Sensor calibration and diagnostics under parameter uncertainty: A smart building application

One key problem with control systems in general, and control systems for buildings in particular, is that mis-calibration of sensors/actuators is commonplace and causes significant problems, such as suboptimal performance and diagnostics false alarms. This paper describes a methodology for calibrating sensors that can reduce these problems. We show how we can take sensor outputs and continuously calibrate them by applying expectation-maximization (EM) learning and recent gossip-based algorithms. We apply our approach to the domain of sustainable buildings, in particular temperature sensors in shared zones in a large commercial building. We empirically show that our approach can correctly either diagnose faults that render sensors impossible to calibrate, or can perform appropriate calibration.