Neuromorphic olfaction: event-based gas sensing systems

Abstract

Gas detection is important in different domains like gas leakage monitoring, pollution control, disease diagnosis, and so on. To prevent health hazards, it is equally important to detect the gas concentration level rapidly. Out of different types of gas sensors, Metal oxide (MOx) sensors are extensively used for gas detection because of their portability, low cost, and high sensitivity for specific gases. However, how to efficiently sample and process the sensor data remains an open question. In a conventional electronic nose (e-nose) system with an array of MOx sensors of different sensitivity profiles for target gas detection, analog-to-digital converters (ADCs) of fixed sampling frequency are used to convert the sensor data into samples which are then fed to computers or microprocessors for further processing. Samples are generated at uniform intervals irrespective of the presence or absence of gas in the environment whenever the e-nose system is turned on which could lead to data explosion at the processing stage for bigger e-nose systems with many sensors. Our nervous system also samples and processes sensory information from the external world. However, the external stimulus signal is not sampled and transmitted uniformly but only when there is some significant variation in this signal to be communicated further. Therefore, only relevant samples are processed by the nervous system, accelerating our decision-making process. Also, the important stimulus features are encoded in the samples such that further processing is greatly simplified. Therefore, developing the event generation and feature encoding mechanisms inspired by those found in biological neural systems for conventional gas sensing systems could be one of the possible solutions to reduce the computational load on processing stages and accelerate the inference process. This thesis introduces an analog front-end for one MOx sensor which converts the MOx data into two discrete pulses generated from two separate pathways. The output of this frontend shares similarities with the spiking output of a mammalian olfactory bulb and one of the event-based vision sensors. The gas concentration is encoded in the time difference between these pulses similar to the odor concentration encoding in the output spikes emerging out of two distinct output channels of the olfactory bulb. We show that for a gas pulse injected in a constant airflow, the time difference between pulses decreases with increasing gas concentration, similar to the spike time difference between two output neurons in the olfactory bulb. The circuit design was further extended to the MOx sensor array and this sensor array front-end was tested in the same environment for gas identification and concentration estimation. The next part of the thesis explores MOx data recorded for artificial gas plumes deployed in a constant airflow and different features were analyzed for gas concentration encoding. The feature most robust to noise was chosen and the sensor array front-end circuit design proposed earlier was modified to encode this feature into analog spikes for gas concentration estimation. This new design was also tested for the single gas pulse environment and produced promising results.

Date of Award	2025
Original language	English
Awarding Institution	Western Sydney University
Supervisor	Andre van Schaik (Supervisor), Michael Schmuker (Supervisor) & Gregory Cohen (Supervisor)