Biologically-inspired pattern recognition for odor detection Thaddeus Roppel a, * , Denise M. Wilson b a Department of Electrical and Computer Engineering, 200 Broun Hall, Auburn University, AL 36849-5201, USA b University of Washington, Washington, USA Abstract It is shown that biologically-inspired data aggregation methods result in improved odor discrimination in a gas sensor array. In one typical group of experiments, the correct identi®cation rate achieved by a neural network was improved from 92% without aggregation to 98% with aggregation. Ó 2000 Elsevier Science B.V. All rights reserved. Keywords: Odor recognition; Sensor aggregation; Data fusion 1. Introduction This letter addresses the issue of biologically- inspired pattern recognition applied to odor rec- ognition. In recent years considerable attention has been focused on the scienti®c and engineering principles needed for the development of machines that can perform automated odor detection and odor source localization. These devices are envi- sioned to supplement the use of dogs and other animals in diverse detection applications and to assist manufacturers of consumer goods in re- peatably producing products with odorant com- ponents such as food, plastics, cosmetics, etc. In all cases, the use of nonlinear, adaptive pattern rec- ognition is key since (i) the available sensors have overlapping speci®cities, (ii) the available sensors are subject to large performance variations and (iii) no general agreement has been reached on what constitutes the fundamental components of odor space, i.e., the ``primary colors'' of odor. Therefore it is still problematic to transform and distill the outputs of sensor arrays into informa- tion that is usable by people in a wide spectrum of potential applications. 2. Background Pre-processing of sensor signals has been stud- ied in biological olfaction, leading to the system- level understanding illustrated in Fig. 1 (Dodd and Castellucci, 1991; Kauer et al., 1991). Individual sensory receptors in the biological system possess many of the same types of vari- ability as available arti®cial sensor technologies, a similarity that supports the use of biological in- spiration in designing both preprocessing networks and pattern recognition algorithms. The olfactory system in most higher-level animals contains a huge number of receptors (between 2 and 20 mil- lion in the human olfactory system) that are ag- gregated, compared and contrasted before being presented to a content-addressable memory in the www.elsevier.nl/locate/patrec Pattern Recognition Letters 21 (2000) 213±219 * Corresponding author. E-mail address: troppel@eng.auburn.edu (T. Roppel). 0167-8655/00/$ - see front matter Ó 2000 Elsevier Science B.V. All rights reserved. PII: S 0 1 6 7 - 8 6 5 5 ( 9 9 ) 0 0 1 5 0 - 6