面向不平衡样本的流形学习故障诊断方法

The serious imbalance characteristic of sample data leads to great troubles for fault feature extraction which is one of the key issues of fault diagnosis in the complex equipment and production process. The fault diagnosis for imbalanced samples is a scientific problem which has important theoretical significance and high engineering practical value. The project is to study the semi-supervised and manifold learning method for the complex nonlinear fault feature extraction problem as the core of the theories and methods for fault diagnosis, which is on the basis of an in-depth analysis for the four performances of imbalanced data characteristics: not complete, strong noise, the relative redundancy and samples unequal in a typical industrial system as an object. The research content includes: the semi-supervised data reconstruction methods for imbalance samples; the modeling technology based on DAPK manifold learning; the feature extraction method based on cost-sensitive wavelet network. The research is distinguished as: 1) the data reconstruction method based on resample and semi-supervised learning is proposed. The method reduces the distortion influence to manifold structure, induced by the sparse fault samples. The difficulty, caused by data imbalance for feature extraction, can be solved. 2) the manifold learning method based on priori oriented local optimization and overall arrangement is proposed. The method solves the fault feature overlap problem by mining structural information and geometric rule in samples data and improves the early warning capability and accuracy of fault diagnosis.

复杂设备及生产过程的样本数据呈现的严重不平衡特性对故障诊断的关键环节-故障特征的提取造成了很大的困难，面向不平衡样本的故障诊断研究是一个具有重要理论意义和很高工程实用价值的科学问题。项目以典型工业系统为对象，在深入分析数据不平衡特性的不完整、强噪声、相对冗余和样本不均四种表现的基础上，针对复杂非线性故障的特征提取问题，研究以半监督和流形学习方法为核心的故障诊断理论与方法。研究内容包括：面向不平衡样本的半监督数据重构方法；基于DAPK流形学习建模技术；基于代价敏感小波网络的故障特征提取方法。研究的显著特点是：1)提出了基于重采样和半监督学习的数据重构方法，减轻了因故障样本稀少对流形结构产生的扭曲影响，解决了数据不平衡对特征提取带来的困难；2)提出了具有先验导向的局部优化和整体排列流形学习方法挖掘样本数据蕴含的结构信息和几何规律，解决了故障特征重叠问题，提高了故障诊断的早期预报能力和准确率。

本项目从研究分析复杂工业过程工况数据特性入手，总结了大部分工业系统中工况数据样本的三个不平衡特性，即样本标记不平衡、样本信息不平衡和样本类别不平衡，提出了面向具有数据样本不平衡特性的工业过程故障诊断的特征提取、数据重构和诊断决策方法，设计了实时监测系统实现了故障的及时、准确诊断，并进一步揭示了基于静态数据样本的故障诊断方法未能有效利用系统变量动态关系这一局限性，提出了新的科学问题，指引了后续的研究方向。项目以基于半监督学习方法、基于流形学习方法和基于敏感代价学习方法的特征提取算法为核心研究内容，以全局提取和局部保持思想为基础建立了面向不平衡数据样本的特征提取机制，提高了数据样本的利用率，克服了工况数据冗余造成的“维数灾难”问题，解决了目前数据驱动诊断方法对小样本学习建模不准确以及对噪声敏感影响诊断准确率的弱点，适合于工程实际应用。研究成果的显著特点是：引入基于全局和局部的流形学习算法准确计算了数据中心点实现了冗余信息的有效剔除；解决了带有不平衡特性的线性不可分数据特征信息有效提取困难的难题；优化平衡了诊断方法性能与计算复杂度之间的关系。研究成果将简捷地实现复杂工业过程工况数据的特征提取与故障决策，为生产过程的高效稳定运行提供了一种基础理论和关键技术方法。