Abstract: In order to solve the problem of incomplete and inaccurate feature extraction in analog circuit soft fault diagnosis, a new method of two classes feature extraction based on Kernel Linear Discriminant Analysis（KLDA） following integrated time and frequency domains is proposed in this paper. At first, the statistics features in time domain such as mean, standard deviation and energies within different frequency bands by wavelet packet decomposition as frequency domain features are extracted from the acquired fault response signals. Then the kernel linear discriminant analysis method is used to further optimize and select fault features, which guarantees the validity on fault feature. Finally, the obtained optimal fault features are inputted into support vector machine to distinguish different faults. Experimental results on Sallen-Key band-pass filter circuit show that this method can reflect the essential characteristics of fault response signal and improve the performance of fault diagnosis effectively.

Key words: soft fault diagnosis of analog circuit, feature extraction, wavelet packet energy spectrum, statistic feature in time domain, kernel linear discriminant analysis, directed acyclic graph support vector machine

摘要： 为了解决模拟电路软故障诊断中特征提取不全面准确的问题，提出了一种基于综合时频域及核判别分析的两级特征提取新方法。首先，对采集到的故障响应信号分别提取均值、方差等时域统计特征和小波包分解后不同频带的能量作为频域特征；然后，通过核判别分析方法对故障特征进一步优选，从而保证故障特征的准确有效性；最后，将所得到的最优故障特征输入支持向量机进行故障诊断。对Sallen-Key带通滤波器电路的仿真实验结果表明，该方法能够很好地反映故障响应信号的本质特征，有效提高故障诊断的性能。

关键词: 模拟电路软故障诊断, 特征提取, 小波包能量谱, 时域统计特征, 核判别分析, 有向无环图支持向量机