“Signal Processing and Machine Learning for Gas Sensors: Gas Source Localization with a Nano-Drone”.
Javier Burgués, Signal and information processing for sensing systems group
Chemical source localization (CSL) by autonomous robots has been a topic of research since the early 1990s and still today remains elusive beyond simple scenarios. It has numerous potential applications, such as the localization of toxic emissions, malodors, gas leaks and hazardous substances in general, without risking human lives. An intuitive CSL approach is to mimic the known chemo-orientation behaviour of some flying insects, such as moths and mosquitos, which effectively use odor plumes for mating and foraging. However, terrestrial robots are too slow to perform insect-like movements and the response time and limit of detection (LOD) of current odor sensors for key compounds of biological relevance for plume navigation is orders of magnitude higher than in biological chemoreceptors. Instead of using a slow terrestrial robot equipped with complex instrumentation, in this thesis we address the CSL problem with a nano-drone, i.e. a miniaturized aerial robot, equipped with a simple metal oxide semiconductor (MOX) sensor. Improving key specifications of MOX sensors for this application is one of the core parts of this thesis. Specifically, we introduce novel signal processing methods for estimating and optimizing the LOD, reducing the power consumption and improving the response time. We propose a univariate LOD optimization method based on linearized calibration models and a multivariate approach based on orthogonal partial least squares (O-PLS). To improve the response time, we use high-frequency features extracted from the MOX signal derivative, which are optimized for changing wind conditions and real-time operation. A novel setup consisting on a 3D grid of MOX sensors is proposed for real-time visualization of the gas distribution. Two map-based CSL strategies are finally evaluated using the nano-drone in experiments performed in a large indoor environment (160 m2) where a chemical source is placed in challenging positions for the drone. The experimental results demonstrate that the proposed nano-drone can quickly (< 3 min) build a rough gas distribution map (3D) of the environment and localize the main chemical source within it with small errors.