Variable-condition fault diagnosis for building chiller based on deep feature extraction and discrepancy minimization

Data-driven fault detection and diagnosis (FDD) has the potential to be applied in building energy systems of smart city, which heavily relies on massive known data and consistent data distribution between training and application. However, due to the diversity of operating conditions (OCs) and faults, on-site chillers always generate data of various distributions, which fatally undermines the validity of the FDD model. This study, based on the innovative idea of deep feature extraction and discrepancy minimization (DFE-DM), proposes a novel chiller fault diagnosis strategy with extrapolation capability to accommodate variable OCs. Comprehensive validation was conducted on the seven types of chiller faults from ASHRAE RP-1043 project. It is demonstrated that compared with the traditional and common adaptable models, the proposed strategy exhibits an excellent FDD performance in terms of overall accuracy, individual F-measure, and diagnostic efficiency. Even when other methods are invalid, it still behaves quite well (F-measure up to 92.42 %), and the average accuracy on 30 tasks is improved by 45.91 % from random forest (RF), with the largest improvement reaching 81.18 %. The number of source OCs and the levels of fault severity are also investigated. The outcome may provide feasible guidelines for chiller online-FDD solutions adaptable to new OCs.