User-friendly fault detection method for building chilled water flowmeters: Field data validation

Abstract

In building HVAC systems, a chilled water flowmeter is an important sensor. Its reading can be used to measure the real-time cooling load, which is a critical variable for the automated control of building HVAC systems, especially central chiller plants. To maintain accurate reading data, fault detection and diagnosis (FDD) of flowmeters are necessary. The available FD/FDD methods for chilled water flowmeters have several common disadvantages: (1) high requirements on sensor integrity: multiple sensors are usually involved in the construction of energy balance models; (2) complex methodology: the fault of any monitored sensor could trigger a detection hit, and a diagnosis procedure must be used to isolate the faulty sensor; and (3) the more sensors that are involved, the more difficult it is to collect fault-free historical data for the establishment of a fault-free benchmark. To address these problems, a user-friendly fault detection (FD) method for building chilled water flowmeters is proposed in this study. The proposed method requires three types of variables to function: the pump frequency, the pump power, and the measured chilled water flow rate on the header pipe. The real-time pump frequency and electrical power data that are collected by VFDs/power meters are used to estimate the true value of the chilled water flow rate with random forest regressor. An auxiliary alarm rule is proposed to tackle the tradeoff issue between detector sensitivity and false alarm rate. Field data of a real HVAC system are used in a case study to evaluate the performance of the proposed method in comparison with three existing FD methods. The results of the validation case study suggest that compared to existing FD methods, which are based on the physical pump model, ensemble learning and clustering techniques, the proposed method could realize higher hit rates and higher alarm rates when confronting different faults (bias, noise and drift) at various levels. Besides, due to the proposed alarm rule, false alarms caused by occasional data errors are completely evaded. Sensitivity analysis to the proposed method indicates that the random forest with default hyperparameters is good enough for the FD task. In brief, the good features of the proposed method (i.e., the simple workflow, low sensor requirements, accessible robust input variable data, handy estimation model, auxiliary alarm rule with tight detection threshold) render the proposed method feasible and user-friendly for engineering practice.