【作者】 连可；

【导师】 王厚军；

【作者基本信息】 电子科技大学 ， 检测技术与自动化装置， 2009， 博士

【摘要】 复杂电子系统故障诊断是一项十分复杂、困难的工作。与一般的系统不同，复杂电子系统的故障有层次性、传播性、相关性和不确定性等独有的特征。因此，单一地利用目前常见的基于信号处理的方法、基于解析模型的方法或基于知识的诊断方法都不足以对复杂电子系统进行及时有效的故障诊断，给出全面合理的结论。本文针对复杂电子系统的特点，提出了一种基于状态监测的复杂电子系统故障诊断方法，将复杂电子系统故障诊断过程划分为“故障检测”、“故障定位”、“故障辨识”三步进行的策略，实现复杂电子系统由表及里，由粗到精的故障诊断。论文主要的研究工作有：1、基于小波变换奇异性分析的故障检测技术研究。故障检测是故障诊断的基础，故障的产生往往表现为检测信号中奇异信号的出现。信号奇异点的位置和奇异值的大小包含着丰富的故障信息，对故障检测至关重要。本文对基于小波变换模极大值的信号奇异性检测方法进行了深入研究，在此基础上对Mallat等人经典的Lipschitz指数估计算法进行了改进，提出了一种新的估计算法。仿真结果表明，本文提出的新算法比经典算法具有更高的估计精度和鲁棒性。2、基于贝叶斯最大后验概率准则的故障定位推理算法研究。本文研究了一种基于贝叶斯最大后验概率准则的故障定位推理算法，该算法以系统多信号流图模型为基础，结合系统先验故障概率计算系统各组件的后验故障概率，以后验故障概率最大为故障定位的准则。将此问题归纳为集合覆盖问题(Set Cover Problem，SCP)，利用拉格朗日松弛算法进行求解，避免了复杂系统故障字典的组合爆炸，实现了复杂系统的故障定位推理。3、针对复杂电子系统整机故障先验分布概率难以获得的情况，本文提出了以复杂电子系统各组件故障随时间的分布函数代替整机故障分布函数确定系统组件当前时刻故障先验概率的思路，对基于贝叶斯最大后验概率准则的故障定位推理算法进行了改进。由于组件故障随时间的分布函数包含了时间信息，因此以此为依据得到的故障定位结论更切合系统运行实际。4、对于系统组件故障随时间的分布函数也无法获得的情况，本文提出采用故障检测传感器检测概率和虚警概率作为故障诊断依据的模糊故障诊断方法。首先将N个故障观测传感器张成N维故障观测空间，然后在故障观测空间设计了模糊函数以描述实际观测向量与故障特征向量的相似度，将故障诊断问题归纳为实际观测向量在故障观测空间中的归类问题，设计了一种模糊多故障诊断算法。5、基于支持向量机(Support Vector Machine，SVM)的故障模式辨识。故障模式辨识本质上是一个多类别模式识别问题。本文针对复杂电子系统完备故障样本集获取困难，需要进行小子样学习的特点采用支持向量机进行模式识别。在研究标准支持向量机算法、分析常见多类分类支持向量机存在不足的基础之上提出了一种基于遗传算法(Genetic Algorithm，GA)的多分类支持向量机决策树优化方法。该方法可以根据具体问题自适应地生成最优或近优SVM决策树。实验结果表明，新方法在提高分类效率、保证分类精度的同时可以大大降低SVM决策树“误差积累”的影响。6、基于状态监测的复杂电子系统诊断方法应用研究。将本文提出的复杂电子系统故障诊断方法应用到某型雷达接收机故障诊断实例中，设计了诊断演示平台，验证了本文提出的相关理论方法的有效性，为开发实用的复杂电子系统故障诊断平台打下了基础。更多还原

【Abstract】 Fault diagnosis of complicated electronic system is an intricate and difficult work. Different from usual fault diagnosis, the faults of complicated electronic systems are hierarchical, propagable, correlative and uncertain. Because of those particular characteristics, any single accustomed diagnosis method such as signal processing based method, analytic model based method or knowledge based method is not good enough to diagnose effectively and get comprehensive and logical fault conclusions. According to the characteristics of complicated electronic systems, this paper presents a state detecting based fault diagnosis method, which disassembles the whole fault diagnosis process into three phases: fault detecting, fault locating and fault recognizing. The main works of this paper include:1. Study on fault detection technology based on wavelet transform singularities analysis. Fault detection is the base of fault diagnosis. Faults of systems can be detected by observing different signals such as voltage, current, temperature, image, et al.. If fault occurs, it will make singularities in those signals. The locations and Lipschitz exponents of singularities contain abundant information about faults, and are important for faults diagnosis. In this paper, the method of measuring singularities based on wavelet transform modulus maxima is studied. Then, we improve the classical method of estimating Lipschitz exponents invented by Mallat, and present a novel algorithm. The result of experiment demonstrates that the method of this paper is more precise and robust than that of classical methods.2. Study on fault locating algorithm based on Bayes maximal posteriori probability principle. In this paper, a fault location algorithm based on Bayes maximal posteriori probability principle is studied. First, the multi-signal model of complicated electronic system is built. Based on the model, the algorithm utilizes the information of apriori probabilitys of faults to compute the maximal posteriori probabilitys according to Bayes theory. This is induced as set cover problem (SCP), and solved by Lagrange relax algorithm. The component that has the maximal posteriori probability is diagnosed as the fault component. This fault location algorithm has ratiocinative ability and can avoid combination blast.3. The apriori fault probabilities of Components are important for the fault locating algorithm based on Bayes maximal posteriori probability principle. In this paper, an idea that uses faults apriori probabilities distributing functions of components to replace that of system is presented to improve the original fault locating algorithm. Because the probabilities distributing functions of components include time information, the results of fault location are more suitable to actual running state of systems.4. For the case that can not get faults distributing functions of components, a fuzzy fault locating algorithm based on false alarm probabilities and detection probabilities of sensors is presented. First, N sensors can create N dimensions fault observation space, in which a fuzzy function is designed to describe the comparability between actual observation vectors and fault character vectors. Finally, a fuzzy multi-fault diagnosis algorithm is presented which induces the problem of fault diagnosis as classification problem of fault observation vectors in fault observation space.5. SVM based fault recognition. The essence of fault recognition is pattern recognition. It is difficult for complicated electronic systems to get sufficient fault samples. In this paper, SVM is used to do fault recognition because of its excellent little samples study ability. First, the standard SVM is studied, and the disadvantages of usual SVM multi-classification methods are analysed. Then, a genetic algorithm (GA) based SVM multi-classification decision-tree optimization algorithm is presented. This algorithm can create optimal or near-optimal decision-tree self-adaptively according to actual instances. Experiment results show that the proposed algorithm can improve recognition efficiency, control error accumulation and at the same time ensure recognition precision.6. Application of the state detecting based fault diagnosis method. The methods proposed in this paper are applied in a fault diagnosis instance of radar receiver. A radar receiver fault diagnosis experiment system is designed, which demonstrates how the theories and methods studied in this paper work in actual fault diagnosis of complicated electronic systems.更多还原