Refrigerant charge fault diagnosis for variable refrigerant flow system based on ReliefF and mRMR coupling feature selection

For the refrigerant charge fault, presents a fault detection and diagnosis method combined with ReliefF and mRMR feature selection algorithm. Firstly, calculates the weight coefficient W(A) of each feature variable by the ReliefF algorithm, and removes the features with weight coefficient lower than the threshold. Then, selects the features which have the max-relevance with the target class and min-redundancy with each other by the ensemble mRMR algorithm. Finally, develops a BP neural network model based on the newly selected variables to achieve fault detection and diagnosis, and compares the results with that of single feature selection algorithm. The result indicates that the new method can improve the accuracy and efficiency of the diagnostic model.