A feature importance ranking based fault diagnosis method for variable-speed screw chiller

Abstract

Fault detection and diagnosis (FDD) technology plays an important role in maintaining the stable and efficient operation of the chiller system, and screening the most important parameters for FDD is the key step that is far more critical than the selection of diagnosis methods. This study proposes a feature importance ranking method based on random forest (RF) to improve the FDD performance of a chiller while using fewer sensors, which is proved to be effective and efficient. When the product type or sensor configuration changes, the proposed method can be used to rank the feature importance efficiently based on the new fault simulation data, the FDD model using the features that are highly significant to fault indication and diagnosis can be established and the diagnosis performance can be promoted, accordingly. For verification and validation, fault simulation experiments of normal operation, refrigerant leakage and refrigerant overcharge have been carried out on a 200-ton variable-speed screw chiller. The importance of 15 parameters obtained from the experiments is ranked according to their contributions to the diagnosis of the corresponding faults using the proposed RF importance ranking method. Different critical features have also been discussed in terms of FDD performance and expertise knowledge for screw chiller. The optimal feature set for the investigated faults of the chiller is found to have just four parameters: Refrigerant Discharge Temperature (TR_dis), Condenser Water Temperature Difference (TWCD), Evaporator Water Temperature Difference (TWED) and Compressor Power (kW). The overall diagnostic accuracy reaches 99.90%, higher than using all the parameters. Fewer sensors achieve better performance.