Using Machine Learning for Optical Spectroscopy Data Analysis

Living a healthy lifestyle is an ever-increasing priority. To facilitate such a healthy lifestyle, accurate, quick, and inexpensive feedback on diet quality is essential. Sensors based on multiple spatially resolved reflection spectroscopy aim to provide such feedback. However, current data processing algorithms require highly accurate hardware. This requirement for accuracy causes production costs of the sensors to be too expensive, while the application scope is too small to be viable for end-customers. In order to keep production costs low, new algorithms capable of handling production inaccuracies need to be developed. This thesis proposes such a novel neural network architecture called a continuous feature network. In addition to being wellsuited for the sensor data at hand, continuous feature networks are capable of compensating for sensor inaccuracies. A continuous feature network is also capable of predicting results from an input sample with partially missing data, allowing it to ignore certain production defects. In this thesis, continuous feature networks are proposed, implemented, trained, and investigated using real-world sensor data. To improve training, a novel method for semi-supervised learning based on the available datasets is introduced and evaluated. Based on the ability of the continuous feature network to operate on partially missing data, a novel explainable AI method is introduced, allowing to accurately quantify possible error sources for a measurement. The newly introduced methods are applied to the processing of sensor data, relaxing the requirement for highly accurate sensor hardware while increasing prediction accuracy. This enables a significant reduction in production rejects and thus sensor cost, while also allowing for the detection and prediction of new vitality parameters.