Localizing low-velocity impacts on wind turbine blades using multi-sensor ensemble learning with dynamic arbitration

Composite material structures, such as wind turbine blades, are vulnerable to low-speed impacts throughout their entire life cycle, encompassing manufacturing, in-service operation, and maintenance. Consequently, developing efficient impact localization methods is crucial for ensuring structural safety. Traditional acoustic emission (AE) methods are challenging to implement due to their dependence on precise wave propagation models. Meanwhile, existing data-driven deep learning approaches commonly suffer from the performance bottlenecks of single models, thereby limiting localization accuracy. To this end, this study proposes a multi-sensor decision fusion framework based on ensemble learning. The framework first converts the one-dimensional time-series AE signals into two-dimensional grayscale images by combining complete ensemble empirical mode decomposition with adaptive noise and permutation entropy. Subsequently, multiple expert models are independently trained on data from different sensors, using a novel multi-scale pyramid attention network as the base model. Finally, a dynamic arbitration ensemble strategy (DAES) is employed for efficient decision-level fusion of the multi-model predictions. Experimental results demonstrate that the localization accuracy of the proposed framework significantly outperforms that of the best single-sensor model and data-level fusion models. Further generalization analysis indicates that the superiority of DAES does not stem from a coincidental combination of specific sensors, but rather represents a fusion paradigm with high generalization capability. This study offers a high-accuracy solution for the region-level impact localization problem in complex structures.