Abstract
Prolonged operation of Variable Refrigerant Flow (VRF) systems invariably leads to malfunctions that can significantly increase energy usage and operational costs, while reducing system efficiency and comfort. Data-driven methods are increasingly recognized for their efficacy in fault diagnosis for VRF systems. Nonetheless, acquiring comprehensive training datasets encapsulating the full spectrum of potential operational anomalies remains a formidable challenge. This limitation invariably undermines the model's generalization capacity across diverse real-world scenarios. Unsupervised transfer learning is a prospective method that enables leveraging extensive unlabeled datasets to refine and enhance fault diagnosis models under conditions of scarce labeled data. However, few studies explore the impact of strategic training data categorization within source and target domains on the efficacy of transfer learning, particularly concerning transfer tasks across varying operational conditions, load rates within identical systems, and model stability amidst datasets featuring disparate fault categories. This paper aims to fill the research gap by thoroughly exploring domain adversarial method under unsupervised transfer learning framework for VRF system fault diagnosis, and providing strategies to enhance model performance and verify model stability amidst these challenges. The findings indicate that for VRF systems with constrained datasets, the unsupervised transfer learning model achieved accuracies, false alarm rates, and missing alarm rates of 89.71%, 5.24%, and 7.74% respectively. This represents an average enhancement of 25.01%, 34.38%, and 4.56% over the baseline model. The proposed enhancement strategy for complex transfer tasks yielded an average improvement of 10.95%, 16.55%, and 2.58% across three evaluation metrics. Furthermore, the study confirmed the transfer learning model's stability across datasets with varied number of fault categories.
Original language | English |
---|---|
Article number | 111917 |
Journal | Journal of Building Engineering |
Early online date | 24 Jan 2025 |
DOIs | |
Publication status | E-pub ahead of print - 24 Jan 2025 |
Keywords
- Variable refrigerant flow
- Fault diagnosis
- Information poor systems
- Transfer learning
- Unsupervised learning