Refrigeration is a fast-growing industry and has become an integral part of various industries such as food storage, pharmaceutical, residential, chemical, and data centers. Rising global temperatures have increased the need for more energy-efficient and environmentally conscious refrigeration systems. A consistent functioning refrigeration system requires good fault detection and diagnosis (FDD) system to detect faults before failures occur. A good FDD system can help reduce maintenance costs and increase energy savings. The refrigerant present in the chillers consists of greenhouse gases. Certain faults, such as refrigerant leakage, result in the release of refrigerant into the atmosphere, which has an environmental impact. Hence, an effective FDD system for chillers is important for accurately detecting and diagnosing faults. This thesis aims to build a data-driven FDD system for vapor-compression chillers. The purpose of this thesis is to address the variable operating conditions and fault conditions that occur during the operation of a chiller system. The operating conditions vary depending on the control logic, climate, and other factors that can occur in a system with multiple components. Faults can occur to varying degrees within a system. Early fault detection and diagnosis lead to prompt maintenance dispatching. It is imperative that an FDD system accommodates these conditions and accurately detect faults. This thesis uses two types of data to build an FDD model. The experimental data provided by ASHRAE RP-1043, which contained both normal and fault conditions, were used. The ASHRAE dataset contains normal conditions and seven fault conditions at the four severity levels. Other types of data used were simulated data generated using the ASHRAE RP-1043 model and a small chiller model developed by the author. Simulated data supplemented the experimental dataset with different normal operating conditions and fault severity levels. A hybrid architecture consisting of a dimensionality reduction method and classifier was proposed. This architecture facilitates the comparison of machine learning and deep learning techniques, and may be employed to develop a hybrid model that incorporates both approaches. Time series data was used to train and test this architecture. A deviation matrix method, which is a preprocessing method applied to the training and testing datasets, was proposed. This method was used with steady-state data to develop a 2D Convolutional Neural Network (CNN), Artificial Neural Network (ANN), and Support Vector Machine (SVM). The deviation matrix method proved effective for machine learning and deep learning models to detect and diagnose faults for different normal and fault severity levels. A study using steady-state time-series data and complete test cycle data was conducted to build and test the hybrid architecture. It was concluded that using steady-state time-series data yielded a higher classification accuracy for faults. The deviation matrix method was applied to the simulation and experimental data, and 2D CNN, ANN, and SVM were used to study these datasets. It was concluded that some parts of the training data could be substituted with the simulation data to obtain acceptable classification accuracy. The 2D CNN, ANN, and SVM were able to diagnose faults in the test data containing different severity levels from the training set and small-chiller data. The SVM was also effective in detecting near-normal operations. / Dissertation / Doctor of Philosophy (PhD)