【作者】 金瑜；

【导师】 陈光（礻禹）；

【作者基本信息】 电子科技大学 ， 测试计量技术及仪器， 2008， 博士

【摘要】 模拟电路故障诊断理论和方法研究是目前研究的热门课题,现代电子技术的发展对模拟电路的测试和故障诊断提出了更高的要求。由于模拟电路故障诊断有其自身的许多困难,因此传统的故障诊断理论和方法在实际工程中很难达到预期的效果。而以神经网络为代表的计算智能技术为模拟电路故障诊断提供了一条有效的途径,受到学术界的广泛关注。本论文以现代测试技术、系统辨识、神经网络和小波分析等理论为基础,深入研究了小波神经网络和多小波神经网络诊断模拟电路故障的方法。本论文的工作主要体现在以下三个方面:1.研究了基于神经网络的模拟电路故障诊断方法。在分析和阐述神经网络诊断模拟电路故障的系统结构和原理的基础上,研究了BP网络诊断模拟电路故障的方法。针对BP网络易收敛于“局部最小值”这一缺陷,提出了用遗传算法优化BP网络权值的方法,该方法能在较短的时间内得到BP网络的最佳参数。2.研究了多种小波神经网络诊断模拟电路故障的方法。针对BP网络在模拟电路故障诊断中存在的缺陷:收敛速度慢、容易收敛于“局部最小值”,网络的结构设计仅凭经验以及存在误诊。提出了以下几种小波神经网络诊断模拟电路故障的方法:1)基于BP小波神经网络的模拟电路故障诊断。首先用具有良好时频局部特性的小波基函数替代传统BP网络的激励函数构造了BP小波神经网络。然后用它诊断模拟电路的故障。诊断结果表明,BP小波神经网络的诊断效果比BP神经网络要好,且收敛速度比BP网络明显加快。2)基于小波框架神经网络的模拟电路故障诊断。根据离散小波框架理论构造了小波框架神经网络,并用于诊断模拟电路的故障。结果表明,小波框架神经网络的故障诊断率明显比BP网络要高,收敛速度更快,且网络结构设计有理论依据——小波框架理论。3)基于多分辨小波神经网络的模拟电路故障诊断。从多分辨分析思想出发构造了多分辨小波神经网络。把待诊断电路的样本输入到该网络中进行训练及测试,结果表明:该网络在诊断模拟电路故障过程中,收敛速度快,避免了收敛于“局部最小值”,且网络结构设计有理论依据,更重要的是该网络不仅对已有的故障诊断不存在误诊,而且对新出现的故障也能够全部正确地分类。3.研究了多分辨多小波神经网络诊断模拟电路故障的方法。和小波相比,多小波同时具有正交性、紧支撑性、对称性和高阶消失矩等优点。本文构造了一种多分辨多小波神经网络,并用于模拟电路的故障诊断。诊断结果表明,该网络保留了多分辨小波神经网络的所有优点,同时克服了多分辨小波神经网络的“维数灾”问题,但良好性质的多小波构造是难点,目前还不成熟。本论文在小波神经网络诊断模拟电路故障的方法研究中已经取得了一些成果。在接下来的阶段里将继续针对这些工作开展深入的研究。更多还原

【Abstract】 As modern electronic technology develops very rapidly, the research on the fault diagnosis theories and methods for analog circuits is becoming very popular and also challenging. However, traditional fault diagnosis theories and methods cannot meet actual requirement due to difficulties inherent in analog circuits fault diagnosis. On the other hand, computation intelligence technologies including neural network (NN) method and something else are great interest to researchers, which might provide potential solution for fault diagnosis. In this dissertation, based on the theories of modern test technology, system identification, neural network, wavelet analysis, and so on, the application of wavelet neural network (WNN) and multi-wavelet neural network (MWNN) in fault diagnosis for analog circuits is studied. In short, the main works of this dissertation are as follows:1. The NN-based fault diagnosis for analog circuits is studied. Firstly, the structure and principle of analog circuits fault diagnosis system based on the NN are analyzed in detail, on which the application of the BP neural network (BPNN) in fault diagnosis is studied. Then, due to the fact that the BPNN usually converges to local minimum, a new way is considered in this dissertation by using the genetic algorithm (GA) to optimize the weights in the hidden layers of the BPNN. The results show that the optimal parameter of the BPNN can be obtained in short time.2. The research on the different WNNs-based fault diagnosis for analog circuits has been done. Because the traditional BPNN-based fault diagnosis has the following disadvantages such as low convergence rate, easy convergence to local minimum, structure design by experience and false diagnosis. In this dissertation, several WNNs-based fault diagnosis methods for analog circuits are proposed.1) Analog circuits fault diagnosis method based on the BP wavelet neural network (BPWNN) is proposed. At first, wavelets with good time-frequency localization property instead of the activation functions of the traditional BPNN, is used to construct the BPWNN. Then, it is applied in fault diagnosis for analog circuits. Finally, experimental results indicate that the diagnosis efficiency of the method using the BPWNN is better than that of the method using the BPNN, and that the convergence rate of the former is faster than that of the latter.2) Analog circuits fault diagnosis method based on the frame wavelet neural network (FWNN) is proposed. Firstly, a FWNN is constructed according to the theory of wavelet frame. Then, it is applied in analog circuits fault diagnosis. Experimental results show that this method is better than the method using the traditional BPNN. And what’s more important, wavelet frame theory can be used as theoretical principle in network structure design.3) Analog circuits fault diagnosis method based on multi-resolution wavelet neural network (MRWNN) is proposed. According to the multi-resolution theory, the MRWNN is constructed in this dissertation. Experimental results show that the method using this MRWNN is characterized by many advantages such as fast convergence rate, avoiding of converging to local minimum, and sound theoretical foundation for structure design. What’s more, there is no false diagnosis for existent faults and those new faults can be exactly classified.3. Analog circuits fault diagnosis method based on multi-resolution multi-wavelet neural network (MRMWNN) is proposed. Compared with conventional wavelet, multi-wavelet is simultaneously orthogonality, compact support, symmetry and vanishing moment. In this dissertation, the MRMWNN is constructed and applied in fault diagnosis for analog circuits. This new fault diagnosis method not only keeps the same advantages as those of the MRWNN, but also overcomes the disadvantage inherent in the MRWNN, i.e., dimension disaster. However, it is not easy to construct good multi-wavelet at present.In this dissertation, some research works on the application of wavelet neural network in fault diagnosis for analog circuits have been done and some achievements have been made. In the future, further research will continue to be done.更多还原