AHU fault diagnosis strategy using transfer learning based on similarity measurement

Transfer learning methods have gained increasing attention in the field of fault detection and diagnosis (FDD) for HVAC systems due to their ability to solve the problem of bad model performance caused by data shortage and different operational patterns. However, the lack of quantitative analysis of correlations between source-domain and target-domain tasks results in insufficient generalization and poor interpretability of existing research results. In this study, an air handling unit (AHU) FDD model based on image and residual neural network (ResNet) is proposed. Experiments were carried out using the operational data of seven independent AHU systems. The experimental results show that the model converges quickly and achieves high diagnostic accuracy up to 100% when there is enough data. In the case of extremely limited data, the posttransfer model improves the diagnostic accuracy on each test dataset by about 20-60%. When the volume of data is raised, the accuracy of the posttransfer model stabilizes around 99%. In summary, this study demonstrates the reliability of the proposed fault diagnosis model and provides quantitative theoretical support for model transfer strategies.