Dimensionality expansion for motif recognition in time series dataBertalanič, Blaž (Avtor) Meža, Marko (Mentor) Fortuna, Carolina (Komentor) time series classificationdimensionality expansiondeep learningtime series imaginggraph neural networkstime series to graph transformationwireless anomaly detectionnon-intrusive load monitoringappliance classificationanomaly classification.This thesis explores the application and optimization of dimensionality expansion techniques for time series classification, focusing on two critical smart infrastructure scenarios: wireless link layer anomaly detection and appliance recognition in smart grids via non-intrusive load monitoring. Initial investigations reveal that dimensionality expansion methods substantially enhance classification accuracy. However, a significant challenge arises with scalability, as these methods typically increase computational complexity quadratically with the length of the time series. To overcome this, we developed a novel approach that not only ensures linear scalability but also surpasses traditional imaging techniques in classification accuracy. Further, recognising the need for rapid response in time-sensitive applications, we aimed to deploy our models on constrained edge devices. We achieved this through the integration of Graph Neural Networks with Visibility Graph transformations. The Visibility Graph transformation effectively converts time series data into a graph data structure, adeptly capturing complex patterns and temporal dependencies that are less discernible with conventional imaging methods. Our results demonstrate that combining Graph Neural Networks with Visibility Graph not only enhances classification accuracy but also reduces computational demand, making these models suitable for deployment on edge devices. This thesis confirms the potential of advanced dimensionality expansion techniques as a powerful tool for improving the accuracy and efficiency of time series data classification in smart infrastructure environments.20242024-09-11 11:55:01Doktorsko delo/naloga161466sl