| The combination of an array of gas sensors with partially overlapping chemical sensitivities and an appropriate pattern recognition system can be used to identify odors. When exposed to a particular odorous compound or mixture, the response of the gas sensor array contains information that can be processed by a pattern recognition system in order to identify the odor. This emerging technology, called the electronic nose, offers a wide range of applications: environmental pollution monitoring, quality control of food products and rapid diagnosis of diseases, to mention a few. In order for the electronic nose to become a practical solution to real-world problems, advances in gas sensor technology and pattern recognition are required.; This dissertation focuses on the signal processing and pattern recognition issues related to the electronic nose. We present a modular pattern recognition architecture composed of four building blocks: data preprocessing, feature extraction, classification and decision making. The preprocessing stage transforms the raw sensor data in order to compensate for the effects of sensor drift, compress the sensor transient response and improve sample to sample repeatability. The feature extraction stage makes use of inexpensive linear techniques to simultaneously reduce the dimensionality of the problem and emphasize the most discriminatory information for classification purposes. The classification stage takes the feature vector from an unknown sample and produces a label assignment along with a confidence level. The decision-making module oversees the classifier's operation and can modify the assignment if there is any application-specific information available, such as confidence thresholds or misclassification risks. Finally, we adopt a machine-learning approach to feature subset selection that improves the predictive accuracy of the pattern recognition system, reduces its complexity, speeds up the learning task and provides a way to customize sensor arrays for particular odor applications. |
