Fouling of surfaces within the heat exchangers of heating, ventilating, and air conditioning (HVAC) systems of buildings is an equipment fault that wastes appreciable amounts of energy, however, it escapes detection under current building automation technology. A novel concept is introduced to automatically detect this fouling on waterside and airside surfaces of watercooled HVAC air coils. A machinelearning algorithm (the agent) embedded in microcontroller hardware receives and stores realtime data sampled from coil instrumentation. The agent maintains an adaptive predictive datadriven model of the coil process, assimilating all data sampled from the instruments whether the coil is in thermal equilibrium or nonequilibrium. A specific system identification transient (query) is exercised virtually on the model as a surrogate for exercising actual system identification transients on the real coil. During system operation, the agent periodically queries its model offline from the stream of sampled data, and the output response from the model is recorded. A difference between the current response of the model and a null baseline (i.e., clean coil) response archived earlier signals the coil thermal process has changed since the baseline query. Human expert knowledge derived from numerical simulations of clean and fouled coils is tabulated and embedded with the agent. The tabulated expert information enables the agent to discern fouling from other changes such as instrument drift, and to distinguish between airside and waterside fouling, estimate the severity of fouling, and estimate the uncertainty of its classification. The tabulated values essentially describe threedimensional surfaces over a plane defined by coil input quantities. The contours of these surfaces characterize the varied impact fouling has on coil thermal effectiveness, as considered over the state space of coil operation. This varied impact is explained by analysis in the effectivenessNTU state plane. A companion paper (Veronica 2010) presents results of employing one form of adaptive predictive model, a multilayer perceptron, to exercise the concept introduced here on simulated data.
Citation: International Journal of Heating, Ventilating, Air-conditioning and Refrigerating Research
Pub Type: Journals
fault detection FDD automation cooling coil HVAC heat exchanger fouling