Details

Autor(en) / Beteiligte

• Alves, Rita Alexandra Nunes

Titel

Time Series Analysis and Machine Learning Algorithms for an E-Nose Sensor

Ort / Verlag

ProQuest Dissertations & Theses

Erscheinungsjahr

2021

Quelle

ProQuest Dissertations & Theses A&I

Beschreibungen/Notizen

• Studies revealed that there are volatile biomarkers associated with diseases, which can be found in patients’ body fluids, constituting possible sources of information for disease diagnostics. Electronic noses mimic the biological olfaction process by combining arrays of gas sensors’ dedicated signal processing and pattern recognition algorithms. Due to their portability, fast response and non-invasive operation, these devices have gained importance as a potential non-invasive diagnostics tool.This work aims to develop data processing and pattern recognition algorithms for optical and electrical signals acquired with e-noses under development at the Biomolecular Engineering Group. Three time series analysis techniques were developed based on: features extracted from the signals; the euclidean distance between signals; and a novel method based on the morphology of the generated signals. Samples of known Volatile Organic Compounds (VOCs) were used to test the methods and the one based on features extraction was applied to urine samples to separate healthy individuals from cancer patients. Signals generated by optical and electrical e-noses were used to assess the performance of each type of sensor and device in differentiating the VOCs. The results demonstrate that the optical e-nose is the most accurate device together with the features extraction method. The reproducibility of the signals over multiple experiments was also studied. In this case, the average accuracy score is 56.6% (between 40.3% and 94.0%) for the distance approach. The information of multiple e-nose sensors was combined which led to a higher classification performance than that of the individual sensors to particular VOCs. The signal analysis of urine samples produced random classifications. Further research is needed to optimize the classification model and the signal acquisition, as more variability of signal morphologies is required.With this work, the performance of the sensors and devices under development was fully characterized. The developed signal processing and machine learning tools, together with their application in a practical use case, set the basis for further research and the potential development of fast and non-invasive diagnostics tools.