A robust fault diagnosis approach for large-scale production process

Abstract Due to the importance and dangerousness of large-scale production processes, the accuracy and reliability of fault diagnosis approaches are critical for safe operation. In this paper, a robust fault diagnosis approach is proposed to realize the reliable classification while ensuring high accuracy. The feature importance distribution is proposed to select appropriate dimension reduction methods, and the real data structure is remained by the same-scale standardization and same-criterion dimension reduction. By means of the whole procedure optimization for datasets and classifiers, the performances of traditional Support Vector Machine and Naive Bayes can achieve the level of ensemble learning. Next the parallel classifier which consists of different classification theories is able to improve the reliability of final prediction. Experimental results show that the proposed approach outperforms the traditional approaches with accuracy that exceeds 92% for Tennessee Eastman benchmark (18 faults) and exceeds 87% for a real-world three-phase flow process (2 faults).