Exploratory data analysis and modelling of building sensor networks

Abstract

Modern building energy management systems generate large volumes of sensor data, and the reliable operation of these systems depends on detecting anomalies, such as sensor faults and abnormal behaviour. In this thesis, anomaly detection is investigated in the sensor network of a single public building, instrumented with 393 active sensors recorded at 15-minute intervals. The main research question is whether learning the sensor graph from data improves anomaly detection compared to using a fixed, domain-derived graph. A Graph Convolutional Network autoencoder is developed for implementing reconstruction-based anomaly detection. The sensor graph is made fully learnable through a parameterization that preserves edge-weight non-negativity and sparsity without explicit regularization. Six models are compared under a shared architecture and evaluation protocol, ranging from a graph-free baseline to fixed graphs initialized from building structure, correlation, and their learnable version, along with graphs initialized from random noise, and are evaluated against three synthetic anomaly types representing sensor noise, sparse interference, and sustained failure. The results show that reconstruction accuracy alone does not directly correlate to detection capability, as the most accurate baseline model detects no anomalies, and that the learnable graph consistently outperforms the fixed one. The learnable graph initialized from the building structure is the overall best performer. It is concluded that the treatment of the sensor graph matters more than the raw reconstruction accuracy for reconstruction-based monitoring of building sensor networks.