Hybrid CNN-GRU framework for wind turbine blade defect classification and data-driven severity assessment for predictive maintenance

Abstract

Abstract The integrity of wind turbine blades is essential for the continuous production of renewable energy. If left unchecked, surface pollution and fissures hasten structural deterioration, lower aerodynamic efficiency, and result in catastrophic blade failure with serious safety and economic ramifications. The hybrid Convolutional Neural Network and Gated Recurrent Unit framework presented in this study was verified using a confidential industrial dataset consisting of 2,995 DJI drone-captured wind turbine blade inspection images. This study aims to use structured dimensionality reduction to preprocess and classify blade surface defects by fault cause and create a Convolutional Neural Network and Gated Recurrent Unit model to anticipate possible blade failures from drone images. MobileNetV2 was used as the spatial feature extractor. The statistical distribution of annotation bounding box regions was used to directly determine crack severity levels, substituting a data-driven method for arbitrary threshold selection. On the held-out test set, the Convolutional Neural Network and Gated Recurrent Unit achieved 95.67% accuracy, 87.13% F1-Score, and an area under the receiver operating characteristic curve of 0.9756. Removing the Gated Recurrent Unit decreases the F1-Score from 87.13% to 47.50% and the dirt-class sensitivity from 89.80% to 38.78%, according to a critical ablation investigation, offering conclusive experimental support for the Gated Recurrent Unit contribution. The superiority of the proposed model under class imbalance is demonstrated by a comparative evaluation against five baselines, including the Coordinate Attention Convolutional Neural Network, MobileNetV2-only, ResNet50, EfficientNetB0, and Yolov5s. A comprehensive Predictive Maintenance decision framework creates a deployable end-to-end solution that supports Sustainable Development Goals 7 and 9 by mapping model outputs to practical industrial maintenance responses.