Turbulent plume analysis
Chemical plumes occur when a chemical substance is released into a fluid with a relatively stable flow, such as wind or water. Everyone is familiar with the smoke plumes released by cigarettes or chimneys; nonetheless, most chemical plumes are invisible. Plumes are important because they provide key information about the location and strength of the chemical source. People often use their nose to detect and track chemical plumes, for example to find rotten food in the kitchen. Some animals (e.g., dogs, insects, bacteria, …) routinely use plumes for mating, foraging or detecting the presence of predators. The idea of using a mobile robot equipped with chemical sensors for plume tracking, e.g. in applications such as gas leak detection or environmental monitoring, has been a focus of research since the 1990s; however, with limited success. The main challenges come from the rapid dilution of concentration with increasing downwind distance, the chaotic dispersion due to atmospheric turbulence, and the limitations of current chemical sensors in terms of sensitivity, selectivity, and response time.
Throughout the last years, we have explored signal processing and machine learning techniques to improve the performance of low-cost chemical sensors for decoding the hidden information contained in chemical plumes. In this paper, we propose and optimize a linear-phase digital differentiator to extract transient features from chaotic gas sensor signals responding to a turbulent chemical plume.
In a wind tunnel scenario and assuming a constant chemical emission rate, these features can predict the distance of a gas source with high accuracy (8 cm error in a distance range of 1.5 m). Compared to statistical descriptors of the sensor signal (mean, variance, maximum, etc.), this represents a 3-fold improvement. In this other paper, we focus on the optimization of the filter parameters and the noise threshold to make the predictions robust against changing wind conditions, without the use of an anemometer. This represents an important advantage over previous approaches, which require wind measurements to produce accurate predictions. The proposed method is succesfully demonstrated in a wind tunnel scenario where the wind speed varies in the range 10-34 cm/s.