Fault Diagnosis Strategy Of VRF System Refrigerant Charge Amount Based On Ensemble Feature Selection And LightGBM

The VRF system has the characteristics of reliable operation,advanced control,and good adaptability of the unit.It has been widely used in small and medium-sized buildings and commercial buildings in recent years.Typical failures of the system mainly include refrigerant charge failure,condenser dirt failure and sensor failure.Diagnosing various types of system failures in time is of great significance for energy saving and emission reduction.This paper uses data mining methods for VRF system refrigeration charge amount failure detection.First,conduct experiments in the enthalpy difference laboratory to obtain various operating data of the VRF system under normal and fault conditions.The experimental data includes ten charge levels.In order to effectively distinguish the initial faults,this article divides the experimental data into seven Kinds of fault categories.Through the DBSCAN algorithm,the abnormal samples in the experimental data are detected,and the abnormal values are eliminated,which improves the data quality.Then use five feature selection methods such as maximum correlation minimum redundancy(mRMR)algorithm,Relief F algorithm,recursive feature elimination(RFE)algorithm,Boruta algorithm and Xgboost algorithm to filter the variables in the experimental data,and then use the mean method and voting method The integrated feature selection method integrates a single ranking.This paper proposes a forward search integrated feature selection method,which brings the variables obtained by the five single feature selection methods and the three integration methods into the LightGBM prediction model,and compares different feature selection methods The fault diagnosis ability of the obtained variables is finally optimized by grid search method to improve the model’s fault diagnosis abilityThe fault diagnosis results show that the variables obtained by the forward search integration method have the best fault diagnosis performance.The first eight variables obtained by the forward search constitute the optimal feature subset.After the feature selection and optimization,the running time of the model is reduced by 50%.The efficiency is improved,and the diagnostic accuracy is 94.71%.After the parameter optimization,the fault diagnosis accuracy rate of the model reached 96.91%,and the optimal feature subset was brought into the comparative models such as Adaboost,Xgboost and GBDT.The fault diagnosis of the comparative models was higher than 96%,proving the forward search integrated feature selection method The applicability is strong.In addition,the fault diagnosis accuracy of the LightGBM model is the highest,and its running time is only 27.5% of the Adaboost model,12.49% of the GBDT model,and 13.52% of the Xgboost model,illustrating the superiority of combining ensemble feature selection with the LightGBM fault diagnosis strategy.