Enhancing Sensor Array Intelligence by Bayesian Fusion of Information Multiplicity Generated by Multiple Processors

The paper presents a new data processing method for sensor array based pattern recognition problem. The primary motivation is to improve the odor recognition efficiency of electronic nose systems. The method creates a set of virtual experts in which individual expert members are defined by a different combination of a feature extractor and a radial basis function (RBF) neural network classifier. In this work the outputs from five different linear feature extraction methods: principal component analysis (PCA), independent component analysis (ICA), singular value decomposition (SVD), linear discriminant analysis (LDA) and partial-least-square regression (PLSR), are fed separately as inputs to five different RBF neural networks. The parameters defining each RBF network are optimized separately by training them as independent decision makers. Since a given feature extractor processes raw data with specific perspective about the data structure, and RBF network generates a set of class likelihood values, the set of virtual experts generate alternate sets of class likelihood values. Bayesian product rule for fusion is then applied for combining these class likelihood values into class posterior probabilities. The class declaration is finally done by maximum posterior probability. The method has been validated by analyzing 9 chemical and 7 non-chemical data sets. The enhancement in classification rate up to 33.3 % has been found. The reason for system performance improvement is that the multiple feature extractors generate varied representations of raw data by exploring diversity of hidden attributes, and Bayesian fusion works on the extended information provided by several experts. Copyright © 2014 IFSA Publishing, S. L.