Estimating odor source proximity via two MOS sensors based on deep learning method

Abstract

Odor source proximity estimation (OSPE) is essential in many chemical sensing applications, such as chemical leak accident rescue or odor-guided robot navigation. To improve the accuracy and generalization of OSPE, a new odor source proximity indicator (PI) called dual-sensor distance ratio (DR) is proposed in this paper, and an odor source proximity deep neural network (OP-NET) is established, which can estimate the DR by automatically decoding the source proximity information from the signals of two metal oxide semiconductor (MOS) sensors. A dataset is generated by the simulation and experimental platform for the deep learning-based OSPE research, where the simulation data is used for model hyperparameter optimization, while the experiment data is used for OSPE method comparison, crosswind validation, and generalization test. The results show that the best OSPE errors of OP-NET are 0.49 m, 0.69 m and 0.99 m in the downwind, crosswind, generalization test, respectively. The proposed deep learning-based OSPE model is expected to improve the ability of robot to locate odor sources. © 2001 Elsevier Science. All rights reserved.