【作者】 刘震；

【导师】 林辉；

【作者基本信息】 西北工业大学 ， 检测技术与自动化装置， 2007， 博士

【摘要】 机内测试(Built-in Test，BIT)技术是改善系统测试性和诊断能力的重要途径，在保障装备的战备完好性、提高维修效率方面发挥了重要作用，但目前广泛应用在机载电源中的传统BIT技术由于诊断方法过于单一，对诊断信息的利用能力也非常有限，因此在使用过程中表现出的故障检测、隔离能力差、虚警率高的问题严重制约着其效能的充分发挥。尤其是下一代多电飞机，由于大量采用电能来提升飞机的整体性能，使得多电飞机对电源系统的可靠性、测试性、故障诊断和容错能力的要求都远高于常规飞机。因此，在下一代多电及全电飞机中，研究新型有效的智能BIT技术对提升飞机的整体性能具有重要意义。本文以“多电飞机电气系统关键技术研究”课题为背景，对其关键技术之一——机内测试技术进行了探索与研究，以期通过采用智能化方法来提高电源系统BIT的故障诊断性能。论文的主要研究成果及创新如下：1．系统地分析和总结了多电飞机电源系统可能的故障模式，建立了相应的FMEA分析表。根据特征检测点的选取原则，对系统中的特征信号进行分析，确定了合理的BIT特征检测点，为电源系统智能BIT诊断技术打下基础；2．建立了飞机电源BIT动态系统的数学描述模型，在此基础上，针对目前电源BIT故障诊断在方法选取上缺乏合理依据的现状，从理论上分析了当前所采用的BIT诊断方法的不足，给出了提高电源BIT系统智能化诊断性能的几种策略；3．针对飞机电源BIT系统传统诊断方法的缺陷，研究了基于竞争学习思想的无监督聚类神经网络，实现了一种能够广泛用于飞机电源智能BIT系统的新型诊断方法：①针对标准无监督GLVQ模型在分类上存在的不足，提出了一种改进型IGLVQ。通过对网络亏损因子进行修正，并推导出网络在新的亏损函数下的学习规则，有效的克服了输入数据范围及类别数变化对分类的影响；②在改进型IGLVQ模型的基础上，引入了有监督的学习矢量量化层，构成了一种新的无监督混合聚类模型(HIGLVQ)，提高了无监督聚类网络对相近故障模式类的区分能力；③将新型的基于HIGLVQ混合聚类网络的诊断模型用于多电飞机电源BIT系统，结果表明，本章提出的方法用于飞机电源BIT状态识别时其准确率较高，可有效的提高电源BIT系统的故障诊断性能；4．针对目前电子系统BIT虚警机理的研究和多电飞机电源系统的自身特点，分析了电源系统中两类暂时性故障的产生机理，并从两类故障的产生条件、持续时间、发生概率和表现特性上对其特点进行了分析和阐述；从BIT虚警率的概率数学模型角度，分析了识别两类故障对虚警率的影响，从理论上证明了识别两类故障状态可有效的降低BIT系统的固有虚警率；提出了基于HIGLVQ—优化Bayes风险决策的电源系统BIT智能虚警滤波模型。将诊断中的概率因素融入诊断器中，通过对LRU级的诊断结果做进一步判定，以决策结果的真实性。结果表明，本章提出的方法能够有效消除瞬态或间歇故障产生的虚警；5．为增强电源BIT系统的智能诊断水平，研究了故障预测技术在电源BIT系统中的应用。针对电源系统渐变故障的特点，通过建立电源系统的隐马尔可夫(HMM)多阶预测模型，提出了基于特征频谱和一维时序信号的故障预报策略；针对HMM模型在线预测不能实时更新参数的问题，提出了一种新型的径向基HMM预测模型(RBHMM)，通过在线数据实时更新模型参数，以使预测模型能够自适应地跟踪系统的状态变化；将基于RBHMM的预测模型用于飞机电源系统BIT中，对一维时序导通率故障信号进行预测，实验表明，基于RBHMM的在线实时预测模型自适应能力强，预测性能高于原有HMM模型离线训练／在线趋势预测的方式，能够有效地增强电源BIT系统的智能故障诊断性能。更多还原

