A nearly end-to-end deep learning approach to fault diagnosis of wind turbine gearboxes under nonstationary conditions. (March 2023)

• Record Type: Journal Article
• Authors: Zhang, Liangwei; Fan, Qi; Lin, Jing; Zhang, Zhicong; Yan, Xiaohui; Li, Chuan
• Is Part Of: Engineering applications of artificial intelligence. Volume 119(2023)
• Publication Date: 2023-03
• Subjects: Fault diagnosis -- Deep learning -- End-to-end learning -- Empirical mode decomposition -- Convolutional neural network; 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.2022.105735
• Languages: English
• ISSNs: 0952-1976

Abstract

Abstract: Fault diagnosis of wind turbine gearboxes is crucial in ensuring wind farms' reliability and safety. However, nonstationary working conditions, such as load change or speed regulation, may result in an accuracy deterioration of many existing fault diagnosis approaches. To overcome the issue, this research proposes a nearly end-to-end deep learning approach to fault diagnosis of wind turbine gearboxes using vibration signals. Concretely, we adopt Empirical Mode Decomposition (EMD) to decompose vibration signals into a series of Intrinsic Mode Functions (IMFs). Then, the multi-channel IMFs are fed into a 1D Convolutional Neural Network (CNN) for automatic feature learning and fault classification. Since EMD is a signal processing technique requiring no prior knowledge, the model architecture can be viewed as nearly end-to-end. The proposed approach was validated in a real-world dataset; it proved deep learning models have an overwhelming advantage in representation capacity over traditional shallow models. It also demonstrated that the introduction of EMD as a preprocessing step improves both the training efficiency and the generalization ability of a deep model, thus leading to a better fault diagnosis efficacy under variable working conditions. Highlights: Combining Empirical Mode Decomposition and 1D CNN, a nearly end-to-end deep learning approach to fault diagnosis is proposed. The proposed approach is empirically validated on a wind turbine gearbox dataset under nonstationary working conditions. To demystify the trained model, we made an in-depth study of the low-level convolutional kernels and high-level features.