A novel gas turbine fault diagnosis method based on transfer learning with CNN. (April 2019)

• Record Type: Journal Article
• Authors: Zhong, Shi-sheng; Fu, Song; Lin, Lin
• Is Part Of: Measurement. Volume 137(2019)
• Page Start: 435
• Page End: 453
• Publication Date: 2019-04
• Subjects: Transfer learning -- Fault diagnosis -- Gas turbine -- Small sample -- CNN -- SVM; Weights and measures -- Periodicals; Measurement -- Periodicals; Measurement; Weights and measures; Periodicals; 530.8
• DOI: 10.1016/j.measurement.2019.01.022
• Languages: English
• ISSNs: 0263-2241

Abstract

Highlights: Large individual differences, obvious data noises and fewer fault samples. Tendency of performance parameters have changed obviously when a fault occurred. The arrangement order of performance parameters has impact on fault diagnosis. A feature mapping method is established by transfer learning with CNN. The mapped feature in new feature space is classified by SVM. Abstract: A transfer learning method based on CNN and SVM is investigated for gas turbine fault diagnosis. The excellent classification ability of CNNs is attributed to their ability to learn rich feature representations from a large number of annotated samples. This property, however, currently prevents application of CNNs to problems with fewer samples. This paper shows how feature representations learned with CNN on large-scale annotated gas turbine normal dataset can be efficiently transferred to fault diagnosis task with limited fault data. A feature mapping method to extract the feature representations for fault dataset by reusing the internal layers of CNN trained on the normal dataset is designed, and SVM is used for fault diagnosis. The influence of gas turbine performance parameters arrangement order on proposed method is theoretically analyzed. Finally, the proposed method is validated by the real-life operation data of a gas turbine sample fleet. The experimental results show that despite difference in the two datasets, the transferred feature representations lead to significant improved results for fault diagnosis as well as obviously weaken the individual difference and data noises. The experimental results also confirm that the proposed method has excellent ability for fault diagnosis under small sample condition.