Data augmentation based open set fault diagnosis on building energy systems

Abstract Deep learning has been extensively applied in the health management of building energy systems. Yet, the majority of existing research is based on the assumption of closed set, which may classify the samples from unknown faults into the known fault categories with high confidence, leading to a drastic performance reduction in real application. In this research, open set fault diagnosis problem is proposed, aiming for effective recognition of unknown faults and correct classification of known faults. To tackle this problem, a novel data augmentation based open set fault diagnosis framework is developed. The core idea is to artificially generate several samples by Soft Brownian Offset sampling method to describe the boundary between known and unknown faults, and then develop a K+1 classifier to directly classify the testing samples into known class or unknown class according to the Softmax probability. Validation experiments are conducted on LBLN FDD dataset. Results show that the proposed method achieves the highest fault diagnosis accuracy of 0.9974 compared with other existing open set FDD methods. This research would expand the application scope of deep learning model in real scenarios.