Fault diagnosis using fused reference model and Bayesian network for building energy systems

Abstract Fault diagnosis is crucial to address the growing fault issue, with benefits of reducing energy use and improving the energy efficiency for building energy systems. However, the poor performance and the limitations of the individual method used separately are the key issues for the field implementation of fault diagnosis method. The aims of this research are to obtain a better performance than the individual method and to overcome such the limitations. To achieve the two aims, a hybrid method is proposed to fuse the reference model-based approach, driven by non-faulty (normal) data, and Bayesian network-based approach, driven by faulty data, into a single framework to make them work together, thus making their advantages complementary to each other, and from the view of methodology, a generic framework about how to develop the hybrid method is given. The fault diagnosis performance is evaluated by using the experimental data. Test results show that the proposed hybrid method can still diagnose faults with quite high accuracies when the adequate faulty data or the accurate reference models are unavailable in the fields. At that moment, the reference model-based or the faulty data-driven method will not work. Test results also show that the accuracies are significantly improved by 45.7% at most (for condenser fouling) when the reference model-based and faulty data-driven parts work together. Obviously, a novel hybrid method is proposed and proven to be effective for fault diagnosis for building energy systems, and the limitations of individual method used separately are overcome effectively, and a better fault diagnosis performance is obtained finally.