改进SpotTune迁移学习的风电机组状态监测

doi:10.19595/j.cnki.1000-6753.tces.260878

摘要
图/表
参考文献
相关文章 (15)

全文: PDF (2026 KB)
输出: BibTeX | EndNote (RIS)

摘要

时空图模型能够有效提取风电机组SCADA时序数据的时空特征,在状态监测中具有较高可靠性,但其性能依赖于大量训练数据。针对目标机组数据不足而难以构建可靠监测模型的问题,引入迁移学习利用源域机组中大量历史数据辅助目标机组建模,同时提出一种改进的SpotTune微调策略用于有效应对传统微调策略的适应性不足问题。首先,构建GAT-LSTM时空图模型,对风电机组关键部件SCADA变量进行图结构建模与时序特征提取;其次,在迁移阶段构建结构一致的双分支模块,其中一支保留源域模型参数并保持冻结,另一支根据目标域数据进行参数更新,输入数据同时经过两条分支进行特征提取,其输出由策略网络自适应生成融合权重并加权融合,从而在保留源域判别能力的同时增强模型对目标域数据的适应能力。在目标域数据不足与跨域分布差异场景下,改进的SpotTune迁移方法能够协助目标域机组模型形成故障传递链,具有更高的精度与可解释性。相比于无迁移方法的RMSE_all降低了74.7%,相比于顶层微调方法降低了60.3%。

	服务

	把本文推荐给朋友
	加入我的书架
	加入引用管理器
	E-mail Alert
	RSS
	作者相关文章
	裴植
	朱晶鑫

关键词 ：风电机组, 状态监测, 时空图模型, SpotTune, 迁移学习

Abstract：

Wind turbine maintenance costs account for a considerable proportion of total expenditure in the wind energy industry. To reduce economic losses caused by unexpected failures and downtime, real-time condition monitoring is essential. Spatio-temporal graph models can effectively extract temporal features from sensor data and spatial correlations among turbine components, but their performance usually relies on abundant training data. Newly commissioned turbines, however, often lack sufficient operating data. To address this issue, transfer learning is introduced to leverage well-labeled healthy data from source-domain turbines for target-domain modeling. Since conventional fine-tuning usually adjusts only the final layer and cannot adequately adapt to distribution differences between source and target domains, this paper proposes an improved SpotTune strategy. It preserves the discriminative knowledge learned from the source domain while improving adaptability to target-domain operating characteristics.
First, time-series samples are constructed from turbine operating data through data cleaning, normalization, and sliding-window segmentation. A graph is then built according to the physical connections among turbine components to embed prior knowledge and help the graph neural network learn spatial relationships among sensor nodes. Next, a source-domain spatio-temporal graph model is trained using healthy operating data, and its parameters are saved for transfer. These parameters are transferred to the frozen and fine-tunable modules of the target-domain model. The frozen module retains source-domain knowledge, whereas the fine-tunable module is adaptively updated using limited target-domain data. Meanwhile, a policy network generates layer-wise decision weights for the two modules, dynamically balancing knowledge retention and target-domain adaptation. Finally, the fine-tuned model performs node-level online condition monitoring to identify sensor deviations and support fault propagation analysis.
The method is validated on both a generated dataset and a public dataset. The main conclusions are as follows. (1) The amount of target-domain training data significantly affects monitoring performance. As training data decrease, MSE_all, RMSE_all, and MAE_all increase, indicating lower prediction accuracy and reliability. Under data-insufficient conditions, the model also struggles to form complete fault propagation chains, weakening its ability to determine fault range, severity, and source. (2) The improved SpotTune method effectively alleviates insufficient target-domain data and cross-unit distribution differences. It provides fault warnings 13 h in advance on the generated dataset and detects abnormal conditions at an early stage on the public dataset. Compared with the non-transfer and top-layer fine-tuning methods, it achieves higher prediction accuracy in both single-source and dual-source scenarios. On the public dataset, it reduces target-domain validation MSE_all, RMSE_all, and MAE_all by 92.9%, 74.7%, and 78.4%, respectively, relative to the non-transfer method, and by 82.4%, 60.3%, and 65.7% relative to top-layer fine-tuning. It also supports more reliable formation of fault propagation chains, improving the interpretability of early-warning results. (3) Compared with top-layer, partial-layer, and full-layer fine-tuning, the improved SpotTune method achieves higher prediction accuracy and demonstrates effectiveness and superiority in cross-unit transfer prediction. (4) Compared with representative transfer learning methods, its RMSE_all is reduced by 44.1% and 30.9% relative to Deep CORAL and DANN on the generated dataset, and by 16.2% and 16.4%, respectively, on the public dataset. These results confirm that the proposed method effectively improves the predictive performance of target-domain models.

