Comprehensive study on sensitive parameters for chiller fault diagnosis

For the fault diagnosis of a building chiller, the selection of suitable features/parameters is potentially more critical than the selection of diagnosis methods. This study explores the sensitive parameters for the seven typical faults in chillers by performing global sensitivity analysis (GSA) based on a Random Forest (RF) meta-model, and proposes a novel hybrid feature screening strategy of cascade feature cleaning and supplement (CFCS) based on correlation analysis and experience. Compared with the traditional experience-based selection, the proposed methods can enable an insight from all dimensions; hence reduce the number of sensors significantly while retaining as much useful information as possible. The results show that the fourteen (SPCS-2 set) or nine parameters (SPDM-set) screened out from the original 64 parameters have an excellent indication to the seven faults, with the diagnosis accuracy reaching 99.67% and 99.79%, respectively. The generalization performance and diagnostic reliability of the feature sets has also been verified under different diagnosis methods. Besides, it is demonstrated that the parameters of lubricating oil exhibit an immediate and salient indication to the system status, more significant for fault diagnosis than those of refrigerant. Therefore, sensors for oil parameters are recommended to be installed for an early recognition of faults. (C) 2021 Elsevier B.V. All rights reserved.