Evaluating Fault Detection and Diagnostics Tools with Simulations of Multiple Vapor Compression Systems

A methodology for evaluating the performance of fault detection and diagnostics (FDD) tools applied to unitary air conditioners has been developed by Yuill and Braun (2013). Data from faulted and unfaulted systems operating over a range of driving conditions are fed to the FDD tools, and the FDD responses are compared to the known operating conditions. The methodology originally relied upon experimental measurement data, but the amount of available data is limited, and evaluations can be far more meaningful if the operating conditions of the inputs can be controlled. Furthermore, a finite input set can be learned by an FDD algorithm, and the evaluation can be thereby gamed. To solve these problems, a large library of data from multiple systems under a wide range of conditions, with and without faults of varying magnitude, was generated with simulations from a novel gray-box modeling approach (Cheung and Braun 2013a, 2013b). The simulation outputs are being used to train neural network models, which can be coupled to software that executes the evaluation method. The neural network models are simpler than the semi-empirical approach, so they can produce evaluation inputs very quickly. This will facilitate the evaluator generating semi-random conditions to provide a unique set of evaluation data that are sufficiently accurate and numerous to provide repeatable results. Some evaluation results from three FDD protocols are used to demonstrate the advantages of simulation data over measurement data.