【作者】 胡雷；

【导师】 胡茑庆；

【作者基本信息】 国防科学技术大学 ， 机械工程， 2010， 博士

【摘要】 飞行器包括航天飞机、运载火箭、航空固定翼飞行器和低空旋翼直升机等等。飞行器健康管理技术对于监测飞行器关键部件工作状态、评估其状态退化趋势、预测其剩余使用寿命、保证其安全运行具有重要意义,是执行状态基维修策略乃至自治性维修保障的基础和前提。时有发生的航空航天事故也不断地给人们以警醒:开展飞行器健康管理关键技术研究的需求十分迫切。在飞行器健康管理体系中,监测是一个极其重要的主题,也是诊断、预测和减缓等其它主题的基础。但在飞行器状态监测的实施过程中,常常存在先验知识缺乏、故障样本稀缺和故障模式不完备的问题。针对这些问题,本文以某型液体火箭发动机的涡轮泵为研究对象,开展了面向飞行器健康管理的新异类检测方法研究,主要研究工作包括:1.深入分析了提高新异类检测器推广能力需要考虑的因素。较为系统地研究了常见新异类检测方法的基本原理,对一些性能优异的新异类检测器进行了数值仿真分析。面向健康管理,总结了新异类检测器的设计原则,提出了三种借助新异类检测开展状态监测的故障检测策略。研究表明,新异类检测可以依据对已知观测样本的学习,实现对已知和未知异常的识别。新异类检测器的设计应综合考虑稳健性的折衷、推广性能、计算的复杂度和适用于不同情况的故障检测策略等。2.深入研究了涡轮泵状态监测的单类支持向量机新异类检测方法。(1)详细介绍了两种具有不同几何解释的单类支持向量机基本原理,结合数值仿真,深入分析了单类支持向量机的关键参数对其性能的影响,并据此给出了单类支持向量机故障检测方法的参数设置原则。(2)应用单类支持向量机对涡轮泵历史试车数据进行了检测分析,结合单类支持向量机的参数设置原则和对涡轮泵试车数据分析的交叉验证,优化了单类支持向量机的参数,降低了虚警率。研究表明,单类支持向量机的性能与其参数设置密切相关;单类支持向量机的原理性误差和特征量的一致性不足会造成较大的检测虚警,通过增大高斯带宽系数和使用偏置缩放因子可以显著地降低虚警率。3.提出了一种用于涡轮泵试车数据在线检测的双偏置单类支持向量机自适应在线检测算法。(1)提出了用于求解单类支持向量机的序贯最小优化算法,降低了求解单类支持向量机的时间和空间复杂度。(2)在序贯最小优化算法的基础上,提出了一种双偏置单类支持向量机自适应在线检测算法,消除了单类支持向量机的原理性误差和因工况变化等外界随机因素造成的误差,剔除了异类样本对检测模型自适应更新的贡献,避免了检测器随异类样本自适应更新的现象。(3)利用提出的双偏置单类支持向量机自适应在线检测算法对涡轮泵历史试车数据进行了检测分析,同时也完成了对该方法的验证。研究表明,序贯最小优化算法可以显著提高单类支持向量机的训练效率,双偏置单类支持向量机自适应在线检测算法能够在无虚警的情况下有效识别涡轮泵的异常状态,包括叶片断裂/脱落和转子碰摩。4.提出了一种基于增量聚类和单类支持向量机的完备样本区域描述方法,用于新异类检测和涡轮泵状态监测。(1)提出了一种具有样本压缩功能的增量聚类算法,用于解决完备样本区域描述所涉及的大样本学习问题。(2)提出了一种集成增量聚类和单类支持向量机的完备样本区域描述方法,建立了涡轮泵时域统计特征的完备样本区域描述模型。(3)构建了涡轮泵故障检测系统,将提出的完备样本区域描述方法在该系统中进行了集成,使用涡轮泵历史试车数据验证了该方法的有效性。研究表明,集成增量聚类和单类支持向量机的完备样本区域描述方法能够从大量样本中增量式地提取均匀分布、张满目标区域且大小可控的代表样本集,实现样本集的压缩;能够建立用以描述完备样本分布区域的边界。对涡轮泵试车数据的检测结果表明,该方法能够在无虚警的情况下有效识别涡轮泵的叶片断裂/脱落、转子碰摩、传感器失效等故障。5.提出了基于并联单类支持向量机和串联单类支持向量机的故障诊断方法,将新异类检测拓展应用到了故障诊断,使用涡轮泵试车数据验证了单类支持向量机故障诊断方法的有效性。研究表明,与各种基于支持向量机两类分类器的故障诊断方法相比,单类支持向量机故障诊断方法的训练样本重复使用率更小,诊断效率更高,扩展性更好,而且能够识别未知状态和已知状态。仿真数据及涡轮泵试车数据的分析结果表明了该方法的有效性。更多还原

