Rapid Inference of Geographical Location with an Event-based Electronic Nose

Abstract

Sensory information determines successful interaction of an agent with its environment. Animals have been shown to be able to encode rapid fluctuations in odour plumes, but this has yet received little attention in electronic gas sensing. State-of-the-art gas sensors actively modify the sensing site using temperature modulation, which decreases the integration time and increases the discriminability. In this work, we propose a novel approach for asynchronous event sampling for temperature-modulated gas sensor data and investigate the effectiveness of different event encoding schemes for solving an inference problem. A multichannel heater-modulated electronic nose was used to record field data at 1kHz. The data consisted of an approx. 90-minute walk in Lisbon covering multiple chemical environments and olfactory sceneries. Single temperature-cycle sensor conductance windows of 140ms were normalised and a model curve was subtracted from each sensor response. Using send-on-delta sampling, on- and off-events were generated and further compressed by considering either their rate, time-to-first-spike, firing-order, or a reconstructed signal. The different representations were assessed by their performance in inferring geographical location on unseen single 140 ms cycles using a linear SVM (75%/25% training/test split). We found that with a small spiking threshold the event-reconstructed signal achieved 82.5±1.0% accuracy, very close to the raw data (84.2±1.2%), and gracefully degraded when reducing the event count by increasing the spike threshold. The rate, latency and rank-order codes could not match that of the reconstructed signal, suggesting that temporal dynamics of the intra-cycle signal contain essential information. We conclude that heater-modulated gas sensors lend themselves to event-based processing, allowing for rapid inference in the sub-second regime. Our work could pave the way from distinguishing broad olfactory scenes to recognising individual odorants in turbulent plumes, with the potential to break new ground in traditional and neuromorphic gas sensing.