Anomaly detection using autoencoders for Ambient Assisted Living

Abstract

One of the main goals of the Ambient Assisted Living field of research is to extend the time of independent living for elderly people through the use of sensors and computational systems. This thesis focuses on building a system which models the power consumption of appliances to build a notion of normality concerning user habits. Once the system learns this notion, it can detect major divergences from it. It is important to detect these major divergences because they can indicate issues in the life of the user. The main research question examines whether autoencoder neural networks are well-suited to detect anomalies in appliance-level power data. For this examination, several autoencoder models with different sets of hyperparameters were trained on appliance-level power data coming from a real-world dataset. Then, the trained models were tested on artificially generated anomalies to evaluate their sensitivity to such data. Autoencoders were shown to per- form well with most kinds of anomalies which were deemed relevant in this context. Overall, this thesis analysed important aspects of autoencoders for anomaly detection. Through this analysis, several strengths of the models were found, as well as weaknesses. Moreover, improvements to the current models were proposed to overcome the limitations which emerged from the analysis.