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

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

TL;DR

The paper introduces FedHPb, a novel federated learning framework designed to detect wind turbine blade icing while preserving data privacy. By exchanging abstract class prototypes instead of raw model parameters and incorporating a dynamically weighted supervised contrastive loss, FedHPb successfully overcomes severe data heterogeneity and extreme class imbalance across different wind farms.

What Problem Does This Paper Address?

Wind turbines operating in high-latitude or high-altitude regions face a significant threat from blade icing, which can reduce annual power generation by up to 30% and cause severe safety hazards. While data-driven methods using Supervisory Control and Data Acquisition (SCADA) systems can detect icing early, traditional centralized training requires aggregating sensitive operational data from various wind farms into a single server, raising major data privacy concerns.

Federated Learning (FL) solves the privacy issue by allowing local turbines to train models independently and share only model updates. However, standard FL fails in real-world wind turbine deployments due to two major roadblocks:

  1. Data Heterogeneity (Non-IID Data): Turbines are scattered across different geographical locations with varying environmental conditions and hardware configurations, causing conflicting gradient updates that hinder global model convergence.
  2. Extreme Class Imbalance: Turbines operate normally most of the time; icing events are rare anomalies. This causes standard models to heavily favor the majority "non-icing" class, frequently missing critical icing events.

Methods

To tackle these twin challenges, the authors propose FedHPb (Federated Heterogeneous Prototype-based learning for imbalanced data), which fundamentally changes how turbines communicate with the central server.

Fig. 1. Overview of FedHPb: In the initial stage, clients update their local prototype sets by reducing classification errors and minimizing the differences between local and global prototypes through supervised contrastive learning. Once the updates are completed, these prototype sets are sent to the central server. The central server creates a global prototype based on the received data and sends this global prototype back to all clients.

1. Client Feature Extractor (LCNN)

Each local turbine client utilizes a hybrid network called LCNN, which combines Long Short-Term Memory (LSTM) networks to capture time-series dependencies from SCADA data with a 3-layer Convolutional Neural Network (CNN). The CNN layers integrate a Squeeze-and-Excitation (SE) attention mechanism to refine and enhance critical features before compressing them into a flattened feature vector.

Fig. 2. Neural network architecture of the client model.

2. Federated Prototype Learning

Instead of transmitting massive model weights or gradients, clients compute and upload class prototypes—the average embedding vectors for the "icing" and "non-icing" classes. The central server aggregates these prototypes by calculating a weighted average based on dataset sizes. This allows the framework to easily accommodate heterogeneous local network architectures and eliminates gradient conflict issues.

3. Supervised Contrastive Loss ($L_c$) with Dynamic Weights

To combat class imbalance, the authors introduce a specialized local loss function: $$L_{all} = (1 - \lambda)L_s + \lambda L_c$$ Where $L_s$ is a balanced supervised learning loss, and $L_c$ is a supervised contrastive loss. $L_c$ applies dynamic weights ($w_{pos}$ and $w_{neg}$) derived from class counts to penalize misclassifications of the minority icing class more severely. This setup forces the model to pull similar class prototypes closer together while pushing dissimilar class boundaries apart in the embedding space.

Datasets

The framework was evaluated using real-world SCADA data from 20 wind turbines across two distinct wind farms in China:

  • Hongshimu Wind Farm (Shaanxi Province): 10 turbines, data spanning 15 days.
  • Yanling Wind Farm (Henan Province): 10 turbines, data spanning 16 days.

The dataset includes 16 expert-selected variables related to blade icing. To test robustness against class imbalance, the authors synthesized three training imbalance ratios (non-icing to icing samples): 20:1, 50:1, and 100:1, while maintaining a fixed 10:1 ratio in the testing set.

Key Insights and Ideas

  • Prototypes over Parameters: Traditional FL averages model weights, which can completely corrupt performance when local data distributions diverge wildly. Averaging abstract class embeddings (prototypes) bypasses this issue entirely, allowing the global server to find a shared representation space without destroying local model specialization.
  • Contrastive Learning for Imbalance: Rather than relying on traditional data-level sampling (like SMOTE or undersampling, which can cause information loss), FedHPb uses contrastive loss to reshape the embedding space itself. It actively forces the network to maximize the distance between normal and anomalous states, making the model highly sensitive to rare icing events even at a 100:1 imbalance ratio.

Key Results

FedHPb was compared against five state-of-the-art FL frameworks (including FedAvg, FedProto, and BiFL) and five class-imbalance algorithms.

  • Performance Boost: FedHPb significantly outperformed all competitors. Compared to the second-best method (BiFL), FedHPb achieved an average improvement of 19.64% in macro-$F_\beta$ ($mF_\beta$) and 5.73% in macro-Balanced Accuracy ($mBA$) across all imbalance ratios. Even at the harshest 100:1 imbalance ratio, it maintained a high $mF_\beta$ of 73.08% and an $mBA$ of 85.35%.
  • Communication Efficiency: Because prototypes are compact vectors compared to full neural network weights, FedHPb drastically slashes network overhead. The average data transmitted per client per round was only 1.18 MB, compared to 5.09 MB for BiFL—a reduction of over 76%.
Fig. 3. Approximate Model Update Size (MB)

Limitations and Future Work

While sharing prototypes instead of raw data provides excellent privacy, the authors acknowledge a residual data exposure risk: sophisticated inversion attacks could theoretically reconstruct aspects of the raw data from the shared feature embeddings.

To mitigate this, the authors state that future work will focus on integrating advanced privacy-preserving techniques, such as confusion matrix methods and homomorphic encryption, directly into the prototype aggregation phase.

Summary updated

Abstract

Wind farms, typically in high-latitude regions, face a high risk of blade icing. Traditional centralized training methods raise serious privacy concerns. To enhance data privacy in detecting wind turbine blade icing, traditional federated learning (FL) is employed. However, data heterogeneity, resulting from collections across wind farms in varying environmental conditions, impacts the model's optimization capabilities. Moreover, imbalances in wind turbine data lead to models that tend to favor recognizing majority classes, thus neglecting critical icing anomalies. To tackle these challenges, we propose a federated prototype learning model for class-imbalanced data in heterogeneous environments to detect wind turbine blade icing. We also propose a contrastive supervised loss function to address the class imbalance problem. Experiments on real data from 20 turbines across two wind farms show our method outperforms five FL models and five class imbalance methods, with an average improvement of 19.64\% in \( mF_β \) and 5.73\% in \( m \)BA compared to the second-best method, BiFL.
BibTeX
@misc{qi2025prototype,
  author = {Lele Qi and Mengna Liu 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