Abstract: The refrigerant charge ( ＲC) breakdown of chiller is a high-risk fault that directly affects the operational efficiency of air conditioning system． Due to the mass representative parameters，it is difficult to detect the fault on- line effectively，quickly and accurately． Aiming at these problems，a fault diagnosis method based on random fo- rests ( ＲF) algorithm was proposed for the ＲC breakdown by using the fault database of refrigeration provided by ASHＲAE in 1999． Directly measured feature parameters related to ＲC were analyzed． Under the premise of retai- ning the physical meaning，contribution rate of each fault feature was studied with ＲF algorithm． Then the detection accuracy of ＲC fault diagnosis based on ＲF，support vector machines ( SVM) and decision tree ( DT) algorithms were compared under the condition of different dimensions and sample size． The results show that the ＲF algorithm has the best recognition efficiency and highest classification accuracy． Compared with DT and SVM，the average detection accuracy of ＲF increases by 3. 3% and 2. 9% ，respectively． Furthermore，three important fault charac- teristics which have a high impact on the diagnosis of ＲC were analyzed． This can provide a theoretical basis for en- suring operating performance and safe operation of refrigeration system．

Key words: fault diagnosis, refrigerant charge, random forests algorithm, air conditioning system