Improved SMOTE based LSDA for Class Imbalanced Fault Diagnosis in TE Industrial Process

Abstract

With the technological progress, industrial processes are developing towards intelligence, and the different classes of data collected show imbalanced features due to the new technologies put into the production process. Due to the imbalance of data classes, it brings difficulties to fault diagnosis. To address the imbalance, high-dimensional and non-linear characteristics of data in the actual production process, this paper proposes a novel fault diagnosis method with improved SMOTE-LSDA (ISMOTE-LSDA). Firstly, in ISMOTE-LSDA, an improved synthetic minority over-sampling technique (ISMOTE) algorithm is used to resample the imbalanced samples and expand the few classes of fault samples. The ISMOTE uses not only the Euclidean distance but also the cosine distance to calculate the K nearest neighbors (KNN) distance, and the nearest neighbor effect is more accurate. Then the manifold learning locality sensitive discriminant analysis (LSDA) algorithm is used to dimensionality reduction of high and non-linear balance samples to extract meaningful data features. Finally, an integrated learning Adaboost classifier is used for fault classification. The TE chemistry process is used to validate the effectiveness of the proposed ISMOTE-LSDA method. The experimental simulation results show that the proposed method has better fault diagnosis accuracy compared to the traditional feature extraction methods which is suitable for the fault diagnosis process with unbalanced data classes.