Quantitative fault detection and diagnosis methods for vapour compression chillers: Exploring the potential for field-implementation

Abstract

Chiller fault detection and diagnosis can optimize building energy and maintenance costs. Previous reviews have largely overlooked the detailed assessment of chiller fault detection methods; this study focuses on chiller methods proposed since 2004. Categorized into regression-based, classification-based, and knowledge-based approaches, the study delves into their procedural aspects and practical implications for field-installed chillers such as availability of the required training parameters and information, accuracy versus energy performance fault impact, and complexity of required data. Classification-based methods are mostly supervised learning, and few have been validated with faulty chiller data from real-world installations. Over 90% of classification-based, over 70% of regression-based, and all knowledge-based surveyed methods were tested using experimental data only. Some measured parameters like the subcooling temperature, oil feed pressure, and oil sump temperature that are commonly used in model training for detection and diagnosis algorithms are rarely measured in field installations. Despite significant research efforts to enhance the early detection of refrigerant leakage and condenser fouling faults, studies indicate minimal impact of these faults at low severities. Justifying their early detection may primarily rely on environmental considerations. To bolster field implementation, incorporating common factory-installed sensor parameters in new method development or testing of existing ones is recommended. Means that can provide faulty data for classification-based methods should also be devised. Hybrid methods that incorporate experts’ knowledge in the detection and diagnosis algorithms are encouraged and more testing of existing methods with real-world installations to ensure applicability is recommended.