Blade icing represents a substantial impediment to the efficacy of wind turbines, particularly in high-latitude regions where wind farms are commonly located. This phenomenon diminishes turbine performance, curtails energy production, and escalates maintenance expenditures. Data-centric methodologies offer a promising avenue for real-time icing detection; however, their success hinges on access to extensive and varied datasets, which are frequently scarce in remote wind farm deployments. This research introduces a novel heterogeneous federated learning (FL) paradigm specifically designed for blade icing detection. This paradigm addresses critical challenges encompassing data privacy, storage constraints, and computational demands. In contrast to conventional FL approaches that mandate uniform model architectures across both server and clients, our proposed model leverages disparate architectures to achieve enhanced learning flexibility and efficiency. Furthermore, this study confronts the prevalent issue of class imbalance within blade icing datasets by integrating data balancing techniques during training. Rigorous experimentation utilizing real-world data from 20 turbines across two distinct wind farms substantiates the superiority of our proposed model over existing FL methodologies. The model demonstrates enhanced class-imbalance handling, thereby presenting a robust solution for improving blade icing detection, optimizing operational effectiveness, and minimizing maintenance costs within wind farm operations.