A Two-Stage Transfer Learning Approach for Wind Turbine Blade Fault Classification with EfficientNetB0

Abstract

Wind energy is one of the fastest-growing sources of renewable electricity, and the reliable operation of wind turbines depends strongly on the condition of their blades. Blades are exposed to continuous mechanical and environmental stress, and surface defects that go unnoticed can lead to costly repairs and unplanned downtime. Manual visual inspection is still widely used, but it is slow, subjective, and hard to scale across large installations. This paper presents an image-based method for distinguishing healthy and faulty wind turbine blades using transfer learning with EfficientNetB0. The publicly available Small Wind Turbine Blade Dataset (CAI-SWTB), containing 6,000 balanced RGB images, was used with its predefined training, validation, and test partitions. A two-stage training strategy was adopted: in the first stage the ImageNet-pretrained backbone was frozen and only a new classification head was trained, and in the second stage the last twenty convolutional layers were unfrozen and fine-tuned at a lower learning rate while the batch-normalization statistics were kept fixed. On the independent test set of 1,200 images, the fine-tuned model reached an accuracy of 91.42%, a macro F1-score of 0.914, and a ROC-AUC of 0.978. Healthy blades were recognized more reliably (recall 0.962) than faulty ones (recall 0.867), meaning that a portion of damaged blades were still labelled as healthy. The results indicate that transfer learning with EfficientNetB0 offers a practical basis for automated blade inspection, while also pointing to the reduction of missed faults as the main direction for further work.