Fusing system configuration information for building cooling plant Fault Detection and severity level identification

Abstract

Building Fault Detection and Diagnosis (FDD) technology is indispensable for energy saving and performance improvement of built environment regulation. Recent data-driven methods have shown the advantage in dealing with complex systems with random inputs. However, existing works on data-driven FDD merely considers the task as a simple classification problem to identify fault types. Prior knowledge on system configuration and fault severity levels have long been ignored. This paper proposes a novel data-driven strategy that adopts hierarchical labeling to fuse system structure information and describes the severity levels in a unified learning framework. A Large Margin Information Fusion (LMIF) method is derived and an on-line learning algorithm for streaming data is developed. Following the ASHRAE Research Project 1043 (RP-1043), the proposed strategy is applied to the FDD of a 90-ton centrifugal water-cooled chiller. Experimental results show that LMIF can greatly improve the FDD performance as well as recognize the fault severity levels with high accuracy, justifying the benefit of fusing prior knowledge of fault dependence information into the learning method.