Abstract: During the operation of wind turbines, blade vibration data exhibit significant differences in feature distributions under varying operating conditions and are often affected by noise, making fault feature extraction challenging and accurate health state identification difficult. To effectively address this challenge, this study proposes a blade health state monitoring method based on gramian angular field and multi-head attention convolutional neural network. First, a modal transformation approach based on Gramian Angular Field is employed to encode the one-dimensional vibration signal into two types of two-dimensional image representations: Gramian angular difference field (GADF) and gramian angular summation field (GASF). By using polar coordinate mapping, this method effectively preserves the amplitude and phase characteristics of the time-series signals. Second, a parallel dual-channel Convolutional Neural Network (CNN) is constructed to extract spatial features from GADF and GASF images separately, incorporating depthwise separable convolutions to reduce model complexity. Finally, a multi-head attention mechanism is introduced to perform cross-channel feature fusion, enhancing the expression of critical fault features through adaptive weight allocation.To verify the effectiveness of the proposed method, experimental validation is conducted using real-world data on blade fatigue damage from wind farms, ensuring its applicability and reliability in practical scenarios. The experimental results demonstrate that the proposed method achieves a diagnostic accuracy of 93.7%, significantly outperforming other comparative methods, thereby confirming its effectiveness and superiority in real-world applications.

Key words: wind turbine blades, assessment of health status, gramian angular field, convolutional neural network, multi-head attention