Non-parametric semi-supervised chiller fault diagnosis via variational compressor under severe few labeled samples

Intelligent fault diagnosis for chiller is essential for energy efficiency optimization and health management within Heating, Ventilation, and Air Conditioning systems. Deep learning-based chiller fault diagnosis methods have demonstrated competitive performance. However, certain challenges remain: (1) in large-scale chiller systems, obtaining labeled data is costly, leading to scarce labeled datasets; (2) large amounts of unlabeled data have not been fully explored. This paper proposes a semi-supervised fault diagnosis algorithm based on system complexity quantification. Firstly, a variational autoencoder-based lossless compressor is trained using unlabeled data in an unsupervised manner. Subsequently, Kolmogorov complexity is approximated via the compressor to achieve entropy-based quantification of complexity. Using the normalized information distance algorithm, an information distance matrix is then calculated, which is combined with a K-nearest neighbors classifier for fault diagnosis. The proposed method is validated using the ASHRAE 1043-RP and HY-31C datasets, and the experimental results indicate that the proposed method can use the severe few labeled samples to obtain a better diagnostic accuracy.