A timeseries supervised learning framework for fault prediction in chiller systems

Abstract

Chiller systems play a critical role in large-scale commercial buildings, providing air-conditioning to occupied spaces and possibly also providing cooling to other process equipment. When faults occur in these systems, significant energy can be wasted, and costly maintenance may be necessary to restore the equipment to working conditions. Meanwhile, occupant comfort or operational efficiency may be sacrificed due to the lost cooling capacity. To help avoid these issues, various fault detection and diagnosis (FDD) approaches have been proposed and developed to automatically detect and diagnose faulty conditions in chiller systems. Unfortunately, merely detecting faults may not be sufficient to fully mitigate the ill effects. If instead faults can be predicted ahead of time, then operators may be able to take corrective action to prevent the faults from occurring in the first place. Nevertheless, fault prediction has rarely been investigated, perhaps due to its inherent challenges compared to standard FDD. In this work, we develop a machine-learning-based modular framework to predict faults in chiller systems based on timeseries sensor data. Specifically, the overall model consists of autoencoder, classifier, and thresholder components, with multiple variants of each component type. These models undergo supervised training using labeled data containing both normal and faulty instances. To validate the proposed approach, we leverage various deep and shallow components (see the Terminology defined in Section 4) in the modular framework with data collected from multiple chillers over one or more years of operation. Sample results show the efficacy and applicability of the framework for fault prediction, thus illustrating the potential for real-world application.