Fault Diagnosis of Chillers using Dimensionality Reduction Methods

Refrigeration systems, such as chillers, consist of interconnected mechanical components. These interdependent systems may have various faults that can be detected using comprehensive sensor measurements and system parameters. This study utilizes the benchmark chiller data collected by ASHRAE project 1043-rp and investigates machine learning and deep learning approaches for fault detection. Dimension reduction on this high dimensional data is investigated, and the compressed features are evaluated to classify various faults. Specifically, Principal Component Analysis (PCA) and Long Short-Term Memory Auto-Encoder (LSTM-AE) is applied as dimensionality reduction methods, and Support Vector Machine (SVM) and Neural Network (NN) are selected as classifiers. An LSTM model using the original data is used as a baseline. This study proposes a model architecture that enables combination of dimensionality reduction and classification choices.