Key words： Wind turbines status monitoring spatiotemporal graph model SpotTune transfer learning

收稿日期: 2026-06-03

PACS:	TM315
	TH17

基金资助:

国家自然科学基金资助项目（72271222,W2411062,72501261）

通讯作者: 裴植男,1982年生,教授,博士生导师,研究方向为智能制造,服务型制造,智慧物流系统等。E-mail：peizhi82@163.com

作者简介: 朱晶鑫男,2001年生,硕士研究生,研究方向风电机组状态监测。E-mail：2465651323@qq.com

引用本文:

裴植, 朱晶鑫. 改进SpotTune迁移学习的风电机组状态监测[J]. 电工技术学报, 0, (): 260878-. Pei Zhi, Zhu Jingxin. Improving SpotTune Transfer Learning for Wind Turbine State Monitoring. Transactions of China Electrotechnical Society, 0, (): 260878-.

链接本文:

https://dgjsxb.ces-transaction.com/CN/10.19595/j.cnki.1000-6753.tces.260878 https://dgjsxb.ces-transaction.com/CN/Y0/V/I/260878

[1] 尹彦博, 许伯强, 吴京隆, 等. 考虑潜变量影响的风电机组双馈感应发电机故障检测方法[J]. 电工技术学报, 2026, 41(2): 527-540.
Yin Yanbo, Xu Boqiang, Wu Jinglong, et al.Fault detection method for doubly fed induction generator of wind turbine considering the influence of latent variables[J]. Transactions of China Electrotechnical Society, 2026, 41(2): 527-540.
[2] Li Z, Wen B, Peng Z, et al.Dynamic modeling and analysis of wind turbine drivetrain considering the effects of non-torque loads[J]. Applied Mathematical Modelling, 2020, 83: 146-168.
[3] Dangi N, Sodja J, Ferreira C S, et al.The effect of turbulent coherent structures in atmospheric flow on wind turbine loads[J]. Renewable Energy, 2025, 241: 122248.
[4] 陈俊生, 李剑, 陈伟根, 等. 采用滑动窗口及多重加噪比堆栈降噪自编码的风电机组状态异常检测方法[J]. 电工技术学报, 2020, 35(2): 346-358.
Chen Junsheng, Li Jian, Chen Weigen, et al.A method for detecting abnormal states of wind turbines using sliding window and multi noise ratio stack denoising autoencoder[J]. Transactions of China Electrotechnical Society, 2020, 35(2): 346-358.
[5] 许伯强, 王彪, 孙丽玲, 等. 基于改进集合经验模态分解和强化视觉 Transformer 模型的风电机组故障预警[J]. 电工技术学报, 2025, 40(20): 6537-6551.
Xu Boqiang, Wang Biao, Sun Liling, et al.Wind turbine fault warning based on improved ensemble empirical mode decomposition and enhanced visual Transformer model[J]. Transactions of China Electrotechnical Society, 2025, 40(20): 6537-6551.
[6] 马良玉, 程善珍. 基于支持向量数据描述和XGBoost的风电机组异常工况预警研究[J]. 电工技术学报, 2022, 37(13): 3241-3249.
Ma Liangyu, Cheng Shanzhen Research on abnormal operating condition warning of wind turbine based on support vector data description and XGBoost[J]. Transactions of China Electrotechnical Society, 2022, 37(13): 3241-3249.
[7] Han H, Yang D.Correlation analysis based relevant variable selection for wind turbine condition monitoring and fault diagnosis[J]. Sustainable Energy Technologies and Assessments, 2023, 60: 103439.
[8] Constantino-Robles C D, Castillo Leonardo F A, Galván J H, et al. Vibration-Based Predictive Maintenance for Wind Turbines: A PRISMA-Guided Systematic Review on Methods, Applications, and Remaining Useful Life Prediction[J]. Applied Mechanics, 2026, 7(1): 11.
[9] Tao H, Zhong Y, Yang G, et al.An online digital imaging excitation sensor for wind turbine gearbox wear condition monitoring based on adaptive deep learning method[J]. Sensors, 2024, 24(8): 2481.
[10] Zhao D, Shao D, Cui L.CTNet: A data-driven time-frequency technique for wind turbines fault diagnosis under time-varying speeds[J]. ISA Transactions, 2024, 154: 335-351.
[11] Li S, Li X, Jiang Y, et al.A novel frequency-domain physics-informed neural network for accurate prediction of 3D Spatio-temporal wind fields in wind turbine applications[J]. Applied Energy, 2025, 386: 125526.
[12] Xie T, Xu Q, Jiang C, et al.The fault frequency priors fusion deep learning framework with application to fault diagnosis of offshore wind turbines[J]. Renewable Energy, 2023, 202: 143-153.
