Unsupervised Fault Detection for Refrigeration Showcase Systems with Kernel Principal Component Analysis based Multivariate Statistical Process Control using Feature Selection with Maximal Information Coefficient

This paper proposes a kernel principal component analysis (KPCA) based multivariate statistical process control (KPCA-MSPC) method for fault detection of refrigeration showcase systems using a feature selection method with maximal information coefficient (MIC). Refrigeration showcase system data include non-linear relationships among pairs of features, and only normal data can be available for training generally. KPCA-MSPC is suitable for the fault detection because it is an unsupervised method and can handle non-linear relationships. In showcase systems, a large number of measured data can be obtained and they can be utilized as features for fault detection. However, considering system costs, the number of sensors installed in the showcase systems and the amount of data stored in data centers are limited. Therefore, a feature selection method based on MIC and k-nearest neighbor algorithm (KNN) (MIC-KNN-FS) suitable for KPCA-MSPC is proposed. The effectiveness of the combination of KPCA-MSPC and the proposed MIC-KNN-FS for showcase systems is verified by comparison with the Laplacian Score feature selection method (LS-FS) and the KNN feature selection method (KNN-FS), which are typically utilized as feature selection methods, and cumulative autoencoders (CAE) and MSPC based on PCA (PCA-MSPC), which are unsupervised fault detection methods.