isalso used for contrast as a typical method. At last, all the three methods are found tohave certain compensation effects. Among them, SVR model performs best, RWLSmodel ranks second. The effect of PCA–CC is relatively not that obvious.For the adaptive drift compensation, an adaptive linear calibration algorithm basedon covariance matrix adaptation evolution strategy (CMA-ES) is adopted in this thesis.A short-term time interval is defined as a time window in this method. It is assumed thatthe drift during a time window can be regarded as linear, which can be compensated bya linear model. The long-term drift then can be compensated according to time windows.CMA-ES is used for periodically updating the linear calibration matrix for samples in a time window, so as to achieve the purpose of adaptively tracking the change of drift. Inaddition, in the application of compensating concentration predicting accuracy, ICA isused to extract the component corresponding to the source of formaldehyde response, soas to eliminate the interference of concentration information when calibrating driftresponses. This method is used in the application of long-term baseline calibration andconcentration predicting accuracy compensation respectively, and shows good ability ofdrift compensation in both situations.For further validating the actual application effect of the algorithms in this thesis,the algorithms were programmed and downloaded into the air quality monitoringE-nose respectively, and experiments were designed for formaldehyde measurementverification. At the same time, PCA-CC is also used for reference and comparison.These three algorithms are verified by experiments and their effectiveness ofcompensation from good to poor can be arranged as adaptive linear calibration, SVR,RWLS and PCA-CC.