Abstract

The thesis tackles major challenges in developing fault detection and diagnosis (FDD) models for heating, ventilation, and air conditioning (HVAC) systems. HVAC systems are crucial for ensuring thermal and air comfort in buildings, but they frequently face inefficiencies due to inadequate maintenance, component wear and tear, and control issues, leading to considerable energy waste. The research highlights several key challenges, such as data contamination, lack of labeled data, and high dimensionality, which limit the effectiveness of traditional FDD methods.
To overcome these challenges, the thesis introduces innovative approaches: a supervised multiclass version of the deep autoencoding Gaussian mixture model (DAGMM) that improves outlier detection by leveraging label information; a neural network-based approach for mixtures of probabilistic principal component analyzers (NN-MPPCA) with a robust loss function for enhanced anomaly detection; a new framework combining variational autoencoders (VAE) with NN-MPPCA to handle high-dimensional data and incomplete datasets; and the adaptive adversarial autoencoder (AdaAAE), which refines anomaly detection with deep support vector data description (DSVDD). Comprehensive validation against cutting-edge algorithms highlights the superior performance of the proposed methods, leading to more efficient and reliable HVAC systems.