Transfer learning-based fault detection and diagnosis for heating, ventilation, and air conditioning systems

Naghmeh Ghalamsiah

doi:10.17918/00011387

Although data-driven methods have shown great potential for fault detection and diagnosis (FDD) in heating, ventilation, and air conditioning (HVAC) systems, their performance heavily depends on the availability of high-quality labeled data. In practice, such data is difficult to collect from real systems, especially under faulted operation. As a result, simulated data or data from similar systems are often used for training, which can create significant distribution differences between training and testing data due to variations in outdoor conditions, control strategies, internal loads, and occupant behavior. Transfer learning (TL) offers a solution by leveraging rich labeled data from source domain, to diagnose faults in a shifted target domain. This study focuses on unsupervised domain adaptation (UDA), a branch of TL that aims to transfer knowledge to a fully unlabeled target domain. Within the reported UDA methods for HVAC FDD, there exists a few research gaps: 1) There lacks a comprehensive study to identify HVAC data pairs for evaluating TL algorithms. 2) While convolutional neural network (CNN)-based approaches remain most popular, their performance is limited due to their simplistic architecture. The critical step of properly preparing HVAC data for CNN-based models is also overlooked. 3) Although complex interactions between HVAC variables can better be represented by graphs, no prior research has explored the transformation of HVAC data into graph structures and the application of graph convolutional networks (GCNs). 4) There is limited effort to interpret the performance of developed models. 7) No work has investigated UDA's application to district-level systems. This research has addressed the above-mentioned research gaps through the following tasks: • Data pairs identification: Suitable HVAC fault dataset pairs and their associated faults for evaluating TL algorithms are identified through a comprehensive analysis. • CNN-based approach: T-CAN, a UDA framework that combines a causality-based method (for converting HVAC data into image representations) and contrastive adaptation network (a CNN-based algorithm originally introduced for image recognition), is developed and evaluated. • Graph-based approach: T-GCN, a UDA framework that systematically constructs graph-based representations of HVAC systems and applies GCN for diagnosing faults, is designed and evaluated. • Model interpretability: To improve transparency, layer-wise relevance propagation (LRP) is applied to both T-CAN and T-GCN to identify the critical features emphasized during diagnosis and to compare their interpretability. • Comparison and district-level evaluation: The performance of the developed methods is systematically compared and evaluated on multiple scenarios. The results show that both frameworks can successfully diagnose faults in unlabeled target domains, while T-GCN generally provides better overall performance in terms of diagnostic accuracy, computational efficiency, and interpretability. Finally, the scalability of the more promising framework (i.e., T-GCN) is evaluated on a district-level heat pump system, where it also demonstrates successful performance. Overall, the research presented in this thesis contributes to the overall objective of developing a TL-based framework for cost-effective and scalable HVAC fault diagnosis. Potential solutions to other challenges in implementing data-driven TL strategies, such as handling building performance disturbances (e.g., changes in weather and internal conditions) and addressing interpretability issues, are also examined.

Transfer learning-based fault detection and diagnosis for heating, ventilation, and air conditioning systems

Files and links (1)

Abstract

Metrics

Details

Transfer learning-based fault detection and diagnosis for heating, ventilation, and air conditioning systems

Files and links (1)

Abstract

Metrics

Details

Drexel University Social media