Fault Diagnosis Of Sensor And Actuator Faults In Multi-Zone Hvac Systems

Abstract

Globally, the buildings sector accounts for 30% of the energy consumption and more than 55% of the electricity demand. Specifically, the Heating, Ventilation, and Air Conditioning (HVAC) system is the most extensively operated component and it is responsible alone for 40% of the final building energy usage. HVAC systems are used to provide healthy and comfortable indoor conditions, and their main objective is to maintain the thermal comfort of occupants with minimum energy usage. HVAC systems include a considerable number of sensors, controlled actuators, and other components. They are at risk of malfunctioning or failure resulting in reduced efficiency, potential interference with the execution of supervision schemes, and equipment deterioration. Hence, Fault Diagnosis (FD) of HVAC systems is essential to improve their reliability, efficiency, and performance, and to provide preventive maintenance. In this thesis work, two neural network-based methods are proposed for sensor and actuator faults in a 3-zone HVAC system. For sensor faults, an online semi-supervised sensor data validation and fault diagnosis method using an Auto-Associative Neural Network (AANN) is developed. The method is based on the implementation of Nonlinear Principal Component Analysis (NPCA) using a Back-Propagation Neural Network (BPNN) and it demonstrates notable capability in sensor fault and inaccuracy correction, measurement noise reduction, missing sensor data replacement, and in both single and multiple sensor faults diagnosis. In addition, a novel on-line supervised multi-model approach for actuator fault diagnosis using Convolutional Neural Networks (CNNs) is developed for single actuator faults. It is based a data transformation in which the 1-dimensional data are configured into a 2-dimensional representation without the use of advanced signal processing techniques. The CNN-based actuator fault diagnosis approach demonstrates improved performance capability compared with the commonly used Machine Learning-based algorithms (i.e., Support Vector Machine and standard Neural Networks). The presented schemes are compared with other commonly used HVAC fault diagnosis methods for benchmarking and they are proven to be superior, effective, accurate, and reliable. The proposed approaches can be applied to large-scale buildings with additional zones.