jaue2021-076: Deep-learning-based fault diagnosis method for centrifugal chillers under imbalanced dataset condition

Abstract

Data-driven fault diagnosis technology for energy-supplying systems can quickly and accurately detect and
diagnose the system faults, effectively avoiding energy waste and keeping system healthy. Due to the low probability
of system faults, the training dataset becomes imbalanced, which will significantly influence the diagnosis accuracy.
This paper proposes a novel method based on deep learning for centrifugal chillers fault diagnosis, which firstly
converts the system operating data into Gramian Angular Field (GAF) images and uses Convolutional Neural
Network (CNN) to classify constructed images according to different operating conditions. In order to solve the
imbalanced dataset problem, this paper adopts conditional Wasserstein Generation Adversarial Network with
gradient penalty (CWGAN-GP) method to expand the fewer samples to balance the training dataset. The result
shows that the proposed method has 89.80% diagnosis accuracy for typical faults of centrifugal chillers，much
higher than general machine learning methods. After balancing the training dataset through the method of generating
fewer samples, the diagnostic accuracy will be obviously improved, reaching more than 96%, which is more suitable
for actual engineering situation.