Fatigue condition diagnosis of offshore wind turbine bearings in noisy environment based on multi-scale integrated dual-channel CNN and improved GWO optimized ESN

Abstract Offshore wind turbines are a key component of renewable energy systems, and their reliable operation is essential for sustainable power generation. In particular, bearing fatigue condition diagnosis (FCD), as a critical task for offshore wind turbines, directly affects operational safety and maintenance costs. However, collected bearing vibration signals often contain significant noise, which impacts diagnostic accuracy. Additionally, a single FCD model struggles with the complexity of diversified and nonlinear data like these vibration signals. To address these issues, we propose a novel FCD method for offshore wind turbine bearings, called MIDCIGE, based on a multi-scale integrated dual-channel convolutional neural network (DCNN) and an improved grey wolf optimization (GWO)-optimized echo state network (ESN). MIDCIGE consists of three main stages: noise reduction, feature extraction, and model optimization. The noise reduction stage employs a method based on multi-indicator weighted complementary ensemble empirical mode decomposition with adaptive noise, utilizing kurtosis and margin factor as metrics for signal filtering and reconstruction. Secondly, a multi-scale sliding window sampling approach transforms the denoised data into grayscale images for input into a DCNN model. The model optimization stage combines improved GWO-optimized ESN, enhancing FCD model accuracy. Finally, on offshore wind turbine dataset, the integrated MIDCIGE model achieves 99.53% average accuracy, outperforming baseline models by 4.65%–18.97%.