Novel transformer-based self-supervised learning methods for improved HVAC fault diagnosis performance with limited labeled data

Existing data-driven HVAC fault diagnosis methods mainly adopt supervised learning paradigms, making them less feasible/implementable for individual buildings with limited labeled data. Considering the demanding requirements of domain expertise and labor work associated in data labeling, advanced data analytics are urgently needed to utilize massive unlabeled operational data for reliable predictive modeling. Therefore, this study proposes a novel transformer-based self-supervised learning methodology for improved HVAC fault diagnosis performance using limited labeled data. Three self-supervised learning approaches are developed to extract knowledge from unlabeled operational data through self-prediction and contrastive learning tasks. A customized transformer-based neural network is designed to ensure the efficiency and effectiveness in tabular data analysis and knowledge transfer. Data experiments have been conducted using multiple HVAC datasets considering different data availabilities, self-supervised learning approaches and model architectures. The results validate the capabilities of self-supervised learning in developing reliable HVAC fault classification models. Compared with conventional supervised learning solutions, the methodology proposed not only substantially reduce the data labelling works required, but also improves the fault diagnosis performance by up to 8.44%. The research outcomes are valuable for upgrading predictive modeling protocols in the building field for developing easy-implementation and high-performance data-driven solutions with limited labeled data.