【作者】 申争光；

【导师】 王祁；

【作者基本信息】 哈尔滨工业大学 ， 仪器科学与技术， 2013， 博士

【摘要】 自确认多功能传感器是一种不但能同时检测多个被测量，而且可在线确认自身工作状态的新型传感器，主要功能是：可对传感器故障进行检测和隔离，并用最佳估计值代替错误输出值实现数据恢复；可在线输出确认不确定度（Validated Uncertainty，VU）来指示确认测量值的准确度范围；可对多功能传感器健康评估。本课题得到国家自然科学基金资助，旨在研究各种状态自确认方法，解决自确认多功能传感器的若干关键技术问题。论文的主要研究内容如下：为验证研究的自确认多功能传感器关键技术的可行性，设计一种自确认多功能传感器实验系统，对多功能传感器进行标定和测试，在分析各敏感单元的失效机理及故障模式基础上，设计故障仿真和故障叠加电路来模拟产生各种真实故障，进而论证多故障时各种状态自确认方法的有效性。针对自确认多功能传感器的多故障检测及数据恢复问题，研究一种基于主元分析—小波相关向量机的数据自确认方法，将传统的单一故障扩展到多个故障。利用主元分析方法分析多敏感单元间的内在关系，研究测试样本在残差空间内投影量的变化次数与多故障检测的关系，以及该模型的故障检测能力。通过比较相关向量机（Relevance Vector Machine，RVM）在径向基、MexicanHat及Morlet小波等不同类型核函数下的预测性能，选取基于小波平移不变核函数的小波相关向量机预测器（Wavelet RVM，WRVM）来提高建模速度和抗噪性能，并利用WRVM模型进行多故障的在线隔离及其数据恢复，与神经网络方法相比，该方法在小样本条件下显著提高了数据恢复的精度和实时性。针对自确认多功能传感器的信号重构及其VU的在线评定问题，研究一种基于多变量相关向量机（Multivariate RVM，MVRVM）和确认的随机模糊变量（Validated Random Fuzzy Variables，VRFV）的确认测量值及其VU计算方法。在多功能传感器信号重构中的小样本和非线性条件下，MVRVM方法具有泛化性能好、稀疏性强和单模型多输出等优点，为此利用该方法进行多个被测物理量的确认测量值计算，与复合式RVM相比，提高了状态自确认效率。为进一步获取确认测量值的不确定度，在分析联合的数学变量VRFV的α cuts与不确定度的关系基础上，研究不同类型的故障对确认测量值的差异性影响，并针对传感器正常工作和故障两种情况，提出一种基于VRFV的VU评定方法。实验结果表明，该方法适用于自确认多功能传感器的在线VU评定，并利用传统的GUM方法对其有效性进行论证。针对自确认多功能传感器的健康评估问题，提出一种定量的基于健康可信度的健康评估方法。该方法将传统自确认传感器对单敏感单元测量值状态的评价，进一步扩展到对多功能传感器健康状况的综合评估，并分析多个敏感单元的相关性对传感器健康水平的影响。从局部的单敏感单元和整体的多功能传感器两个层面，本文重点分析和研究健康可信度的概念及其方法论原理，以量化形式直观表示其健康水平。实验结果表明，该方法能够如实反映不同健康水平的多功能传感器性能变化。更多还原

【Abstract】 The self-validating multifunctional sensor is a novel sensor, which can detectmultiple measurands as well as validate its own status. Its main functions are asfollows: detect and isolate multiple faults of sensors; replace the incorrectmeasurements by an optional estimated value to implement the data recovery underfaults; indicate the accurate range of the validated mearurements value by means ofon-line validated uncertainty (VU); evaluate the health levels of the multifunctionalsensors. Our research is supported by National Natural Science Foundation of China,and it centers on the status self-validation approaches to solve several key issues ofself-validating multifunctional sensors. Major work has been done as follows:To verify the feasibility of the status self-validation methods, a self-validatingmultifunctional sensor experimental system is designed. The faults mechanism andmodes of all the sensitive units are analyzed; the faults simulation and additioncircuits are designed to produce and simulate real failures. The status self-validationfunctions are then tested and demonstrated under multiple faults.Aiming at the on-line multi-fault detection and recovery of self-validatingmultifunctional sensors, the principal component analysis (PCA) coupled withwavelet relevance vector machine (WRVM) strategy is studied to implement thedata validation, which has been expanded from traditional single fault to multipleones. The inner relationship among sensitive units is analyzed deeply by using PCAtheory, and then multi-faults detection issue as well as the PCA-based faultsdetection ability is stuied by means of monitoring the projection changes of testsamples in residual space. Compared with the predictive performance of relevancevector machine (RVM) under different kernel types such as radical basis function,Mexican Hat, and Morlet wavelet, the Mexican Hat-based WRVM predictor isselected to improve the generalization ability, training speed and performance ofsuppressing noise. The WRVM is emplyed to implement on-line multiple faultsisolation and recovery. Compared with neural network methods, WRVM canachieve higher accuracy and fewer burdens, and it is suitable for data recoveryunder small sample.Aiming at the signal reconstruction and on-line VU estimation of self-validating multifunctional sensors, the multivariate RVM (MVRVM) and validatedrandom fuzzy variables (VRFV) are propsoed to compute the valiated measurmentvalues and VU respectively. Under the small sample and non-linear problem ofmultifunctional sensor reconstruction, the status self-validation efficiency is stuied by using the MVRVM-based reconstruction mode, which has the goodgeneralization ability and can output multiple values with one model simultaneously.Compared with the compound multi-RVM model, MVRVM-based reconstructiontechnology has a30%reduction in computation burden. To solve the on-lineuncertainty estimation, the relationship between α cutsof the proposed VRFVand measurement uncertainty is deeply analyzed and the negative effects fromdifferent faults types are also studied. The VRFV-based VU estimation methodsunder both fault-free and faults situation are studied. The fuzzy logic rules are quitesimple; therefore, the computation burden among VRFVs is low enough to allow anonline estimation of the uncertainty. The experiment results have verified thatVRFV is very suitable for on-line VU estimation and the validity has been provenby the traditional GUM method under normal situation.Aiming at the novel health evaluation function of self-validatingmultifunctional sensors, the health reliability degree (HRD) is proposed to describethe health level in a quantitative way. The measurement value status estimation ofsingle sensitive unit in traditional self-validating technology has been expanded tothe comprehensive health estimation of multifunctional sensor, in which therelationship among multiple sensitive units is deeply analyzed. From views of singlesensitive unit and overall multifunctional sensor, the HRD concept and methodologyare emphasized to express the health level in a more direct way. Experimentalresults show that HRD is quite suitable for the quantitative health evaluation, and ithas a rapid response to the health changes of multifunctional sensor on differenthealth levels.更多还原