【Abstract】 The Built-in Test (BIT) is an integral capability of the mission equipment whichprovides an on-board, automated test capability to detect, diagnose, or isolate systemfailures. It is an effective approach to improving testability and maintainability of acomplex system. However, with the increasing requirements of fault detection and themaintenance time, many problems of this conventional BIT technique are manifested,such as notorious False Alarm (FA), Cannot Duplicate (CND) and Retest OK (RTOK),which have some strong effects on the readiness of military aircraft, especially on that ofMore-Electric Aircraft (MEA). Due to the electric power widespread using in the MEA,the reliability and fault tolerance capability of More-Electric Aircraft Electrical PowerSystem (MEAEPS) must be significantly higher than conventional aircrafts. Furthermore,the MEA requires that more effective and efficient testing and fault diagnosis techniquesbe developed to improve the reliability of the EPS. So it is very important to study theintelligent BIT technique and its application in MEAEPS in order to improve theintegrated capability of MEA. Supported by the National Project—Research on keytechnologies of more-electric aircraft electrical system, this paper is aiming to in-depthresearch on theories and methods on intelligent BIT fault diagnosis in order to enhance thediagnostic capability of the BIT system of MEAEPS. The main contents of the dissertationare as follows:1. All of faults which may happen in MEAEPS are analyzed and summarized in thisdissertation, and a Failure Mode and Effect Analysis (FMEA) table is established. On thebasis of that, according to the selecting principle of fault detecting points, some soundfault detecting points of BIT system are selected.2. A mathematical description model of BIT dynamic system of aircraft electricalpower system is established. Aimed to the lack of reasonable basis in terms of selectingfault diagnosis methods employed in aircraft electrical power system, the shortage of BITfault diagnosis methods which now applied in aircraft electrical power system is analyzedbased on the given BIT mathematical model, and several methods which can improve thediagnosis capability of BIT system are given.3. Aimed to the shortage of-conditional BIT techniques employed in aircraftelectrical power system, an unsupervised clustering neural network based on competinglearning method is studied. Firstly, aimed to the drawbacks of original GLVQ network interms of classification; an improved IGLVQ is proposed. The IGLVQ algorithm adopts anew form of loss factor, and its learning rules are derived through finding a minimum ofthe loss function, which avoid the influence of the input space scale and the class numberto classification. Secondly, in order to overcome the common drawback of unsupervised networks that it can not using the prior classifying information, a LVQ layer is added tothe, IGLVQ network to construct a hybrid neural network model (HIGLVQ), whichimproves the ability to distinguish the similar classes. Thirdly, the new HIGLVQ networkhas been applied to the intelligent BIT system of the MEAEPS, and the results show thatthe proposed method has a good performance in pattem classification and it is promisingto improve the fault diagnosis capability of the BIT system.4. Through analyzing the characteristics of MEAEPS and the status-in-quo of studyon BIT false alarm of electronic system, the causes which influence BIT decision aregiven. A main cause is that the conditional BIT techniques which applied to the aircraftelectrical power system avoid intermittent faults and transient faults, which induce thehigh FAR of BIT system. On the basis of that, the mechanisms of how the intermittentfaults and the transient faults produce and the characteristics of them are analyzed. Basedon the probability mathematical model of BIT FAR, an important conclusion is provedthat identifying these faults can reduce the FAR of BIT system effectively. And a falsealarm filter model, based on HIGLVQ-optimal Bayes decision, is proposed. Theexperimental results show that this proposed model can eliminate effectively false alarmcaused by the intermittent faults and the transient faults in BIT system of MEAEPS.5. In order to enhance the intelligent diagnosis level, fault predicting theories appliedto BIT system of MEAEPS are studied. Aimed to the slow-emerging faults, a faultpredicting method based on frequency spectrum and one-dimension time series signal isproposed, which uses the Hidden Markov Model (HMM). On the basis of that, a novelRadial Basis Hidden Markov Model (RBHMM) and its online parameter-updatingalgorithm are given since the original HMM model can not update in real time. Using theRBHMM-based fault predicting model to predict the one-dimension time series faultsignals of the BIT system of MEAEPS, the proposed method shows that it has a betterperformance than that of the original HMM model in term of fault predicting, and it canimprove the intelligent fault diagnosis capability of the BIT system of MEAEPSeffectively.更多还原