Adaptive variational mode decomposition with attention-guided deep learning for multi-source fault diagnosis of wind turbine rotor imbalance
TL;DR
A hybrid diagnostic framework integrating the Ivy algorithm-optimized variational mode decomposition with an attention-enhanced temporal convolutional network and support vector machine classifier is proposed, confirming the framework’s high accuracy and generalization capability for multi-source rotor imbalance identification.
Abstract
Rotor imbalance represents a critical fault type affecting the operational stability and power generation efficiency of wind turbines, commonly induced by multiple factors including blade mass deviation, pitch angle abnormalities, and structural damage. To address the intelligent identification of rotor imbalance faults arising from multi-source causes, this paper proposes a hybrid diagnostic framework integrating the Ivy algorithm-optimized variational mode decomposition with an attention-enhanced temporal convolutional network and support vector machine classifier. First, to overcome the empirical parameter selection limitations of variational mode decomposition, the Ivy algorithm adaptively optimizes its key parameters, enabling high-precision decomposition of nacelle vibration signals. Second, intrinsic mode functions with strong discriminative capability are selected via the envelope entropy criterion, which effectively identifies pronounced fault characteristics, and multi-dimensional time-domain statistical features are extracted to construct feature vectors. Third, to capture temporal dependencies and strengthen critical features, a temporal convolutional network enhanced with a convolutional block attention module is utilized. Finally, to enhance generalization in small-sample scenarios, a support vector machine is adopted as the classification decision layer, leveraging its maximum margin principle. The effectiveness of the proposed method was validated using the University of Mustansiriyah nacelle vibration dataset. The proposed model achieved a test accuracy of 99.43% and a macro-F1 score of 99.4%, significantly outperforming baseline models including standard temporal convolutional network (94.29%) and support vector machine (84.00%). These findings confirm the framework’s high accuracy and generalization capability for multi-source rotor imbalance identification.