Study of Autoencoder Neural Networks for Anomaly Detection in Connected Buildings

Nowadays, buildings are equipped with safety devices, to prevent and identify critical events, such as fires using smoke detection devices or intrusions using alarm systems. The expansion of the Internet of Things (IoT) involves new perspectives, especially in terms of data variety and quantity from which critical event detection can benefit. These new perspectives have made it feasible to develop data-driven methods to detect problematic situations that may occur in connected buildings using anomaly detection without necessarily setting up a specific system for each of them. Recent research works in deep learning have shown that autoencoder neural networks have a great capability in analyzing high-dimensional data and can be used for the purpose of anomaly detection in a supervised or unsupervised way, as anomaly labeled data are typically not available. In this paper, we investigate and compare the capability of different kinds of autoencoder neural networks in the detection of anomalies in a large scale Smart Home dataset and we propose a method for evaluating partially labeled anomaly detection models.