Date of Graduation
Master of Science in Computer Science
Determining sensor relationships in smart environments is complex due to the variety and volume of time series information they provide. Moreover, identifying sensor relationships to connect them with actuators is difficult for smart home users who may not have technical experience. Yet, gathering information on sensor relationships is a crucial intermediate step towards more advanced smart home applications such as advanced policy generation or automatic sensor configuration. Therefore, in this thesis, I propose a novel unsupervised learning approach, named SeReIn, to automatically group sensors by their inherent relationships solely using time series data for single resident smart homes. SeReIn extracts three features from smart home time series data - Frequent Next Event (FNE), Time Delta (TD), and Frequency (FQ). It then applies Spectral Clustering, K-Means clustering, and DBSCAN to group the related sensors. The application of unsupervised learning enables this approach to operate anywhere in the smart home domain regardless of the sensor types and deployment scenarios. SeReIn functions on both large deployments consisting of around 70 sensors and small deployments of only 10 sensors. Evaluation of SeReIn on real-world smart home datasets has shown that it can recognize inherent spatial relationships. Using three different unsupervised clustering evaluation metrics: Calinski-Harabasz Score, Silhouette Score, and Davies-Bouldin Score, I ensure that SeReIn successfully builds clusters based on sensor relationships.
IoT, relationship inference, sensor, smart home, time series, unsupervised learnin
Graphics and Human Computer Interfaces | Other Computer Sciences
© Samuel Nack
Nack, Samuel, "Sensor Relationship Inference in Single Resident Smart Homes Using Time Series" (2023). MSU Graduate Theses. 3931.