基于故障敏感度学习和集成学习的可诊断性设计方法

Diagnosability Design Method Based on Fault-sensitive Learning and Integrated Learning

对系统进行可诊断性设计,能优化诊断方案、提升诊断方案与系统的契合程度,对于准确检测或隔离系统故障具有重大意义。为此提出一种基于故障敏感度学习和集成学习的可诊断性设计方法。对于采集到的不同状态下的系统测试点信号进行特征提取,结合系统的可诊断性评价结果,提出特征贡献度量化算法,对信号中不同特征对故障诊断的贡献度进行评估,使用改进的D-S证据理论算法对不同信号的特征进行融合,确定适用于故障检测与隔离的故障敏感特征集合;采用集成学习方法对基分类器的诊断效果进行加强,最终获得对当前系统的诊断方案。仿真实验结果表明,采用新的可诊断性设计方法设计出来的针对系统的诊断方案,可以对系统故障进行良好的诊断,其中与未进行故障敏感度学习环节相比,故障诊断的错误率由5.33%下降到2.66%,与未进行集成学习环节相比,故障诊断的错误率由16.22%(基诊断器的平均值)下降到2.66%。在与其他诊断方案的对比实验中,新方法的故障诊断错误率相较于对比方法的平均值下降了3.34%。

The diagnosability design of system can optimize the diagnostic scheme and improve the the degree of compatibility between diagnostic scheme and system,which is of great significance for the accurate detection or isolation of system faults.For this,a diagnosability design method based on fault-sensitive learning and integrated learning is proposed.The featuresfrom the collected signals of system test points in different states are extracted.Combined with the results of diagnosability evaluation of the system,a quantitative feature contribution algorithm is proposed to evaluate the contribution degree of different features in the signal to fault diagnosis.The improved D-S evidence theory algorithm is used to fuse the features of different signals,which can determine the fault sensitive feature set suitable for fault detection and isolation.The integrated learning method is used to strengthen the diagnostic effect of base classifier,and finally the diagnosis scheme of the current system is obtained.Simulation experiments show that the system diagnosis scheme designed by the proposed diagnosability design method can make a good diagnosis of system faults.Compared with no fault-sensitive learning,the error rate of fault diagnosis decreases from 5.33%to 2.66%.Compared with no integrated learning,the error rate of fault diagnosis decreases from 16.22%(mean value of base diagnostics)to 2.66%.In the comparison experiments with other diagnostic schemes,the fault diagnosis error rate of the proposed method is decreased by 3.34%compared to the average fault diagnosis error rateof other methods.