Wind turbine blade icing diagnosis using Gram Matrix representation and a multi-branch ConvNeXt network

Abstract

Wind turbine blade icing frequently occurs under harsh operating conditions, such as low temperatures, high altitudes, and high humidity, causing aerodynamic degradation, reduced power generation efficiency, and even serious safety risks. However, accurate identification of blade icing states under varying conditions remains a major challenge. To address this issue, this study proposes a high-accuracy blade icing diagnosis method based on Supervisory control and data acquisition (SCADA) data to improve both diagnostic performance and generalization capability. First, key features strongly related to blade icing are selected from multivariate SCADA time-series data by combining random forest feature importance with permutation importance, which enhances both the effectiveness and physical interpretability of the input variables. Next, one-dimensional time-series signals are transformed into two-dimensional Gramian matrix representations to better capture temporal correlation patterns. Based on these representations, a multi-branch ConvNeXt deep convolutional neural network is developed to independently extract features from Gramian matrix images of different variables, followed by high-level feature fusion for collaborative learning of multi-source information. Experimental results using real-world wind farm SCADA data show that the proposed method achieves an icing diagnosis accuracy of 99.44% and demonstrates strong generalization performance in cross-turbine experiments. These results confirm the effectiveness and robustness of the proposed method and indicate its potential as a reliable data-driven solution for blade icing detection and intelligent operation and maintenance of wind turbines in complex environments.