Towards Faster Machine Olfaction: Advances In High-Speed Sensing And Processing

This thesis investigates the principles and methods for achieving high temporal resolution in odour sampling and processing for machine perception. Initially, a comprehensive review of fast olfaction mechanisms in insects and mammals is provided, highlighting the necessity of rapid sensing when exposed to the dynamics of turbulent odour plumes. The discussion extends to artificial olfaction technologies, emphasising current advancements in electronic nose (e-nose) systems and their applications. Initial evaluations of existing datasets and algorithms revealed significant limitations in a widely used gas sensor dataset, particularly due to a non-randomised measurement protocol and severe sensor drift. These issues rendered the dataset unusable for classification benchmarks. Multiple studies that are impacted by this were identified, where the example of a prominent neuromorphic few-shot odour-learning algorithm study was investigated further. In response, a set of best practices for future gas sensor data collection campaigns was established to ensure data reliability and reproducibility. Several data collection campaigns were conducted using a custom-built e-nose system, which was based on MOx gas sensors and fast peripheral electronic devices. A novel approach to data feature acquisition for odour classification was proposed, involving rapid temperature cycling of the gas sensors. This method enabled the recording of two datasets capturing diverse indoor and outdoor olfactory scenes, which were effectively distinguished using the acquired features. An extensive laboratory campaign followed, in which the e-nose system was evaluated against a benchmark previously used to explore the temporal odour discrimination capabilities of mammals. The results demonstrated that the e-nose could distinguish correlated odour pulse trains from anti-correlated ones at modulation frequencies up to 40 Hz and determine frequencies up to 60 Hz, surpassing mammalian capabilities. Additionally, the system achieved odour classification at pulse widths as short as 10 milliseconds when employing 50 millisecond duty cycles for sensor temperature modulation, setting a new precedent in artificial olfaction. Further, for efficient processing of olfactory signals, neuromorphic computing principles were explored. The potential advantages of asynchronous sampling and data processing methods were discussed in the context of the physical characteristics of turbulent odour plumes. Various event generation and processing algorithms were critically reviewed, and discussed in the context of olfactory signals. An example study is provided, in which asynchronous event sampling is applied to heater-cycled MOx sensor e-nose data, and the effectiveness of different event encoding schemes was assessed. Finally, the thesis discusses the results by putting them in perspective with future research directions.