Swarm intelligence and similarity measures for memory efficient electronic nose system

Recursive discrete binary particle swarm optimization (RDBPSO) with Euclidean distance as a fitness function was recently proposed for the electronic nose system (ENS). It demonstrated a significant reduction in the memory requirement and the response time of the system using optimum subset of features. To further enhance the memory requirement and response time of the ENS, this paper proposes some new fitness functions for RDBPSO with following merits: 1) a novel approach based on different similarity measures is proposed with improved memory efficiency and response time of the ENS; 2) feature augmentation strategy that combines the original sensors response and some extracted features is incorporated, which enhances the learning process; 3) the proposed algorithm utilizes only the median values of the training data, which makes the system memory efficient and feasible for portable use; and 4) optimization of features vector using RDBPSO maintains the classification accuracy high, even with a reduced number of features. The proposed algorithm is validated against four different data sets of gases and mixtures of gases. Two of these data sets are experimentally acquired from a laboratory setup, while other two are publically available on the University of California, Irvine machine learning online repository. Validation results reveal an average of 96.77% decrease in the memory requirement, while an average classification accuracy of 98.73% is achieved for all the four data sets.