基于证据推理的分类决策故障诊断方法

发布时间：2018-12-15 03:22

【摘要】：作为Dempster-Shafer(DS)证据理论的一种重要分支,近年来被提出的证据推理(Evidence Reasoning,ER)规则及其融合方法中,定义了证据可靠性和重要性的概念并明确了两者的区别,这对于证据的获取及性能评价十分重要。此外,基于正交和定理给出的证据推理规则,相比于传统的Dempster证据组合规则,提供了一种更为严密的概率推理过程,其重新诠释了贝叶斯推理在辨识框架密集上的推广。本文开展基于证据推理的分类决策故障诊断方法研究,主要工作包括:(1)基于证据推理的电机转子系统故障诊断方法。首先,利用似然函数归一化的方法,从故障样本变化区间的投点结果中获得各故障特征的诊断证据;结合传感器本身固有误差以及各样本区间的诊断证据对各故障模式的诊断能力,计算诊断证据的可靠性因子;构建基于欧式距离度量的双目标优化模型,获得诊断证据的最优权重值;利用ER融合规则合并考虑了可靠性和权重的诊断证据,并根据融合结果进行诊断决策。最后,在电机转子系统故障诊断实验中,该方法表现出了较好的诊断性能。(2)基于证据推理的轨道高低不平顺故障诊断方法。由轨道高低不平顺引起的异常振动轻则引起行车质量下降,重则导致列车脱轨。利用有效的状态监测方法对轨道高低不平顺故障进行检测和诊断,是保证列车行车质量和行车稳定性的前提。为此,建立了一种基于ER规则的推理模型,通过融合车载传感器采集的加速度数据推理出轨道高低不平顺幅值的估计值。通过与经典神经网络方法在完备与不完备测量数据环境下的典型实验对比,验证了方法的有效性。(3)基于证据推理的广义分类器设计方法。经上述面向设备故障诊断应用的ER规则推理方法研究,可知故障诊断本质上是一种基于多源属性信息的分类决策问题。因此,进一步提出基于证据推理的广义分类器设计方法,以期将ER规则推广到解决一般意义上的分类问题。首先从某属性训练数据中获得证据;根据属性的分类能力确定属性及其证据的可靠性,并基于初始参数构建分类器;基于序列线性规划(SLP)训练分类器的参数;根据各属性证据融合结果进行分类决策。最后,选择5种国际上通用的基准数据集,将ER分类器与其它6种经典分类器方法进行对比试验,从而说明ER分类器的有效性和普适性。
[Abstract]:As an important branch of Dempster-Shafer (DS) evidence theory, in recent years, Evidence Reasoning,ER rules and their fusion methods have defined the concepts of reliability and importance of evidence and made clear the difference between them. This is important for obtaining evidence and evaluating performance. In addition, the rule of evidence reasoning based on orthogonal sum theorem provides a more rigorous process of probabilistic reasoning than the traditional Dempster rule of evidence combination, which reinterprets the generalization of Bayesian reasoning in the dense identification framework. In this paper, the classification decision fault diagnosis method based on evidence reasoning is studied. The main work includes: (1) the fault diagnosis method of motor rotor system based on evidence reasoning. Firstly, using the method of likelihood function normalization, the diagnosis evidence of each fault feature is obtained from the result of the fault sample variation interval. The reliability factor of diagnosis evidence is calculated by combining the inherent error of sensor itself and the ability of diagnosis evidence of each sample interval to diagnose each fault mode. A two-objective optimization model based on Euclidean distance metric is constructed to obtain the optimal weight of diagnostic evidence. The ER fusion rule is used to merge the diagnostic evidence considering reliability and weight and then the diagnosis decision is made according to the fusion results. Finally, in the fault diagnosis experiment of motor rotor system, this method shows good diagnosis performance. (2) the track irregularity fault diagnosis method based on evidential reasoning. The abnormal vibration caused by the irregularity of the track leads to the decrease of the driving quality and the derailment of the train. It is a prerequisite to ensure the quality and stability of train operation to detect and diagnose track irregularity by using effective state monitoring method. For this reason, a reasoning model based on ER rule is established, and the estimated value of track irregularity amplitude is deduced by combining acceleration data collected by vehicle sensor. The effectiveness of the proposed method is verified by comparison with the classical neural network method in complete and incomplete measurement data environments. (3) the design method of generalized classifier based on evidential reasoning. Based on the research of ER rule reasoning method for equipment fault diagnosis, it can be concluded that fault diagnosis is essentially a classification decision problem based on multi-source attribute information. Therefore, a design method of generalized classifier based on evidential reasoning is proposed in order to extend the ER rule to solve the general classification problem. Firstly, the evidence is obtained from some attribute training data, the attribute and the reliability of the evidence are determined according to the classification ability of the attribute, and the classifier is constructed based on the initial parameters, and the parameters of the classifier are trained based on the sequential linear programming (SLP). The classification decision is made according to the fusion results of each attribute evidence. Finally, five kinds of international datum data sets are selected, and the ER classifier is compared with other six classical classifier methods to demonstrate the effectiveness and universality of the ER classifier.
【学位授予单位】：杭州电子科技大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP202;TP277

【参考文献】