[13] Kong Z, Tang B, Deng L, et al.Condition monitoring of wind turbines based on spatio-temporal fusion of SCADA data by convolutional neural networks and gated recurrent units[J]. Renewable Energy, 2020, 146: 760-768.
[14] Ma J, Fu Y, Cheng T, et al.SCADA data-driven spatio-temporal graph convolutional neural network for wind turbine fault diagnosis[J]. IEEE Transactions on Instrumentation and Measurement, 2025.
[15] Zhu Y, Xie B, Wang A, et al.Wind turbine fault detection and identification via self-attention-based dynamic graph representation learning and variable-level normalizing flow[J]. Reliability Engineering & System Safety, 2025, 253: 110554.
[16] Misbah I, Lee C K M, Keung K L. Fault diagnosis in rotating machines based on transfer learning: Literature review[J]. Knowledge-Based Systems, 2024, 283: 111158.
[17] Yao S, Kang Q, Zhou M C, et al.A survey of transfer learning for machinery diagnostics and prognostics[J]. Artificial Intelligence Review, 2023, 56(4).
[18] Tang S, Ma J, Yan Z, et al.Deep transfer learning strategy in intelligent fault diagnosis of rotating machinery[J]. Engineering Applications of Artificial Intelligence, 2024, 134: 108678.
[19] Dai B, Frusque G, Li T, et al.Smart filter aided domain adversarial neural network for fault diagnosis in noisy industrial scenarios[J]. Engineering Applications of Artificial Intelligence, 2023, 126: 107202.
[20] 何西花, 魏喆, 陈墨, 等. 基于变尺度卷积注意力的多任务域适应轴承故障诊断[J/OL]. 机械设计, 1-14[2026-06-21].
He Xihua, Wei Zhe, Chen Mo, et al.Bearing fault diagnosis based on multi-task domain adaptation with variable-scale convolutional attention[J/OL]. Journal of Machine Design, 1-14[2026-06-21].
[21] Zhang B, Dong H, Qaid M A A H, et al. Deep domain adaptation with correlation alignment and supervised contrastive learning for intelligent fault diagnosis in bearings and gears of rotating machinery[J]. Actuators, 2024, 13(3): 93-.
[22] Tan C, Sun F, Kong T, et al.A survey on deep transfer learning[C]//International conference on artificial neural networks. Cham: Springer International Publishing, 2018: 270-279.
[23] Alfano P D, Pastore V P, Rosasco L, et al.Top-tuning: A study on transfer learning for an efficient alternative to fine tuning for image classification with fast kernel methods[J]. Image and Vision Computing, 2024, 142: 104894.
[24] Guo Y, Shi H, Kumar A, et al.SpotTune: transfer learning through adaptive fine-tuning[C]//Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2019: 4805-4814.
[25] Güuck C, Roelofs C M A, Faulstich S. CARE to Compare: A real-world dataset for anomaly detection in wind turbine data[J]. arXiv preprint, 2024.
[26] Jin X, Lv S, Kong Z, et al.Graph spatio-temporal networks for condition monitoring of wind turbine[J]. IEEE Transactions on Sustainable Energy, 2024, 15(4): 2276-2286.
[27] Rosenblatt M, Tejavibulya L, Jiang R, et al.Data leakage inflates prediction performance in connectome-based machine learning models[J]. Nature communications, 2024, 15(1): 1829.
[28] Chen X, Yang R, Xue Y, et al.Deep transfer learning for bearing fault diagnosis: A systematic review since 2016[J]. IEEE Transactions on Instrumentation and Measurement, 2023, 72: 1-21.
[29] Zhuang F, Qi Z, Duan K, et al.A comprehensive survey on transfer learning[J]. Proceedings of the IEEE, 2020, 109(1): 43-76.
[30] Xiang S, Qu K, Tian Y, et al.Short-term wind power forecasting with small-sample datasets using an attention-enhanced domain-adversarial neural network[J]. Neurocomputing, 2025: 131308.