FDD, Fault Impact, Data-driven Modeling
Like any electromechanical system, direct-expansion (DX) air conditioners and heat pumps often develop faults over time that contribute to reduced operating efficiency, more frequent comfort violations, or even premature failure. Automated fault detection and diagnosis (AFDD) methods have been developed for these systems and much experimental effort has been undertaken for their evaluation. In order to reduce development costs required for AFDD technologies, additional research related to modeling DX equipment subject to faults has been undertaken. Investigation of AFDD methods in a virtual environment typically requires relatively detailed equipment models based in some part on thermodynamic principles. Because of these embedded constraints, simulation of faulty equipment operating performance can be time consuming and computationally intensive. In this work, meta-models based on previously developed greybox fault impact models for DX equipment have been developed using artificial neural networks. After tuning these neural network meta-models for different equipment, AFDD performance and fault impacts were simulated using a simple building load model. Significant computational speedups were realized over the original greybox equipment models without loss of significant accuracy. Ultimately through careful meta-model training, it is believed that using neural networks to approximate detailed, computationally-intensive equipment or building models may be useful in applications that require frequent model evaluations.