With the demand for more information on the building’s indoor climate, a massive amount of multi-sensors are mounted in buildings. Sensor anomalies in smart buildings lead to higher energy consumption and a less comfortable indoor climate. In the state of practice, rule-based approaches were proposed to detect and diagnose sensor anomalies in the building sensor network. However, as the number of sensors is growing and the types of sensors are becoming more diverse, rule-based approaches become more and more limiting and expensive.

This project proposed a data-driven detection and diagnosis framework that transfer the knowledge from one sensor to the other sensors in the network for detecting and diagnosing sensor anomalies in the smart building. In the proposed framework, all the data from one room that contains targeted anomalies is chosen to be the training set. Principal component analysis (PCA) is used to extract the features, and support vector machine (SVM) is used to model the classifier. This framework is tested specifically on all the CO2 sensors in a smart building. Functions of detecting a single anomaly, detecting mixed anomaly and diagnosis anomaly are evaluated and compared with the state of practice. Moreover, the influence of anomaly rate on the performance of the proposed method is investigated. In order to test the sensitivity to the training dataset, a final experiment was performed where the room that provide the training data was changed to a different room.