Towards Human-Centric Smart Homes: Modeling Sensor-Actuator Interactions with Deep Learning

Recent advancements in the Internet of Things (IoT) have enabled smart home environments to automate daily activities with increasing sophistication. However, existing automation systems rely on predefined rules or static policies, requiring frequent human intervention and failing to adapt to statistical variations in user behavior. In this paper, we propose a novel data-driven approach for learning sensor-actuator activation and deactivation patterns in smart home environments, independent of specific human activities. Using deep learning models, our method extracts information on sensor-actuator interactions, which can be leveraged to generate adaptive operational policies for automation. We evaluate Long Short-Term Memory (LSTM) networks and Transformer models to analyze time-series sensor-actuator interactions, assessing their ability to dynamically anticipate and respond to user needs. Unlike traditional rule-based automation, our approach continuously adapts to variations in user activity sequences, enabling situational predictions that adjust dynamically to evolving behaviors. We validate our method across multiple smart environment testbeds using real human activity data, demonstrating the potential of deep learning to enhance autonomous and adaptive smart home automation while minimizing user intervention.