【作者】 郑红岩；

【导师】 李红波；

【作者基本信息】 解放军信息工程大学 ， 检测技术与自动化装置， 2011， 硕士

【摘要】 随着现代电子技术的高速发展,电子设备的结构越来越复杂,其集成度和密集度不断增加,而相应的电路故障诊断技术,尤其是模拟电路故障诊断技术却发展缓慢,制约着设备保障能力的发展与提高。因此,模拟电路故障诊断技术研究具有十分重要的现实意义,从而成为当前电路故障诊断领域的研究热点。在对当前模拟电路故障诊断领域的研究现状进行归纳总结的基础上,本文采用了基于BP神经网络的模拟电路故障诊断方法。针对该方法诊断模型的建立流程,重点对特征提取算法和BP神经网络优化算法进行了深入研究和探讨,提出了基于非高斯性分析的特征提取算法和改进的GA—BP神经网络算法,主要工作包括:1.从特征提取、智能故障诊断和BP神经网络优化的角度出发,对现有的各类算法进行了较为系统的总结,综合对比了各类算法的性能和优缺点。2.针对特征提取过程中特征向量维数较高,计算复杂度较大的问题,提出了一种基于非高斯性分析的特征提取算法。该算法利用电路在正常模式及各种故障模式下输出信号偏离高斯信号的不同程度,将信号的峭度、负熵和重心作为特征参数进行特征提取。仿真结果和实际验证表明,该算法故障分辨率高,能有效降低特征向量维数,且计算方法简单,易于实现,具有一定通用性。3.深入研究了普通的GA—BP神经网络算法,针对其在同时优化BP神经网络结构和初始权值阈值时,存在编码过于冗长,优化性能较差等问题,提出了一种改进的GA—BP神经网络算法。该算法对原有算法的编码方式和适应度计算方法进行了优化,有效缩短了编码长度。仿真结果和实际验证表明,该算法有效提高了BP神经网络的设计效率、收敛成功率及网络的性能。4.针对实际设备中的较大规模电路,设计了基于故障二叉树和BP神经网络的TPS开发方法,编写了相应的用户开发界面。最后结合XX自动测试诊断系统,将本文算法在WJ8615P超短波接收机的TPS开发中予以实现。实际测试结果达到了预期的要求,进一步验证了本文所提算法的可行性与有效性。更多还原

【Abstract】 With the developing of modern electronic technology, the electronic equipment’s structure becomes more and more complicated, and its integration and density is also increasing continuously. But the corresponding technology of fault diagnosis of circuits, especially that of analog circuits is developing slowly, which restricts the maintainability of equipments. Therefore, it is of great significance to study the fault diagnosis technology of analog circuits, which has become a hotspot in the fault diagnosis field of circuits nowadays.After summarizing the current technologies of fault diagnosis of analog circuits, this thesis adopts BP neural network algorithm. This paper makes a study of the feature extraction algorithm and the optimizition algorithm of BP neural network, which both are critical sections in the fault diagnosis procedure. This thesis proposes a feature extraction algorithm based on the non-gaussianity analysis and an optimized algorithm of GA-BP neural network. The main work is as follows:1 This paper summarizes the current algorithms in the field of feature extraction, fault diagnosis using artificial intelligence and the optimization of BP neural network, the credits and infects of all those technologies above are also illustrated.2 A feature extraction algorithm based on the non-gaussianity analysis is proposed. It can overcome the problem of large dimension and high computing complexity, which exist in the current algorithms of feature extraction. This algorithm selects the centroid, the kurtosis and engentropy of the signal as the features. The last two features reflect the deviations between a gauss signal and signals sampled in the normal mode and various faulty modes of the circuit. According to the outcome of simulations and actual experiments, this algorithm has the advantage of high definition, low feature dimension and computing simplicity. It is also easy to realize and has generality to some extent.3 The normal GA-BP algorithm has the disadvantages of long codes and poor optimization performance. This thesis puts forward an improved GA-BP algorithm. The methods of coding and fitness computing are optimized to have a short coding length. According to the outcome of simulations and actual experiments, this algorithm can achieve a better designing efficiency, a higher ratio of convergence and a better performance of network.4 A developing method of TPS based on fault binary tree and BP neural network is designed to cope with the larger scale circuits in the equipment. The corresponding developing interface for users is also provided. Finally, the algorithms of this thesis are applied to an ATS to develop the TPS of WJ8615P, which is a VHF receiver. The results of experiment achieve the anticipated performance, proving the feasibility and effectiveness of the algorithms proposed in this paper.更多还原