Pattern recognition system design for sensor application

Feature extraction, classification, and sensor performance evaluation and improvement are three of the key topics in gas identification with chemical sensors. This dissertation is particularly focused on the following research topics: searching a generally useful feature extraction method, applying SVM (Support Vector Machine) and Fuzzy ArtMAP (Adaptive resonance theory) algorithms for a better application to overcome the drawback of BPN (BackPropagation Neural network) and investigating the sensor index. For generally useful feature extraction method, the Nyquist sampling plus PCA (Principal Component Analysis) method is talked and compared with currently used geometrical dynamic feature extraction methods in existing lab platform. It is proved that the Nyquist sampling can work better than dynamic feature extraction methods in out applications. Also, this method is from the signal processing theory and can be widely used for any applications where similar geometrical method can be used. Besides the sampling method, it is investigated how to solve the vapor concentration dependence issue and design a reasonable test protocol. For patter recognition classifiers, the SVM method is used to check if or not the training data size can be reduced for the purpose of accelerating the data processing time and lowering the software algorithm optimization cost. Also, the Fuzyy ArtMAP classifier is used to solve the unknown vapor recognition issue. For sensor performance indices, the separability, sensitivity, response time, detection limit and the reliability are studied. For separability, LDA concept is used to quantify it. For sensitivity and detection limit, diabetes detection application is discussed to check them. The test protocol design covers the response time. The reliability of the sensors is checked by proving the aging issue is not a significant factor for using half a year test result; the significant factor is the temperature and accordingly the temperature modulation idea is proposed and implemented to overcome its influence.