Statistical Detection of Alarm Conditions in Building Automation Systems

A method for the automatic detection of abnormal behavior in a building automation system is compared to a standard system for problem detection. The automated method is based on statistical models of sensor behavior. A model of normal behavior is automatically constructed. Model parameters are optimized using an on-line maximum-likelihood algorithm. Incoming sensor values are then compared to the model, and an alarm is generated when the sensor value has a low probability under the model. The alarms generated by the automated system are compared to alarms generated by pre-defined rules in a standard automation system. The performance, strengths and weaknesses of the automated detection system are discussed.