A robust VRF fault diagnosis method based on ensemble BiLSTM with attention mechanism: Considering uncertainties and generalization

• Published in: Energy and buildings 2022-08, Vol.269, p.112243, Article 112243
• Main Authors: Han, Linzhi, Deng, Yuping, Chen, Huanxin, Wei, Gou, KaiSheng, Shi, Jingfeng
• Format: Article
• Language: English
• Subjects: Attention; BiLSTM; Fault diagnosis; Variable refrigerant flow system
• ISSN: 0378-7788
• DOI: 10.1016/j.enbuild.2022.112243

•Proposed an interpretable deep learning model using the attention-BiLSTM.•Four typical faults of VRF are well identified with an overall diagnosis accuracy of 97.5%.•Effective fault diagnosis under incomplete information is realized through the integrated model of different feature combinations.•Realize the quantification of diagnosis uncertainty through the ensemble model.•The model performs well on the unknown data set under different working conditions. Developing effective fault detection and diagnosis (FDD) model is of great significance to the energy efficiency and comfort of variable refrigerant flow (VRF) system. In recent FDD studies, the consideration of incomplete information and uncertainty, as well as the generalization of model and the interpretability of fault action mechanism have become a great concern. This paper accordingly proposes a fault diagnosis strategy based on attention-BiLSTM with ensemble feature sets. Three methods including Correlation Analysis, Information Entropy, and Gini Impurity are used for feature selection and ensemble feature sets combination. The proposed model can achieve interpretability of fault diagnosis and quantification of diagnosis uncertainties through attention mechanism and ensemble feature sets. A total of four typical faults under different working conditions are used to verify the accuracy and generalization of the proposed method. The results show that the ensemble attention-BiLSTM model has a fault diagnosis accuracy rate of 98.3% if the first two most likely outputs are considered, and when some sensors fail, the accuracy rate can still remain above 90%. In addition, a generalization verification idea is proposed, that is, the model is trained under specific operating conditions and tested on datasets with different operating conditions or fault levels. The results show that the average generalization of the optimized model reaches 86%.