【作者】 崔江；

【导师】 王友仁；

【作者基本信息】 南京航空航天大学 ， 测试计量技术及仪器， 2010， 博士

【摘要】 随着微电子技术与半导体技术的快速发展,模拟电子系统的集成度越来越高,功能越来越复杂,而对其要求的可靠性却越来越高,但相应的可测性正变得越来越差。如何运用信号处理和人工智能技术测试和诊断模拟电子系统中的故障元件或子系统,是目前模拟诊断领域的一个热点。本文综合运用小波分解、特征选择等信号处理技术,并结合基于人工神经网络(ANN)、支持向量机分类器(SVC)等模式识别方法对待测模拟电子电路或系统的故障分类技术进行了研究。本文重点对以下几个方面进行了研究和探讨。针对某些模拟电子系统中故障特征样本数量少且难以获取等问题,研究了一种基于云模型方法的模拟故障特征样本产生方法,并采用神经网络对新产生的扩展样本集进行训练。结果表明,利用新样本集训练过的神经网络对噪声干扰具有较好的鲁棒性。针对模拟故障分类问题,研究了基于支持向量机的分类器设计方法,相关的创新点和贡献包括:提出了一种SVC平均计算量的分析指标,能够定性的对各种SVC的计算效率进行评估;对常规的one-against-rest SVC进行了改进,通过在决策阶段引入K近邻方法,实现了分类性能的提高;为提高one-against-rest SVC的测试效率,提出了一种基于故障字典和空间距离分析的混合故障分类器;提出了一种基于自组织特征映射神经网络(SOFM ANN)的聚类二叉树SVC的故障分类器,这种分类器可以根据样本的空间分布自适应的划分分类结构,从而能够减少测试误差,并给出了其平均计算量的计算方法。为解决单一分类器在模拟电子测试中的不足,提出了一种基于分类器融合的模拟电子系统诊断方法,该方法采用神经网络和SVC作为子分类器,采用模糊推理技术对两种子分类器的输出进行决策融合,实验证明该项技术是有效的,并对噪声干扰有一定的鲁棒性。故障特征提取和选择是模拟电子系统测试领域的关键技术,对于后续的故障分类十分重要。目前的相关研究侧重于故障特征提取,但对于特征选择的研究内容偏少。针对此问题,提出了一种基于标量小波系数的故障特征选择方法。该方法能够选择适合后续one-against-rest SVC分类器的特征,同时能够确定后续分类器的核参数,因此这项技术解决了SVC设计中的核参选择问题;为加快特征选择的速度和效率,提出了一种简化的混合one-against-rest SVC,这项技术能够提高特征选择效率,用于小波特征选择中取得了很好的效果。本文还对实验方法进行了研究,制作了两类实验系统,分别基于PC机和数字信号控制器(DSC)。针对基于PC机的实验系统,分别设计了基于采集卡和数字信号处理器(DSP)的两种数据采集系统;基于DSC的实验系统可脱离PC机运行,利用软件实现了全部诊断任务,包括高速数据采集、基于信号处理的特征提取技术、基于改进型SVC的故障分类等。几个典型的实际模拟电路测试结果均较好的验证了所提出的方法的有效性和实用性。更多还原

【Abstract】 With the development of microelectronics and semiconductor technologies, the integration of the analog electronic circuit system presents with high density and complicated functions. A higher reliability for such a system is required, but the system under consideration is always with low testability. Signal processing and artificial intelligence can be combined together to implement the testing and diagnosis of the analog circuit system at the level of components or sub-system. Such a task is also an interesting and hot subject in the domain of analog circuit testing and fault diagnosis.This paper conducts the research of faults classification for analog electronic circuit under test (CUT) or system under test (SUT), using the technologies of signal processing, i.e. wavelet transformation and feature selection, and pattern classification, i.e. artificial neural network (ANN) and support vector machines classifier (SVC). In our study, several important parts are listed as follows. In some cases of analog circuit fault diagnosis, it is hard to obtain a large number of fault feature samples. Focusing on this problem, we present a novel method of extended set generation for fault feature samples based on cloud model. The artificial neural networks are used to train the extended sets. The experimental results show that the neural classifiers trained with these extended sets are robust to the random noise.Focusing on the problem of fault classifier design, we conduct the research of fault classifier design based on support vector machines classifier (SVC) and, the diagnosis performances of these classifiers are also investigated. In this section, the contributions and the inventions are as follows. In order to evaluate the computational complexity of the SVC qualitatively, we propose a specification of average test calculations. We improve the conventional one-against-rest SVC, and the KNN classifier is introduced in the stage of fault decision. Such a design can improve the performance of the SVC. In order to reduce the testing time of the conventional one-against-rest SVC and improve the testing efficiency, a hybrid classifier based on the fault dictionary (FD) and space distance discriminant is proposed. We also invent a novel SVC, whose training structure is from the training of the self organization feature mapping ANN (SOFM ANN), and this classifier can adaptively find a good training structure according to the distributions of the training samples. This classifier can reduce the diagnosis errors and testing time, and in this part, we also demonstrate the calculation method of average test calculations for this classifier.A single classifier has its drawbacks in diagnosing the analog electronic circuits and in our study, we present a fusion method at the level of decision classifiers. Two classifiers, the ANN and the SVC, are regarded as the sub-classifiers of the system. The fuzzy inference measure is adopted to realize the final decision. The experiments validate the effectiveness of the proposed method and this method is also robust to the random noise.Both of fault feature extraction and selection are crucial to the faults classification of analog electronic system. At present, the feature extraction technique has received many attentions, but the feature selection technique has not been exploited with further research. Focusing on this problem, this paper proposes a selection technique of scalar wavelet features. This novel method can select the proper features for the subsequent one-against-rest SVC, and at the same time, the kernel function parameter is also determined. This technique solves the problem of kernel parameter selection without exhaustive searching, thus, a lot of calculations are saved. In order to speed up the feature selection process, a fast classifier based on the conventional one-against-rest SVC is proposed and this simple classifier can reduce the selection time significantly while good performance can be achieved.In addition to these contents, we also conduct some real experiments. Two systems, based on the personal computer (PC) and the digital signal controller (DSC), are designed, respectively. For the first experimental system, the data collection system is realized with the data acquisition card (DAC) and the Digital Signal Processor (DSP), respectively. For the second system in independence of PC, the software is designed to implement all tasks, such as high-speed data collection, feature extraction based on advanced signal processing and fault decision based on improved SVC. Several real circuits are demonstrated to vindicate the effectiveness and validness of our proposed methods.更多还原