Fault classification method based on fast k-nearest neighbor with hybrid feature generation and K-medoids clustering

Fast and accurate classification of the process faults in industry is important for ensuring reliable operation. The traditional K-Nearest Neighbor (KNN) algorithm needs to calculate the distance between the sample to be classified and all the training samples obtained under normal operation condition (NOC). Large computational cost involves in this step when the number of the NOC samples is huge. Aiming at this, existing methods reduce the training samples by using the clustering algorithms, so as to reduce the computational cost of the KNN. However, the reduction of training samples usually leads to a decrease in the accuracy of fault classification. Misclassification directly affects the normal production and safety of the industrial processes. To this end, a fast KNN fault classification method based on hybrid feature generation and K-Medoids is proposed. Firstly, a hybrid feature generation method combining ReliefF algorithm and linear discriminant analysis algorithm is used to select and extract the sample features, thus enhance the separability inter-classes and improve the accuracy of fault classification. Then, K-Medoids clustering algorithm is used to select few representative training samples and reduce the computational complexity of KNN algorithm. Finally, the simulation is performed on the Tennessee-Eastman process, which verifies that the proposed algorithm is superior than other related four methods and only consumes far less running time than the basic KNN algorithm while retaining higher classification accuracy.