A Deep Recurrent Neural Network-Based Method for Automated Building System Fault Diagnosis

Faults in building system operations typically result in building functionality degradations and considerable energy wastes. Conventional approaches mainly rely on domain expertise and engineering experiences for decision makings, which are neither efficient and effective considering the great varieties in individual building system configurations and operating conditions. The wide existence of building operational data has provided ideal platform to develop data-driven approaches for building system fault diagnosis. Such approaches are capable of conducting accurate, automated and in-time controls over building systems and therefore, has drawn increasing attentions from both academia and building professionals. Existing studies mainly treated building operational data as cross-sectional data for developing fault diagnosis methods, while ignoring the temporal dependencies among building variables. To enhance the fault diagnosis performance, it is essential to explore the intrinsic temporal relationships in building operational data. Therefore, this study proposes a deep recurrent neural network-based methodology for building system fault diagnosis. The methodology has been validated using experimental data on building chiller systems. The results indicate that deep recurrent models can achieve an accuracy of at least 95% for seven typical faults in chiller systems. The research outcomes are helpful for enriching analytic tools for building system fault diagnosis.