AbstractIn oxygen concentration sensing based on the physical phenomenon of luminescence, temperature plays a fundamental role since the quality of sensory information of the luminescence signal is influenced dramatically by the temperature. It is crucial to measure the sensing material’s temperature accurately, with positionally precise and fast temperature measurements, especially in very small-scaled sensors. In comparison with the conventional measurement methods such as mathematical models and look-up tables, the problem of the measured oxygen concentration has been addressed with novel hardware design and learning principles, resulting in three contributions as follows in optical oxygen and temperature sensing with machine learning algorithms.
Firstly, this Thesis presents a new sensing hardware design that is able not only to perform measurements, but also to autonomously collect a very large dataset that can be used to train a machine learning model that can then be used for inference. This makes the necessity of having additional hardware for data collection obsolete.
Secondly, this Thesis shows how to extract oxygen concentration from the luminescence measurements with single-task learning neural network architectures at a fixed temperature. Overfitting and hyper-parameter tuning have been studied extensively in this case, as they present themselves in a different form than in other situations.
Finally, thanks to multi-task learning architectures, this work demonstrates how to simultaneously extract oxygen concentration and temperature from the luminescence measurements. This approach, working for a large range of temperatures and oxygen concentrations, reaches an accuracy comparable to that of commercial sensors. Inference in this case is computationally very cheap and almost instantaneous, making this type of fast and accurate sensing extremely attractive for the development of oxygen sensors.
This approach described in this Thesis has the potential to change the way in which oxygen sensors are designed and built, making them simpler to build and more reliable in demanding sensing conditions. Future research directions are in the areas of transfer learning, to be able to reduce the data necessary for training models for new sensing scenarios, and explainability to optimise the data collection process and the models.
|Date of Award||May 2021|
|Supervisor||Honghai Liu (Supervisor)|