基于核主成分分析和RVM的传感器故障诊断算法设计  

Design of Fault Diagnosis Algorithm for Sensor Node Based on Kernel Principal Component Analysis and RVM

作　　者：郭秀峰[1] 任贺宇[2] 

机构地区：[1]河南牧业经济学院计算机应用系,郑州450000 [2]河南交通职业技术学院交通信息工程系,郑州450000

出　　处：《计算机测量与控制》2014年第3期709-712,共4页Computer Measurement &Control

摘　　要：传统的传感器故障诊断模型受限于所采用的机器学习方法需要人为设定参数,诊断精度依赖于参数设置的好坏,且无法实现传感器在线诊断,为此,提出了一种基于核主成分分析和稀疏贝叶斯RVM(relevancevector machine,RVM)的传感器在线故障诊断模型;首先,采用核主成分分析法将故障征兆数据映射到高维空间对数据进行降维,降低数据的复杂度;然后采用稀疏贝叶斯RVM对传感器进行故障诊断,在贝叶斯框架下对诊断函数权重进行推断,从而获得各故障类别的后验概率,量后,根据后验概率和投票致判断最终的故障类别;在NS2仿真环境下对实验进行仿真,结果表明,文中方法具有较高的故障诊断精度,较其它方法具有诊断时效高、泛化能力强和稀疏性好的优点,具有很强的可行性,Traditional Sensor node fault diagnosis model was limited to the machine learning method needing to set the parameters manu- ally, and the diagnosis accuracy was relied to the parameters, and can not realize the on--line diagnosis, therefore, a fault diagnosis model based on kernel principal component analysis and sparse Bayesian RVM was proposed to diagnose the sensor. Firstly, KPCA （Kernel princi- pal component Analysis） was used to map the sample data to the high dimension space to reduce the complexity of the data, then the sparse Bayesian RVM （relevance vector machine） was used to diagnose the sensor, the weight of diagnosis function was inferred in the Bayesian framework, and the probability of the fault was obtained. Finally, the probability and the votes were used to justify the final fault. The sim- ulation experiment is operated in the NS2 simulation environment shows the method in this paper can obtain the higher diagnosis accuracy, and compared with the other methods, it has the higher diagnosis accuracy, generalization ability and higher sparse, so it is proved feasible.

关 键 词：传感器 核主成分 故障诊断 相关向量机 稀疏贝叶斯 

分 类 号：TP31[自动化与计算机技术—计算机软件与理论]