考虑样本多样性的转辙机故障诊断决策模型

Fault diagnosis decision model for switch machines considering sample diversity

为解决现有研究未充分考虑诊断模型在不同转辙机数据样本条件下的适用性,且单一诊断方法难以满足故障诊断精度与决策要求的问题,提出一种考虑样本多样性的故障诊断决策模型.首先,从ZYJ7电液转辙机8种故障模式和正常模式所对应的油压曲线中分别提取时域、频域、时频域特征量,采用基于核函数的主成分分析法对3个域内的特征量分别降维,得到每个域的特征矩阵,进而构成不同类型的数据样本.然后,基于PSO-KNN、SA-PSO-PNN、PSO-SVM算法构建决策模型.当样本是一般数据样本时,决策模型采用3种算法分别做同一域数据分类,并对同一域各算法诊断结果进行三取二表决,分别得到每一域诊断结果;当样本是大数据样本、不均衡数据样本时,决策模型根据不同样本特点采用3种算法中的推荐算法得到每一域诊断结果.最后,利用决策模型对各域诊断结果进行三取二表决得到最终诊断结果 .仿真结果表明:与单一诊断算法相比,决策模型在大数据样本下,平均准确率提升1.01%;在不均衡数据样本条件下,决策模型的平均准确率提升12.82%;在一般数据样本下,决策模型平均准确率提升6.18%.决策模型通过结合多域的多维特征与各算法特点提高了诊断精度,为集成学习在转辙机故障诊断领域应用提供了一种思路.

To address the issue that existing research does not adequately consider the diagnostic model’s applicability across various switch machine data sample conditions,and that a single diagnos-tic method struggles to meet fault diagnosis accuracy and decision-making requirements,this study proposes a fault diagnosis decision model that takes sample diversity into account.Firstly,time do-main,frequency domain,and time-frequency domain features are extracted from the oil pressure curves of eight fault modes and normal mode of the ZYJ7 electro-hydraulic switch machine.The Ker-nel Principal Component Analysis(KPCA)method is employed to dimensionally reduce the feature quantities in the three domains,resulting in the formation of feature matrices for each domain and thus different types of data samples.Then,the decision model is constructed using PSO-KNN,SA-PSO-PNN,and PSO-SVM algorithms.For general data samples,the model applies all three algorithms for data classification within the same domain.A two-out-of-three voting mechanism is then used to consolidate the diagnosis results from each algorithm within the same domain,yielding domain-specific diagnosis outcomes.For big data and unbalanced data samples,the model selects the recom-mended algorithm from the three based on sample characteristics to determine the diagnosis results for each domain.Finally,a final diagnosis is obtained by applying a two-out-of-three voting approach to the domain-specific diagnosis results.Simulation results demonstrate that the decision model achieves an average accuracy improvement of 1.01%for big data samples,12.82%for unbalanced data samples,and 6.18%for general data samples compared to single diagnosis algorithms.These im-provements highlight the model’s enhanced diagnostic precision through the integration of multidimen-sional features across multiple domains and algorithm-specific attributes,offering a novel approach for the application of ensemble learning in switch machine fault diagnosis.