近年来风能利用发展迅速，风电关键装备的部署区域更加广泛、运行环境更加极端，管理要求更加精细。智能运维技术对设备运行的安全性、可靠性和经济性具有重要意义，实际需求日益迫切。针对实际风电机组数据故障信息稀疏、数据质量低，现有研究面临多源异构数据融合困难等挑战。本文基于大量风电机组实际运行数据，采用以转换器网络为基础的深度学习、自监督学习理论方法，围绕风电机组数据清洗、异常预警、故障诊断展开研究。主要内容包括：（1）针对风电机组多源振动信号特征提取和有效融合问题，本文提出了一种基于多源振动信号图像化特征融合的故障诊断方法。通过将一维振动信号转化为二维频域格拉姆角场和马尔科夫转移场特征，利用视觉转换器网络聚合特征信息，实现多源振动信号的特征层融合。在实际风电机组和实验室故障数据集上的诊断结果表明，使用所提特征相比原始振动信号可以获得约8%的性能提升。采用融合特征多次实验结果的标准差显著降低，且在噪声条件下有更好的抗噪性能。（2）针对SCADA数据质量低、异常预警困难的问题，本文首先在小风速区间内进行功率排序实现异常数据聚合，基于贝叶斯变点估计定位异常区间，实现单变量的异常数据清洗；针对分布极端复杂的数据，基于风机知识和运行规则拟合正常状态超曲面，判断潜在异常点，并根据多次拟合中数据异常频率实现多变量的异常数据清洗。在222台风机SCADA数据上的清洗结果表明所提方法在大规模风机应用上有更强的鲁棒性。利用清洗后的数据，基于分块时序转换器网络提出了多源SCADA数据融合的异常预警方法，通过变量独立分块和网络权重共享，实现多尺度故障特征融合。在风电机组齿轮箱油温高故障案例中的应用结果表明，所提方法相比于对比方法提早25个小时预警，在维修后10天内误警率为0。（3）针对多源异构数据融合困难、故障典型样本不足、难以全面了解故障模式等问题，本文提出了基于半监督-自监督学习的振动-工况数据特征融合的已知故障诊断和未知故障识别方法。将视觉转换器网络作为自监督学习基本网络，利用少量标签信息进行对比学习，并添加基于马氏距离的未知故障识别模块，实现了少量标签样本下的已知故障诊断和未知故障识别。研究结果表明，所提方法相比于对比方法在有标签数据充足的条件下准确率提高了1.16%，在有标签数据稀少的条件下准确率提升了84.21%。

In recent years, wind energy utilization has been developing rapidly, with wind turbines deployed in more extensive and extreme environments, demanding more delicate management. Intelligent operation and maintenance (O&M) technology is of great significance to the safety, reliability, and economy of equipment operation. For actual wind turbine data, the scarcity of failure information and low data quality pose significant challenges, while existing research struggles with the difficulty of integrating heterogeneous data from multiple sources. Based on a substantial collection of operational data from wind turbines, this study leverages transformer-based deep learning and self-supervised learning to focus on data cleaning, anomaly early warning, and fault diagnosis methodologies. The main content includes:(1) To address the challenge of multi-source vibration signal feature extraction and fusion for wind turbines, this study proposes an image-based fault diagnosis method. By transforming one-dimensional vibration signals into two-dimensional frequency domain Gramian Angle Field and Markov Transition Field features, a Vision Transformer (ViT) network aggregates global feature information to achieve feature-level fusion. Finally, the classifier constructed by a multilayer perceptron (MLP) enables fault diagnosis for wind turbines by establishing a one-to-one mapping between features and specific faults. Experimental results on real-world and laboratory datasets show that the proposed features improve performance by 8% over the original signals. Multi-source fusion also reduces result variability and enhances noise resistance.(2) To address the challenges of low-quality SCADA data and anomaly early warning, this study proposes a univariate data cleaning method. By constructing a pseudo power series within small wind speed intervals, the location of anomaly data intervals can be identified based on Bayesian change-point estimation theory. To further enhance the performance, a multivariate data cleaning method is introduced by leveraging wind turbine knowledge and operational rules. This method fits multi-dimensional normal data to a hypersurface and statistically analyzes the frequency of abnormal samples based on their distance from the hypersurface, enabling multivariate data cleaning. The results on SCADA data from 222 wind turbines demonstrate that the proposed method exhibits higher robustness in large-scale wind turbine applications. Furthermore, by using quality-enhanced data, this study proposes a time-series transformer-based method for multi-source SCADA data fusion anomaly detection and early warning. The method independently patches variables and shares network weights to fuse multi-scale fault features. In real-world gearbox oil temperature anomaly cases, it achieves 25-hour-earlier warnings than conventional methods with zero false alarms within 10 days post-maintenance.(3) To address the challenges of multi-source heterogeneous data fusion, limited typical fault samples, and incomplete understanding of fault patterns, this paper proposes a known fault diagnosis and unknown fault identification method based on semi-supervised and self-supervised learning for vibration and operation condition data feature fusion. The method employs a ViT network as the backbone for self-supervised learning, utilizes limited labeled data for contrastive learning, and incorporates a Mahalanobis-distance-based unknown fault recognition module to achieve both known fault diagnosis and unknown fault identification. Experimental results demonstrate a 1.16% accuracy improvement under sufficient labeled data conditions and an 84.21% accuracy improvement under few labeled data conditions.