Prototype-Based Heterogeneous Federated Learning for Blade Icing Detection in Wind Turbines with Class Imbalanced Data

Lele Qi, Xu Cheng, Fan Shi, Xiufeng Liu, Shengyong Chen

Introduction

Wind energy is a cornerstone of renewable power, but turbines in high-latitude or high-altitude regions face a major threat: blade icing. Icing reduces aerodynamic efficiency, increases mechanical stress, and can slash annual power generation by up to 30%. While data-driven detection using SCADA (Supervisory Control and Data Acquisition) systems is effective, it faces three massive hurdles: data privacy, geographical heterogeneity (Non-IID data), and extreme class imbalance where icing events are rare compared to normal operations.

Methodology

The researchers propose FedHPb (Federated Heterogeneous Prototype-based learning), a framework designed to overcome these challenges by deviating from standard Federated Learning (FL) in several key ways.

System Architecture and Feature Extraction

Each client (turbine) utilizes a local model combining Long Short-Term Memory (LSTM) to capture time-series dependencies and a Convolutional Neural Network (CNN) with Squeeze-and-Excitation (SE) blocks to extract spatial features. Instead of uploading massive model gradients or weights, clients send "prototypes"—the average feature vectors for the "icing" and "normal" classes.

Overview of the FedHPb framework, illustrating local client training with contrastive learning and prototype aggregation at the central server.

Contrastive Learning and Dynamic Weighting

The server aggregates these prototypes to create a global representation of icing, which is then sent back to clients. To fight class imbalance, the authors designed a Supervised Contrastive Loss ($L_c$) that pulls local prototypes closer to their corresponding global prototypes while pushing dissimilar class prototypes apart. The loss function incorporates dynamic weights that automatically adjust based on class frequency, forcing the model to prioritize rare icing samples.

Key Results

The study utilized real-world SCADA data from 20 wind turbines across two distinct wind farms in China. The performance was compared against five state-of-the-art FL frameworks and five classic class-imbalance algorithms across challenging imbalance ratios (20:1, 50:1, and 100:1).

  • Superior Accuracy: FedHPb outperformed the second-best method (BiFL) by an average of 19.64% in mF$\beta$ score and 5.73% in balanced accuracy (mBA).
  • Resilience to Imbalance: Even at the extreme 100:1 ratio, FedHPb maintained an mBA of 85.35%, whereas standard methods like FedAvg plummeted.
  • Communication Efficiency: The prototype-based approach significantly reduces bandwidth requirements.
Comparison of approximate model update sizes between BiFL and the proposed FedHPb method.

As shown above, the update size for FedHPb was only 1.18 MB, compared to 5.09 MB for the previous state-of-the-art (BiFL)—a nearly 77% reduction in data traffic.

Conclusion

FedHPb demonstrates that sharing mean embeddings (prototypes) is a highly effective strategy for federated learning in heterogeneous environments. It provides global alignment without the privacy risks of raw data or the massive bandwidth requirements of full model weights. Future work will explore integrating homomorphic encryption to further secure prototypes and optimizing data window sizes to refine detection for different icing types.

Summary updated

BibTeX
@misc{qi2025prototype,
  author = {Lele Qi and Xu Cheng and Fan Shi and Xiufeng Liu and Shengyong Chen},
  title = {{Prototype-Based Heterogeneous Federated Learning for Blade Icing Detection in Wind Turbines with Class Imbalanced Data}},
  year = {2025},
  eprint = {2503.08325},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2503.08325},
}
Figure