20 August 2025 in IEEE Transactions on Instrumentation and Measurement
In wind energy generation, monitoring the health of wind turbine blades is essential for ensuring efficient energy conversion and system safety. However, the complex operating environment and diverse failure modes of wind turbine blades pose challenges for traditional monitoring methods, which often struggle with data scarcity and feature extraction, leading to the poor fault classification performance. Consequently, this study introduces an innovative approach that combines a variational autoencoder (VAE) model with adaptive dimensionality selection and perceptual loss. By employing cumulative explained variance ratio (CVR) based on dimensionality reduction techniques and rate-distortion theory from information theory, the model’s latent space dimension is dynamically optimized using a threshold $v_{\text {th}}$ . Furthermore, the model incorporates perceptual loss based on multilayer features of the VGG network to enhance the semantic consistency of defect edges and textures. Experimental results demonstrate that the proposed framework achieves a remarkable accuracy of 98.23% $\pm ~0.35$ % in fault classification tasks while maintaining a lightweight architecture and producing high-quality reconstructions, and compared with the initial model, PSNR and SSIM were improved by 36.24% and 30.32%. Even in simulated complex scenarios, the model’s classification accuracy exceeds 90%. Ablation studies highlight the critical role of perceptual loss in improving the feature extraction. This work offers an efficient and cost-effective solution for wind turbine blade maintenance and supports the intelligent transformation of wind energy operations.
30 November 2025 in Expert Systems with Applications