A fault diagnosis method for wind turbines gearbox based on adaptive loss weighted meta-ResNet under noisy labels. (December 2021)

• Record Type: Journal Article
• Authors: Zhang, Kai; Tang, Baoping; Deng, Lei; Tan, Qian; Yu, Haoshuai
• Is Part Of: Mechanical systems and signal processing. Volume 161(2021)
• Publication Date: 2021-12
• Subjects: Fault diagnosis -- Deep learning -- ResNet -- Meta-learning -- Noisy label; Structural dynamics -- Periodicals; Vibration -- Periodicals; Constructions -- Dynamique -- Périodiques; Vibration -- Périodiques; Structural dynamics; Vibration; Periodicals; 621
• DOI: 10.1016/j.ymssp.2021.107963
• Languages: English
• ISSNs: 0888-3270

Abstract

Highlights: A loss-weighted ResNet is designed to diagnose wind turbine' faults with noisy label. A meta-model is applied to adaptive weight the ResNet's loss guided by clean labels. Datasets from DDS and wind farms verify the ALWM-ResNet' effectiveness. The ALWM-ResNet can effectively identify the gearbox's status with 40% noisy labels. Abstract: The effectiveness of traditional supervised fault diagnosis methods for wind turbine gearboxes typically depends on accurate labels, which are time-consuming and challenging to obtain. However, owing to noisy labels in datasets generated because of uncontrollable artificial or objective reasons, a robust method against label noise interference must be investigated for the engineering application of fault diagnosis. Herein, a novel method based on an adaptive loss-weighted meta-residual network (ALWM-ResNet) is proposed to address fault diagnosis with noisy labels using a weighted network and a meta-network cloned from the original ResNet to establish a weighted function mapping to adaptively learn weights from data with clean labels. The feasibility and effectiveness of the ALWM-ResNet are verified using the simulation gearbox dataset from the drivetrain diagnostic simulator test-bed and the engineering historical data of a wind farm obtained through the condition monitoring system. The results show that the proposed method improves the accuracy of the original ResNet by 30.52% and 22.44% for the simulated and wind farm datasets with 40% noisy labels, respectively.