An unsupervised anomaly detection framework for smart assisted living via growing neural gas networks

Abstract

In this study, we present a novel framework for detecting anomalies in everyday activities within a smart-home environment. Our method utilizes the growing neural gas (GNG) concept to dynamically adapt to the changing behaviors of monitored individuals, eliminating the need for supervised input. To develop and evaluate our framework, we collected real-life data from environmental sensors that tracked the daily activities of 17 elderly subjects over a continuous two-year period. The proposed approach is highly versatile, capable of detecting a wide range of anomalies associated with daily living activities. We focus on activities that exhibit abnormal duration, frequency, or entirely new behaviors that deviate from established routines. The performance evaluation of our framework revolves around two key aspects: reliability and adaptability. Reliability measures the accuracy of detecting unusual events, while adaptability assesses the system’s ability to accommodate changes in user behavior. This involves recognizing recurrent anomalous behaviors as new norms over time and transitioning from persistent anomalies during an initial phase. Our proposed anomaly detection system demonstrates promising results in real-life scenarios. It achieves good reliability, with true negative rate and true positive rate exceeding 90% and 80% respectively, across all activities and users. Additionally, the system swiftly adapts to new individuals or their evolving behaviors, adjusting within a span of 3 to 7 days for new behaviors.