【Abstract】 Health management (HM) technologies for flight vehicle, such as space shuttle, carrier rocket, fixed-wing aircraft and helicopter, are of significance to enable operating condition monitoring, degradation trend assessment, remaining useful life prediction and safe operation assurance. Thus HM is considered the premise and foundation that supposes condition based maintenance, even autonomic logistics. Occasionally occurred accidents warn us constantly that it is imperative to research on key techniques of HM.Condition monitoring is an important theme in the architecture of vehicle HM. And it provides foundation for other themes including diagnosis, prognosis and mitigation. One of the challenges in condition monitoring of flight vehicles or their key components is lack of fault samples and prior knowledge about fault modes. Accordingly, this dissertation takes the turbopump of a liquid rocket engine as object and researches on novelty detection methods oriented to vehicle HM. The detailed contents and innovative work can be summarized as follows.1. Facts associated with the generalization of novelty detection methods are deeply analyzed. Philosophy of popular novelty detection methods is systematically studied and some excellent methods of them are analyzed via numerical simulation. Oriented to HM, design principles are summarized and three strategies for fault detection with novelty detection methods are presented.The research shows that novelty detection is able to recognize known and unknown abnormities according to known samples. Some issues, such as robustness trade-off, generalization, computational complexity, detection strategies that suit for different cases, should be considered when develop novelty detectors.2. One-class support vector machine (OCSVM) based novelty detection method is deeply studied for turbopump condition monitoring.(1) Two kinds of OCSVMs with different geometrical explanations are introduced in detail. The influence of OCSVM’s parameters to its performance is thoroughly analyzed integrating numerical simulation. And principles to set OCSVM’s parameters are given accordingly.(2) OCSVM based novelty detector is applied to detect historical test data of the turbopump. OCSVM’s parameters are optimized according to their set principles and cross-validation of detection results, and by which false alarms are reduced.The above research shows that OCSVM’s performances are closely related to its parameter set. The principle error of OCSVM and the bad consistency of feature vectors are main factors that cause false alarms. These false alarms can be reduced evidently by increasing OCSVM’s Gaussian kernel width and by introducing an offset scaling parameter.3. An online adaptive novelty detection algorithm based on double-offset OCSVM is presented and applied to the detection of turbopump real test data.(1) A sequential minimal optimization (SMO) algorithm is introduced to solve the quadratic optimization problem in OCSVM, which reduced the computational complexity of OCSVM.(2) A double-offset OCSVM online detection algorithm based on SMO algorithm is developed. In this algorithm, detection error caused with OCSVM itself and changing environmental conditions is eliminated. Abnormal samples detected are prevented from contributing to the adaptive update of the detection model.(3) The online detection algorithm presented above is applied to turbopump historical test data detection, which also verifies this algorithm.The research shows that SMO algorithm can improve the training efficiency of OCSVM evidently. The online detection algorithm based on double-offset OCSVM is able to detect kinds of novel events of the turbopum, involving vane shedding and rub-impact. And there are not any false alarms.4. A complete sample region description (CSRD) method that integrates OCSVM with incremental clustering is presented for novelty detection and turbopump condition monitoring.(1) An incremental clustering algorithm that enables sample compression is presented to solve the learning problem of large samples in CSRD.(2) A CSRD method integrating OCSVM with incremental clustering is presented and a complete description model is established for the time-domain statistical features of the turbopump.(3) A turbpump fault detection system is constructed and the CSRD method is integrated into the system. Validity of the CSRD method is validated with turbopump historical test data.The research shows that the CSRD method is able to extract a more representative sample set from large rude samples incrementally. The representative set distributes uniformly and covers the entire target region. And the size of the representative set is under control. The CSRD method is also able to generate a boundary surrounding the target region. Detection results of turbopump historical test data demonstrate that the method can identify different spikes in vibration signals caused by abnormal events such as vane shedding, rub-impact and sensor faults. And there are not any false alarms.5. Novelty detection is exploited to fault diagnosis. Diagnosis methods based on parallel OCSVMs and series OCSVMs are presented and validated.The research shows that compared with diagnosis methods based on SVM, OCSVM based methods have better diagnosis efficiency and expansibility. And they are able to recognize known and unknown conditions. Their validity is verified with the classification of simulation data and the classification of turbopump historical test data.更多还原