Chiller fault detection and diagnosis by knowledge transfer based on adaptive imbalanced processing

The existing fault detection and diagnosis (FDD) model of chillers requires considerable normal and fault data. The acquisition of these data is time-consuming and expensive, and the model is only suitable for special units, which makes it difficult to popularize FDD technology in the operation and management of chillers. At present, a 120-ton chiller has only a small amount of normal and fault data when compared with the abundant data of a 200-ton model of the same series. This study investigates the FDD model of a 120-ton chiller and considers similar characteristics of the refrigeration cycle of the same series of chillers. A training set can be created using the 200-ton prior-knowledge data and the 120-ton data. However, this training set is imbalanced, and the common imbalanced processing synthetic minority oversampling technique (SMOTE) synthesis mechanism has an overlap problem. This study adopts two adaptive imbalance processing technologies called the adaptive synthetic sampling approach (ADASYN) and borderline SMOTE (BSM) that can solve the imbalance problem and SMOTE oversampling overlap problem during knowledge transfer. A support vector machine FDD model with 100% to 400% oversampling ratios is established. The best model is ADASYN with less than 100% oversampling ratio, with a diagnostic accuracy rate of 94.33%.