基于信息融合的机械故障概率盒诊断模型和算法研究

There are two difficult problems existed in traditional mechanical fault diagnosis based on information fusion. One is the information loss from feature extraction, and another is the difficulty of space-time registration and heterogeneous integration. The probability box (p-box) theory has a strong advantage in dealing with uncertainties. Its application has been confirmed in the fields of economic, ecological, nuclear physics, etc. However, researches in the field of fault diagnosis are rarely reported. This project proposes a fault diagnosis method based on data fusion and p-box theory to solve the above problems. Typical parts such as rolling bearings, gears are used as concrete research objects. The p-boxes which can contain the original information are used to solve the information loss from feature extraction. The fault p-box models are established which are composed by original data p-box models and feature p-box models. The p-boxes models are corrected to improve its compactness. The Bayesian and evidence theory fusion methods based on p-boxes as fusion objects are proposed to solve the difficulty of space-time registration. Fusion algorithms are improved to reduce the overlaps among models. The Bayes decision, support vector machine fault diagnosis methods of p-boxes optimized by genetic algorithm, particle swarm algorithm are proposed to improve the correct recognition rate. This project will provide theoretical exploration, guidance and basis for the crossover studys of p-box theory and fault diagnosis field, and further enrich the methods of machinery fault diagnosis.

传统基于信息融合的机械故障诊断存在特征提取带来的信息丢失问题和时空配准及异类融合难点问题。概率盒理论在处理不确定性问题上存在强大优势，其应用已在经济、生态、核物理等领域得到证实，但在故障诊断领域研究报道鲜见，本项目拟提出基于概率盒理论的信息融合故障诊断方法以解决上述难点问题。以滚动轴承、齿轮等典型零部件为具体研究对象。利用概率盒包含原始信息解决特征提取带来的信息丢失问题。建立以原始数据概率盒为主分量、特征概率盒为次分量的故障概率盒模型，提出模型修正方法，以改善其紧致性。提出以概率盒为融合对象的多贝叶斯、证据理论等融合方法，以解决时空配准难点问题，并改进融合算法，减小模式间重叠程度。以提高正确识别率为目标，提出基于遗传算法、粒子群算法优化的Bayes决策、支持向量机等的概率盒故障诊断方法。本项目将为概率盒理论与故障诊断领域的交叉研究探索提供理论指导和依据，并进一步丰富机械故障诊断方法。

传统基于信息融合的机械故障诊断在融合前都需要对原始信号进行特征提取，这将导致原始信号丰富信息的丢失，并且很难有效的解决多源信息的时空配准问题，多数研究都局限于同类信息。概率盒建模能将诸多主观和客观的不确定因素考虑进去，弥补传统特征提取方法丢弃丰富概率统计信息的缺陷。本项目利用概率盒理论在处理不确定性问题上的强大优势来实现基于信息融合的机械故障诊断。. 以滚动轴承、齿轮等典型零部件为具体研究对象。针对故障信号无确定概率分布类型的特点，分别提出了故障特征概率盒建模和原始数据直接建模两种改进算法，以解决特征提取带来的信息丢失问题。建立以原始数据概率盒为主分量、特征概率盒为次分量的故障概率盒模型。提出模型修正方法，改善概率盒模型的紧致性。提出以概率盒为融合对象的多贝叶斯、证据理论等融合方法，以解决时空配准难点问题，并改进融合算法，减小模式间重叠程度。以提高正确识别率为目标，提出基于遗传算法、粒子群算法优化的Bayes 决策、支持向量机等的概率盒故障诊断方法。. 以SVM为模式识别工具，对比传统方法与概率盒建模方法，得到的总正确识别率分别为83.45%和93.33%。修正后的概率盒模型的总正确识别率提升至95.5%。以修正的概率盒为融合对象，采用改进的融合算法，总正确识别率提升至为99.5%。结果表明修正并改进融合后的概率盒模型的总正确识别率明显高于传统方法。. 本项目将为概率盒理论与故障诊断领域的交叉研究探索提供理论指导和依据，并进一步丰富机械故障诊断方法。