Abstract
The cost of operation and maintenance of wind farms has been increased significantly due to insufficient attention to structural damage of wind turbine blades. There is currently a great demand for structural health monitoring of wind turbine blades that could identify the location of damage source timely. The heterogeneity of the wind turbine blades poses a great challenge for damage source localization. A novel damage source localization method based on acoustic emission (AE) and long short-term memory neural networks is developed for heterogeneous structure of wind turbine blades in this paper. Two machine learning models based on long short-term memory (LSTM) neural networks, LSTM-Atten and LACNN, are pre-trained and utilized in a two-step approach to minimize the influence of propagation characteristics of AE signals. The first step is to identify the zone where the damage source is located and the second step is to determine the exact coordinates of the damage source. The time difference of arrival (TDOA) is used as the training feature. TDOA is calculated by the time of arrival (TOA), which is determined by the Akiake information criterion (AIC). Then, an example application of damage source localization for a “spar cap-trailing panel-shear web” heterogeneous structure of a wind turbine blade is performed to evaluate the performance of the method. A comprehensive comparison between the proposed method and the damage source localization method based on “L” shaped sensor cluster is carried out to verify the superiority of the proposed method. The results demonstrate the feasibility and robustness of the proposed damage source localization method for heterogeneous structure of wind turbine blades.
