[Objective] To address the issues that wind turbine rotor imbalance fault features are easily submerged by noise under complex turbulent wind conditions, multi-channel independent signal decomposition tends to cause frequency misalignment, and samples of mild and coupled imbalances are prone to confusion. 

[Method] To systematically address these issues, this paper proposes a collaborative feature extraction and noise reduction method that synergistically integrates Multivariate Variational Mode Decomposition (MVMD) and Permutation Entropy (PE). Instead of treating sensor data in isolation, four-channel multi-modal response signals are synchronously acquired, specifically including blade flapwise load, blade edgewise load, fore-aft acceleration, and lateral acceleration of the nacelle. By leveraging the shared center frequency mechanism inherent in MVMD, the proposed approach constrains multiple channels to share the same mode center frequencies during optimization, thereby realizing precise alignment of homologous fault frequencies and preserving intrinsic physical correlations among multi-source signals. Following decomposition, PE is utilized as a nonlinear dynamical measure to quantify the randomness and complexity of each decomposed Intrinsic Mode Function (IMF). An adaptive entropy threshold is applied to intelligently screen valid IMFs for reconstructing high-fidelity denoised signals, effectively amplifying low-frequency fault-sensitive components. Furthermore, comprehensive indicators covering time-domain statistical parameters, frequency-domain characteristics, and spectral distribution features are extracted. Principal Component Analysis (PCA) is subsequently applied to eliminate data collinearity and redundancies, fusing the extracted metrics into a compact 22-dimensional sensitive feature set. To ensure engineering relevance, a refined aeroelastic model of a 3 MW wind turbine is established via GH-Bladed software, and a representative dataset encompassing 10 typical operating conditions—covering normal states, single imbalance types, and various coupled imbalance scenarios—is constructed for rigorous validation. 

[Results] Quantitative verification demonstrates the superiority of the proposed framework. When the MVMD-PE-PCA fused features are fed into a Support Vector Machine optimized by the Sparrow Search Algorithm (SSA-SVM), the diagnostic model achieves a remarkable average accuracy of 94.11%. Comparative analyses reveal significant performance gains: the accuracy is improved by 21.61 and 11.25 percentage points compared to models utilizing raw, un-denoised features and those relying on conventional single-channel VMD denoised features, respectively. Additionally, the PCA-based feature fusion strategy proves far more effective than using isolated single-modal features or simply concatenating multi-channel features, yielding further accuracy improvements of 17.59 and 11.17 percentage points, respectively. These results confirm that the fused feature set possesses strong discriminative capabilities even for samples with subtle fault signatures. 

[Conclusion] The proposed method effectively resolves the issues of multi-channel frequency misalignment and noise interference. It significantly mitigates the confusion between samples of mild aerodynamic imbalance and low-level coupled imbalance, providing a highly robust feature extraction scheme for the health monitoring of wind turbine drive systems.