Study on a hybrid SVM model for chiller FDD applications

Fault detection and diagnosis (FDD) is the basis for timely maintenance to keep chiller systems operate at a normal and efficient condition. This study investigates a hybrid model that combines support vector machine (SVM) with genetic algorithm (GA) and parameter tuning technique for chiller FDD applications, where GA is responsible for searching potential feature subsets and SVM behaves both as an FDD tool and an evaluation method for feature selection. Subsets of 6, 7, 8, 9, and 10 features were studied, respectively, and compared with the original 64-feature set in terms of overall performance – correct rate (CR), and individual performance – hit rate (HR) and false alarm rate (FAR). The results showed that the eight-feature subset (Feat8) singled out by the hybrid SVM model behaves the best with its testing CR about 2% higher than that of the simple SVM model (64-features) while consuming less computational time. Further validation and comparison analysis with C4.5 FDD model also convalidate the outstanding performance of Feat8. Fewer features also mean fewer sensors required for data acquisition and accordingly less sensor cost. Moreover, a drastic improvement in individual performance (HR, FAR) was observed for the two most-difficult-to-be-identified faults – refrigerant leak (RefLeak) and refrigerant overcharge (RefOver).