New fault diagnostic strategies for refrigerant charge fault in a VRF system using hybrid machine learning method

Abstract Fault diagnosis of a refrigerant charge is significant for the stable operation of a variable refrigerant flow (VRF) air conditioning system. At present, machine learning methods are widely used to diagnose faults associated with the model training. However, new faults that are not associated with the process are rarely considered for diagnosis. On this basis, this study presents two strategies using hybrid back propagation neural network (BPNN) and decision tree (DT) methods, namely, distinguish threshold-based and vector similarity principle-based strategies. This study aims to construct a diagnostic model that can diagnose new faults for a VRF system. The data of the refrigerant charge amount fault are employed to verify the strategies and are divided into training, validating, and testing datasets. The training dataset simply constituting known faults is first employed to train the proposed hybrid model. The validating dataset is employed to evaluate the diagnostic performance of a hybrid model. In a well-trained hybrid model, the BPNN is responsible for diagnosing known and new faults from the testing dataset. Meanwhile, the DT is responsible for diagnosing the specific fault type from the known faults. Results show that the hybrid model using the proposed strategies possesses good diagnostic performance. The hit rates of the new faults can yield 94.47% and 97.44%. The correct rates of the testing dataset can yield 95.2% and 96.63%.