[Paper Review] Intelligent Icing Detection Model of Wind Turbine Blades Based on SCADA data

Wenqian Jiang, Junyang Jin|arXiv (Cornell University)|Jan 20, 2021

Icing and De-icing Technologies37 references4 citations

TL;DR

This paper proposes PGANC and PGANT, two deep learning frameworks using two-stage training with parallel GANs and CNNs or domain adaptation to detect wind turbine blade icing from SCADA data. The models achieve superior accuracy and generalization, especially under limited labeled data, outperforming conventional CNNs and machine learning methods in AUC (up to 0.949) and MCC scores.

ABSTRACT

Diagnosis of ice accretion on wind turbine blades is all the time a hard nut to crack in condition monitoring of wind farms. Existing methods focus on mechanism analysis of icing process, deviation degree analysis of feature engineering. However, there have not been deep researches of neural networks applied in this field at present. Supervisory control and data acquisition (SCADA) makes it possible to train networks through continuously providing not only operation parameters and performance parameters of wind turbines but also environmental parameters and operation modes. This paper explores the possibility that using convolutional neural networks (CNNs), generative adversarial networks (GANs) and domain adaption learning to establish intelligent diagnosis frameworks under different training scenarios. Specifically, PGANC and PGANT are proposed for sufficient and non-sufficient target wind turbine labeled data, respectively. The basic idea is that we consider a two-stage training with parallel GANs, which are aimed at capturing intrinsic features for normal and icing samples, followed by classification CNN or domain adaption module in various training cases. Model validation on three wind turbine SCADA data shows that two-stage training can effectively improve the model performance. Besides, if there is no sufficient labeled data for a target turbine, which is an extremely common phenomenon in real industrial practices, the addition of domain adaption learning makes the trained model show better performance. Overall, our proposed intelligent diagnosis frameworks can achieve more accurate detection on the same wind turbine and more generalized capability on a new wind turbine, compared with other machine learning models and conventional CNNs.

Motivation & Objective

• To address the challenge of accurate and generalized icing diagnosis in wind turbines, especially with scarce labeled data.
• To leverage SCADA data for end-to-end learning of normal and icing state features without relying on physical models.
• To develop a transferable framework that generalizes across different wind turbines despite varying operating conditions.
• To improve detection performance by combining GAN-based feature learning with domain adaptation for zero-shot or few-shot transfer.
• To validate the model on real-world SCADA data from three wind turbines under diverse environmental and operational scenarios.

Proposed method

• Proposes PGANC: a two-stage framework using two parallel GANs to extract intrinsic features from normal and icing samples, followed by a CNN for classification.
• Introduces PGANT: extends PGANC by adding domain adaptation learning to improve performance when labeled data for the target turbine is insufficient.
• Employs two loss functions: MMD-based domain alignment loss (L_ms) and domain discrepancy loss (L_md) to align feature distributions across domains.
• Uses SCADA data including operational parameters (e.g., power, rotor speed) and environmental variables (e.g., temperature, wind speed) as input features.
• Applies a two-stage training process: first pre-train GANs to learn disentangled representations, then fine-tune the classifier or domain adaptation module.
• Utilizes evaluation metrics including AUC, F1-score, and MCC, with emphasis on minimizing false negatives in imbalanced data.

Experimental results

Research questions

• RQ1 Can a two-stage GAN-based framework effectively learn discriminative features for normal and icing states in wind turbine SCADA data?
• RQ2 How does the inclusion of domain adaptation improve model generalization when transferring to a new wind turbine with limited labeled data?
• RQ3 Does GAN-based feature learning outperform standard CNNs in capturing subtle patterns in imbalanced SCADA data for icing detection?
• RQ4 To what extent does the two-stage training strategy enhance model performance and training efficiency compared to end-to-end training?
• RQ5 How do the proposed models compare to traditional machine learning models (e.g., SVC, XGBoost) and other deep learning baselines in cross-turbine transfer learning?

Key findings

• PGANC achieved the highest AUC of 0.940 on turbine 21 and 0.933 on turbine 8, outperforming SVC, CNN, and other baselines in all three turbines.
• The two-stage training (stage 2) significantly improved performance over stage 1, with AUC gains of up to 0.056 on turbine 15 and 0.057 on turbine 8.
• PGANT achieved an AUC of 0.849 on turbine 15-21 and 0.800 on 21-8, outperforming KNN, SVC, XGBoost, and PGANC in cross-turbine transfer learning.
• The ablation study showed that PGANT with both MMD losses (L_ms and L_md) achieved the best MCC (0.769) and AUC, indicating improved minority-class prediction.
• Even without domain adaptation, PGANC showed better transferability than traditional models, proving the robustness of GAN-based feature learning.
• The model maintained high performance under 5% labeled data, with PGANT achieving an AUC of 0.949 on turbine 21-8, demonstrating strong generalization under data scarcity.