Model Identification with Application to Building Control and Fault Detection

Motivated by the high speed of real-time data acquisition, computational power, and low cost of generic PCs and embedded-PCs running Linux, this thesis addresses new methods and approaches to fault detection, model identification, and control. Fault detection. A series of faults was introduced into a 3-Ton roof-top airconditioning unit (RTU). Supply and condenser fan imbalance were detected by changes in amplitude spectrum of real power resulting from the interaction of impeller rotation and the dominant chassis vibration mode. Ingestion of liquid refrigerant by the compressor was identified by detecting power and reactive power transients during compressor starts. An adaptive ARX(5) model was used to detect ingestion during steady compressor operation. Compressor valve or seal leakage were detected by a change in the leakage parameter of a simple evaporator-compressor-condenser model that explains the rise in compressor load from 0.25 to .5 seconds after compressor start, i.e. as shaft speed rises from about 50% to 90% of synchronous speed. Refrigerant undercharge was also detected by changes in start transient shape. Overcharge was detected by steady state compressor power and reduced evaporator and condenser air flow were detected by steady state power draw of the respective fan motors. Model Identification. On-line models are useful for control as well as fault detection. Model-based control of building loads requires a valid plant model and identification of such a model for a specific building or zone is a non-trivial inverse problem. The thesis develops three advances in the thermal diffusion inverse problem. Two involve thermodynamic constraints. The problem is first reformu-lated in such a way that the constraint on temperature coefficients becomes implicit and the problem may still be solved as an unconstrained linear least squares problem. To enforce the constraint on system eigenvalues, the problem is formulated as an unconstrained mixed (linear and non-linear) least-squares problem, which is easier to solve than the corresponding problem with linear objective function and non-linear constraints. The (usually unfounded) assumption on which the normal equations are based-that observations of the independent variables are error free-is relaxed at the cost of one more non-linear term. The resulting model coefficients are valid for predicting heat rate given zone temperature as well as for predicting zone temperature given heat rate. Control. Three important control applications involving transient zone thermal response are HVAC curtailment, optimal start, and night precooling. A general framework for model-based control of zone and whole-building operation is developed. Optimal …