Hybrid model- and learning-based fault diagnosis in adaptive buildings

Abstract

Adaptive buildings offer an enormous potential for saving resources and reducing emissions due to their ability to actively compensate deformations, which allows for a significantly lighter supporting structure. The long-term autonomous operation of an adaptive building, a prerequisite for its efficiency, requires the accurate detection and isolation of faults in its sensors and actuators. However, conventional model-based approaches achieve inadequate performance in case of substantial model errors. In that context, this article investigates the potential of integrating unsupervised learning techniques into model-based diagnosis schemes to improve the diagnostic accuracy. Specifically, we propose to train an autoencoder, a type of neural network, to suppress the effects of model errors in parity space residuals. A publicly available dataset of measurements from an adaptive high-rise building is introduced and used for the experimental validation of the proposed diagnosis method. The results are discussed in relation to a similar approach based on the principal component analysis (PCA), as well as standalone model- or learning-based approaches as reference. In different test scenarios, either the autoencoder- or the PCA-based approach is able to suppress the effects of model errors more effectively, yielding a more accurate fault detection. The PCA-based approach however allows for a more accurate fault isolation due to the exact propagation of the considered probability distributions.