变工况下多源信息深度融合的风电传动系统运行状态评估研究

Problems caused by Wind Turbine (WT) working with failure are becoming an increasingly serious issue, especially drive train malfunction rate staying at a high level, which severely threat the security of WT. Considering the fact that a single type fault information under complex condition is difficult to accurately reflect drive train running states, the multi-view data deep fusion method is proposed, which focuses on several crucial issues including feature fusion, collected data annotation, remaining useful life prognosis and fault localization. The project includes: (1) The multi-view data fusion method based on attention mechanism is studied to solve complementary fusion problems of multi-view features with difference and relevance. (2) The semi-supervised transfer component analysis method is proposed to auto-annotate each single view data according to different conditions, and then construct multi-view features adaptive weighted fusion state evaluation model. (3) The degenerate analysis model and remaining useful life prognosis model are constructed based on the deep recurrent neural network combined with attention mechanism under unsteady operating conditions. (4) After modal filtering and feature extraction, sparse reconstruction theory is further adopted to study the multi-source location algorithm based on vector array frequency focusing and beam-forming methods.. The research achievement is expected to form an innovative method of fault diagnosis and operating conditions evaluation used in WT drive train, and provide theoretical and technical basis to establish the reliable condition monitoring and diagnosis system. It has important practical significance to improve the WT operating in safe and economic conditions.

风电机组运行故障问题日益突出，尤其是传动系统的故障率居高不下，给机组运行带来重大的安全隐患。针对复杂工况下单一类型故障信息难以全面准确反映传动系统运行状态问题，提出多源信息深度融合的评估方法，就特征融合、运行数据标注、剩余寿命预测和故障定位等问题展开研究，包括：（1）研究基于注意力机制的多源信息融合方法，解决不同视角特征的差异性和关联性互补融合问题。（2）基于半监督成分分析的迁移学习方法，解决区分工况的各视角数据自动标注问题；建立多视角特征自适应加权融合的状态评估模型。（3）变工况下深度循环神经网络联合注意力机制的退化分析与剩余寿命预测。（4）在稀疏重构的模态滤波和特征提取基础上，研究频域聚焦的矢量阵波束形成多源定位算法。.本项目研究成果有望形成具有创新性的风电传动系统运行状态评估方法，为构建可靠的风电机组健康管理体系提供理论和技术支持，对提高风电机组安全经济运行具有重要现实意义。

本项目针对复杂工况下单一类型故障信息难以全面准确反映风电传动系统运行状态问题，提出了多源信息深度融合的评估方法，就特征融合、运行数据标注、剩余寿命预测和故障定位等问题展开了深入研究，取得的主要成果包括：.（1）提出了复杂非平稳振动信号的带宽傅立叶分解方法，实现了强噪声环境下滚动轴承早期微弱故障特征的准确提取；建立了时变啮合刚度下齿轮箱故障响应的动力学分析模型，提出了共振基带宽傅立叶分解方法，实现了齿轮箱故障特征的高效提取。在上述成果基础上，采用随机森林和证据理论对多源多视角故障特征进行特征级和决策级融合，为传动系统运行状态的可靠评估奠定了基础。.（2）基于迁移学习方法解决了数据标注问题。提出了基于可迁移注意力的特征自适应强化方法和基于多源域分解的迁移诊断方法，实现了领域信息缺乏下的多源域故障信息迁移；提出了机理模型驱动的阶次谱迁移算法，实现了故障样本缺乏下的跨设备特征迁移，为解决工业场景中故障样本不足的共性问题提供了新思路。.（3）提出了基于长短时记忆神经网络（LSTM）的剩余寿命预测方法。构建了时域、频域、时频域、信息熵、分形维等长期趋势特征集用于表征设备退化状态，建立了基于单调性、鲁棒性、趋势性和可辨识性为准则的趋势特征优化筛选方法，大幅提高了传动系统退化分析和寿命预测建模的准确性。.（4）根据声发射（AE）阵列信号能量的稀疏性构建了空间定位模型，提出了基于频率自动聚焦的多重近场AE源定位算法和频率多快拍稀疏分解AE源定位算法，解决了故障源定位中的传感器阵列孔径约束问题。.上述研究成果为复杂工况下风电传动系统行状态评估提供了新方法，已应用于10MW级大型海上风电机组健康状态评估系统，对于提高风电机组安全运行具有重要价值和广泛的应用前景。.项目成果发表学术论文26篇，其中SCI检索13篇，EI检索10篇；申请发明专利25件，其中已授权3件；培养博士生8名，硕士生22名；获电力行业科技奖励3项（省部级）。