Hybrid method for automatic fault detection and diagnosis in buildings with limited monitoring data

• A novel methodology for Automatic Fault Detection and Diagnosis under limited data is developed • Dynamic and functional variables are fully integrated into the inference as random variables • The methodology is able to identify most faults efficiently, even simultaneous ones • A novel encoder-decoder based emulator is proposed for calibration of dynamic computer models Early Automatic Fault Detection and Diagnosis (AFDD) is essential in energy use and carbon emission reduction to meet the Paris Agreement goals. However, AFDD application scope is usually narrowed down to buildings equipped with costly hard-to-maintain Advanced Metering Infrastructure (AMI). The methodology developed in this paper relies on commonly metered data and numerical simulations of building energy performance to be applicable to most buildings. Besides, while other works in the field leveraged smart meters to measure dynamic solicitations about the building, e.g. building occupancy, a different way of dealing with these random variables is studied by treating them as unknown parameters in the calibration process, further decreasing the dependence to fine-grain metered data. The resulting methodology is promising, as it is able to identify most symptoms and orient an operator towards most likely faults with a reduced data set, and without specifying the faults beforehand. Operationally, this methodology could gain in robustness and efficiency if coupled with simple rule-based AFDD. Furthermore, additional metered data points may prove critical to greatly improve the AFDD efficiency in a cost-effective way.