Indoor Localisation Using Ion Mobility Spectrometry
Hietamäki, Leevi (2025)
2025
Sähkötekniikan DI-ohjelma - Master's Programme in Electrical Engineering
Informaatioteknologian ja viestinnän tiedekunta - Faculty of Information Technology and Communication Sciences
Hyväksymispäivämäärä
2025-08-20
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:tuni-202507317968
https://urn.fi/URN:NBN:fi:tuni-202507317968
Tiivistelmä
This thesis investigates the use of machine learning methods for indoor location classifcation based on scent profles of different locations, measured using Ion Mobility Spectrometry (IMS). The IMS measurements were performed using an electric nose (eNose) device called ChemPro100i. The thesis’s purpose is to assess the efciency of several classifcation algorithms and their capacity to generalize to real-world scenarios. Four primary methods, K-Nearest Neighbours (KNN), Random Forest (RF), Gradient Boosting Decision Tree (GBDT), and Convolutional Neural Networks (CNN), were tested on a dataset collected from various measurement locations mostly within university campus. While this thesis uses existing implementations for KNN, RF and GBDT algorithms, it introduces a novel CNN model specifcally designed for captured IMS measurements.
The results show that RF achieved the highest overall accuracy. While the proposed CNN model demonstrated strong performance in certain runs, its high variance led to a lower average accuracy, making its overall performance less consistent and reliable. In addition to the baseline study, data augmentation approaches were used to increase the training set and improve generalization, especially for our CNN model.
Confusion matrices were analyzed to better understand classifcation results across different locations. The fndings indicate that, while machine learning models can learn to classify IMS-based location data to some extent, external factors present in real-world environments signifcantly disrupt IMS measurements, reducing the classifcation methods ability to effectively distinguish between measurement locations. Future research should concentrate on combining multi-sensor data, enhancing model robustness, and collecting a larger data set before these methods can be reliably implemented in practical applications.
The results show that RF achieved the highest overall accuracy. While the proposed CNN model demonstrated strong performance in certain runs, its high variance led to a lower average accuracy, making its overall performance less consistent and reliable. In addition to the baseline study, data augmentation approaches were used to increase the training set and improve generalization, especially for our CNN model.
Confusion matrices were analyzed to better understand classifcation results across different locations. The fndings indicate that, while machine learning models can learn to classify IMS-based location data to some extent, external factors present in real-world environments signifcantly disrupt IMS measurements, reducing the classifcation methods ability to effectively distinguish between measurement locations. Future research should concentrate on combining multi-sensor data, enhancing model robustness, and collecting a larger data set before these methods can be reliably implemented in practical applications.