Algorithms for enhancing pattern separability, feature selection and incremental learning with applications to gas-sensing electronic nose systems.

摘要

Three major issues in pattern recognition and data analysis have been addressed in this study and applied to the problem of identification of volatile organic compounds (VOC) for gas sensing applications. Various approaches have been proposed and discussed. These approaches are not only applicable to the VOC identification, but also to a variety of pattern recognition and data analysis problems. In particular, (1) enhancing pattern separability for challenging classification problems, (2) optimum feature selection problem, and (3) incremental learning for neural networks have been investigated.; Three different approaches are proposed for enhancing pattern separability for classification of closely spaced, or possibly overlapping clusters. In the neurofuzzy approach, a fuzzy inference system that considers the dynamic ranges of individual features is developed. Feature range stretching (FRS) is introduced as an alternative approach for increasing intercluster distances by mapping the tight dynamic range of each feature to a wider range through a nonlinear function. Finally, a third approach, nonlinear cluster transformation (NCT), is proposed, which increases intercluster distances while preserving intracluster distances. It is shown that NCT achieves comparable, or better, performance than the other two methods at a fraction of the computational burden. The implementation issues and relative advantages and disadvantages of these approaches are systematically investigated.; Selection of optimum features is addressed using both a decision tree based approach, and a wrapper approach. The hill-climb search based wrapper approach is applied for selection of the optimum features for gas sensing problems.; Finally, a new method, Learn++, is proposed that gives classification algorithms, the capability of incrementally learning from new data. Learn++ is introduced for incremental learning of new data, when the original database is no longer available. Learn++ algorithm is based on strategically combining an ensemble of classifiers, each of which is trained to learn only a small portion of the pattern space. Furthermore, Learn++ is capable of learning new data even when new classes are introduced, and it also features a built-in mechanism for estimating the reliability of its classification decision.; All proposed methods are explained in detail and simulation results are discussed along with directions for future work.