A hybrid ICA-BPNN-based FDD strategy for refrigerant charge faults in variable refrigerant flow system

• Published in: Applied thermal engineering 2017-12, Vol.127, p.718-728
• Main Authors: Sun, Shaobo, Li, Guannan, Chen, Huanxin, Huang, Qianyun, Shi, Shubiao, Hu, Wenju
• Format: Article
• Language: English
• Subjects: Air conditioners; Air conditioning; Air conditioning equipment; Algorithms; Artificial neural networks; Back propagation networks; Back-propagation neural network; Fault detection; Fault detection and diagnosis; Fault diagnosis; Independent component analysis; Model accuracy; Neural networks; Refrigerant charge fault; Refrigerants; Studies; Variable refrigerant flow
• ISSN: 1359-4311; 1873-5606
• DOI: 10.1016/j.applthermaleng.2017.08.047

•A data-based method is proposed to detect and diagnose refrigerant charge faults.•The performance of ICA model for fault detection is superior to BPNN model.•ICA model to detect fault and BPNN model to diagnose fault have a good effect.•ICA algorithm performs dimensionality reduction on the original data from 12 to 4.•The accuracy rate improves from 82.7% to 93.8% by the hybrid ICA-BPNN model. Refrigerant charge fault is inevitable in air conditioning systems causing serious negative influences on system performance. This study presented a hybrid ICA-BPNN-based fault detection and diagnosis (FDD) strategy for refrigerant charge faults in variable refrigerant flow (VRF) system. It consists of two steps. Firstly, the independent component analysis (ICA) method is employed to detect the faults, and the normal-charge operating data set is used to train the ICA model. Secondly, a fault diagnosis model is established using the back-propagation neural network (BPNN) method, and the BPNN model is trained by the faulty operating data set with labels. The results show that the original data dimensions are reduced from 12 to 4 by the ICA algorithm. ICA-based method can detect the faults using both I2-statistic and I2-SPE-statistic. The accuracy rates are 93.6% and 95.9% respectively. Combined with the BPNN model, the hybrid ICA-BPNN model shows good fault diagnosis performance. Compared with single BPNN method, the hybrid model improves the accuracy rates from 82.7% to 93.8% for overcharge fault data using I2-SPE-statistic.