泛化流形学习模式下风电机组传动系统早期故障预示方法研究

Based on generalization manifold learning theoretical framework, a novel incipient fault prediction method for wind turbine transmission system is proposed, in order to overcome the limitation which the typical fault diagnosis method does not suit to wind turbine transmission system incipient fault prediction, compensate the deficiencies of the existing manifold learning theoretical system, and solve key issues existed in the incipient fault prediction of wind turbine transmission system, such as weak feature extraction, decoupling, dynamic pattern recognition, feature trend continuous prediction and so on. The research stratagem of this proposed method is as follows. A novel framework theory of generalization manifold learning, which has weak feature extraction, decoupling, dynamic pattern recognition and feature trend continuous prediction together, will be proposed to form the new principle and the new methods for incipient fault prediction. A novel method of weak feature extraction based on main manifold expansion and low-dimensional manifold reconstruction will be proposed to fulfill incipient fault weak feature extraction. A novel method of dynamic pattern recognition based on multi-manifold space embedded in intelligent decision-making mechanism will be proposed to achieve incipient fault dynamic pattern recognition. A novel method of evolution trend prediction based on multi-manifold space combined with evolution prediction mechanism will be proposed to solve trend prediction of incipient fault across time scale. Finally, a multi-agent system prototype of wind turbine transmission system incipient fault prediction based on generalization manifold learning will be developed and system performance will be evaluated. The research on proposed incipient fault prediction has important theoretical and economic values not only for ensuring wind turbine steady and reliable operation, but also for enrichment and development of the incipient fault prediction theory and technology.

提出泛化流形学习模式下风电机组传动系统早期故障预示新方法，目的是克服典型故障诊断方法的局限性，弥补现有流形学习理论体系的不足，在泛化流形学习框架内解决风电机组传动系统早期故障微弱特征提取、解耦、动态模式识别及演变趋势预测等关键问题。该方法研究思路为：提出集微弱特征提取、解耦、动态模式识别及演变趋势预测等功能于一体的泛化流形学习新思想，形成早期故障预示的新理论和新方法；提出主流形拓展和低维流形重构的微弱特征提取新方法，实现早期故障微弱特征提取；提出多流形空间嵌入智能决策机制的动态模式识别新方法，实现早期故障动态模式识别；提出多流形空间融合演变预测机理的趋势预测新方法，解决早期故障跨时间尺度的趋势预测问题；最后研发基于泛化流形学习的风电机组传动系统早期故障预示系统原型，进行系统性能评估。该方法对保证现役风电机组的稳定可靠运行，丰富和发展机械早期故障预示理论和技术，具有重要的理论和经济价值。

风电机组传动系统早期故障特征具有微弱、耦合、时变性强、动态发展等特点，导致风电机组传动系统早期故障预示难度更大、要求更高。针对风电机组传动系统早期故障预示难题，项目按计划进行了泛化流形学习模式下风电机组传动系统早期故障预示方法研究，研究了泛化流形学习模式下风电机组传动系统早期故障微弱特征提取、动态模式识别及退化趋势预测等关键科学问题。.提出了集非线性降噪、微弱特征提取、动态特征提取、早期故障识别、退化趋势预测等功能于一体的泛化流形学习理论，建立了泛化流形学习的风电机组传动系统早期故障预示理论构架。研究了无监督流形学习的非线性降噪方法，解决风电机组传动系统非线性噪声干扰问题。研究了改进的核空间距离测度（IKDM）和基于DSmT证据理论的多准则融合敏感特征选择方法，增强高维多域特征的敏感特性。.根据风电机组传动系统早期故障是否有明确故障类别信息，研究了不同微弱特征提取方法：无明确故障类别信息时，提出了无监督流形学习的微弱特征提取方法，包括正交局部保持映射(OLPP)、正交邻域保持嵌入(ONPE)、线性局部切空间排列(LLTSA)、等距映射(ISOMAP)等算法；有明确故障类别信息时，提出了有监督流形学习的微弱特征提取方法，包括有监督线性局部切空间排列 (SLLTSA)、有监督扩展局部切空间排列 (SELTSA)等算法；只有部分故障类别信息时，提出了基于正交半监督局部Fisher判别分析(OSELF) 的微弱特征提取方法。针对风电机组传动系统不同早期故障可能分布于不同故障流形，提出了基于多故障流形的微弱特征提取方法。.提出了基于连通性或非连通性增殖ONPE流形学习的风电机组传动系统早期故障新增样本动态特征提取方法，解决早期故障样本连续新增的问题。提出了粒子群参数优化（EPSO）最小二乘支持向量机（LS-SVM）的风电机组传动系统早期故障识别方法，解决早期故障样本稀缺的问题。提出了无监督流形学习和最小二乘支持向量机的风电机组传动系统退化趋势预测方法。研发了一套基于B/S和C/S混合架构的风电机组传动系统网络化状态监测与早期故障预示系统，并进行测试应用。.项目研究达到了预期目标，研究成果对保证现役风电机组的稳定可靠运行，丰富和发展机械早期故障预示理论和技术，具有重要的理论和经济价值。