Transfer learning based fault diagnosis of automobile dry clutch system. (January 2023)

• Record Type: Journal Article
• Authors: Chakrapani, G.; Sugumaran, V.
• Is Part Of: Engineering applications of artificial intelligence. Volume 117:Part A(2023)
• Publication Date: 2023-01
• Subjects: Dry clutch -- Transfer learning -- Deep learning -- Fault diagnosis -- AlexNet -- VGG16 -- GoogLeNet -- Resnet-50; Engineering -- Data processing -- Periodicals; Artificial intelligence -- Periodicals; Expert systems (Computer science) -- Periodicals; Ingénierie -- Informatique -- Périodiques; Intelligence artificielle -- Périodiques; Systèmes experts (Informatique) -- Périodiques; Artificial intelligence; Engineering -- Data processing; Expert systems (Computer science); Periodicals; 620.00285
• DOI: 10.1016/j.engappai.2022.105522
• Languages: English
• ISSNs: 0952-1976

Abstract

Abstract: Dry friction clutches are prone to fault occurrences due to their continuous exposure to thermal loading and high abrasive rate during power transmission. Fault occurrences in clutches can lead to damage in internal components, downtime and seizure of transmission system. Thus, it is necessary to detect and diagnose such fault occurrences at an early stage. In the present study, an attempt was made to diagnose various clutch faults like release fingers worn out, pressure plate broken, pressure plate worn out, friction material loss, and tangential strip bent using deep learning technique (transfer learning). Vibration signals were acquired from a test rig operated under various load conditions for different clutch conditions (5 faulty and 1 good) that were further processed and stored as vibration plots. The concept of transfer learning involving four pre-trained network models namely, AlexNet, VGG-16, GoogLenet and ResNet-50 were utilized in the present study. Various hyperparameters such as train–test split ratio, learning rate, optimizer and batch size were altered and the best hyperparameters suitable for achieving high classification accuracy for every pre-trained network was determined. Overall, VGG-16 pre-trained network performed exceptionally well among the counterparts with a classification accuracy of 100% in a computational time of 449 s.