High-speed odour sensing using miniaturised electronic nose

Abstract

Animals have evolved to rapidly detect and recognise brief and intermittent encounters with odour packages, exhibiting recognition capabilities within milliseconds. Artificial olfaction has faced challenges in achieving comparable results -- existing solutions are either slow; or bulky, expensive, and power-intensive -- limiting applicability in real-world scenarios for mobile robotics. Here we introduce a miniaturised high-speed electronic nose; characterised by high-bandwidth sensor readouts, tightly controlled sensing parameters and powerful algorithms. The system is evaluated on a high-fidelity odour delivery benchmark. We showcase successful classification of tens-of-millisecond odour pulses, and demonstrate temporal pattern encoding of stimuli switching with up to 60 Hz. Those timescales are unprecedented in miniaturised low-power settings, and demonstrably exceed the performance observed in mice. For the first time, it is possible to match the temporal resolution of animal olfaction in robotic systems. This will allow for addressing challenges in environmental and industrial monitoring, security, neuroscience, and beyond.

Figures (7)

• Figure 1: Electronic nose and odour delivery system.a, Decoding temporal information of odour plumes requires fast sensing. Top: Two sequential TiCl4 smoke plume photographies, shifted and superimposed, kindly provided by Dr. Paul Szyszka. Bottom: Dual-PID recordings of source-separated odour plumes, from Ackels et al. Ackels2021. Plume and sensor location (red) for illustrative purposes only. b, Experimental setup with odour delivery device and electronic nose. Adapted from Ackels et al. Ackels2021. c, Electronic nose circuitry. d, Microscopy image of the MiCS-6814 NH3 sensor with its housing removed. e, Heater modulation cycle in ambient air. f, PID and flow meter traces for a 20Hz stimulus. Solid / faded (occluded) traces for mean / std. of five trials. g, Resulting olfactometer temporal fidelity, for various frequencies. Odourants abbreviations: IA: isoamyl acetate; EB: ethyl butyrate; Eu: cineol; 2H: 2-heptanone; blank: odourless control.
• Figure 2: Rapid heater modulation enables robust data features.a, Sensor resistance of four MOx sensors with 20 Hz hotplate temperature modulation, responding to a 1s odour pulse of isoamyl acetate (green background). b, 50ms data feature for different gases, selected between 500ms and 550ms after odour pulse onset. Raw sensor response (upper) and normalised sensor response (lower, see Methods for normalisation procedure). Time shifted by cycle phase $\rho$ w.r.t. odour onset, for visual guidance only. c, Principal component analysis (PCA), explained variance (most left) and projections, and d, t-distributed stochastic neighbour embedding (t-SNE) visualisation, for the set of normalised data features extracted between 500ms and 1000ms after odour onset. e, Accuracy scores for a k-nearest neighbours (k-NN) classifier trained on 50ms data features from 1000ms odour pulses at full concentration, and tested on 50ms features from 1000ms odour pulses at different concentration levels (tuned by adjusting the duty cycle of the micro-valves).
• Figure 3: Electronic nose can classify short odour pulses based on 50 ms data features.a, Feature labels for the training set were phase aligned in relation to odour on- and offset. Features that overlapped with transition periods were not considered for training ("rejected", see Methods for parameters). b, Odour stimulus classification over time for odour pulses of various lengths (10ms - 1000ms, as predicted by a RBF-kernel SVM classifier trained on 50ms features from 1000ms second odour pulses. Shown here are 1000ms pulses. For visual clarity only, the trials are sorted by odour, and within each odour are sorted by phase w.r.t. stimulus onset. c, Classification correctness over time (evaluated via the true odour presence), for different pulse durations. d, Test accuracy, onset time and offset time for the prediction over time described in b & c. Onset and offset were extracted using time-to-first non-'blank' and 'blank' prediction respectively, and shown here with respect to theoretical odour onset and offset.
• Figure 4: Decoding temporal structure of rapidly switching odours. a, Odour valve commands. b, PID response. c, Electronic nose response. d, Discrete Fourier Transformation (DFT) of first derivative of the sensor log-resistance. Crosses denote highest-magnitude peaks. DFT bin frequencies were rounded to nearest integers, for visual clarity only. e, Feature visualisation frequency, magnitude and phase of the dominant DFT peaks. Thick lines; means of corresponding trials, thinner lines; single trials. f, Class-balanced accuracies for modulation frequency classification. g, Accuracies for binary modulation frequency classification. h, Class-balanced accuracies for binary modulation mode classification (corr. vs. anti-corr.). i, Subset of g for IA-EB, for mouse performance comparison (described in detail in Ackels et al.Ackels2021). j, Subset of h for IA-EB, for mouse performance comparison. Panels a-e show representative trials only. For f-j, electronic nose accuracy mean and SD (clipped at 1.0) arise from repeated training and testing with different random seeds.
• Figure S1: Supplementary figure for experimental setup.a, Electronic nose design, displaying how the microcontroller unit (MCU) sets and reads out the sensor heaters in a closed loop, while reading out the analyte dependent sensor resistances. Further, the MCU connects to an environmental sensor (PHT) and a micro SD card. b, R-T curve of a 50ms temperature cycle between 150℃ and 400℃ without external stimulus, displaying how the sensor response closely follows the hotplate temperature. c, Different sensor hotplate settings over time. For each experiments, all the stimuli were presented in randomised order. d, Heatmap depicting the distribution of odour presentations over a set of 1 hour time intervals. A ${\chi}^2$ test was performed to assess the randomness of class distribution over time intervals, with the computed p-value indicated as '$p$'. e, PID response to a 1s isoamyl acetate pulse. Grey-dotted and red dotted lines denote mean of pre-stimulus baseline and 4 standard deviations threshold respectively. Where the response crosses the threshold upwards (downwards), the odour onset (offset) is registered. f, Extracted odour onsets (w.r.t. $t=0ms$) and offsets (w.r.t. $t=1000ms$) for 1000ms pulses of different odours. For all experiments, the odourants are abbreviated as follows. IA: isoamyl acetate; EB: ethyl butyrate; Eu: cineol; 2H: 2-heptanone; blank: odourless control.