A neuro-computational approach to chiller fault identification and isolation

Abstract

This thesis develops a steady state model of a reciprocating, single-stage, vapor compression chiller, based on thermodynamic principles and the ASHRAE HVAC-1 primary toolkit model, to predict and diagnose the behavior of the chiller. The characteristics of the chiller are identified by measuring certain parameters. Once the parameters are identified, the system boundaries are established, with certain parameters tainted by faults being isolated from the input parameters and, in turn, being derived from them using a nonlinear equation solver. A model-based approach is followed for chiller fault detection and diagnosis. The model generates residues by comparing the output of the chiller model with that of the same model albeit with some faults embedded. A nonzero residue indicates the presence of a fault. Various neuro-based classification techniques such as Adaptive Neuro-Fuzzy Inference Systems (grid partition and subtractive clustering) and Artificial Neural Networks (ANN) are evaluated, culminating in fault identification and isolation by an ANN system, batch trained with a fault matrix using the Levenberg-Marquardt learning algorithm, on a feed-forward back propagation network. Finally, additive sensor noise is introduced in select parameters and its effect on the overall accuracy of the model is tabulated.