基于DBN特征提取的模拟电路早期故障诊断方法

Analog circuit incipient fault diagnosis method based on DBN feature extraction

针对当前模拟电路早期故障诊断中特征提取方法的不足,提出了应用深度置信网络(deep belief network,DBN)进行特征提取的方法。利用混沌粒子群优化算法,对DBN中受限玻尔兹曼机的学习率开展优化,进一步提升特征提取的性能。相比于其他常用的特征提取方法,提出的DBN特征提取方法可提取出早期故障深度和本质的特征,且具有相同的故障聚集程度高、不同故障的分离能力极为明显的特点。应用二级四运放双二阶低通滤波器仿真电路和Sallen-Key带通滤波器电路板进行早期故障诊断实验,得到的故障诊断正确率分别为98.13%和100%。

Aiming at the deficiency of current feature extraction methods of analog circuit incipient fault diagnosis,the feature extraction method applying deep belief network(DBN)technology is presented.Chaos particle swarm optimization(CPSO)algorithm is employed to optimize the learning rates of the restricted Boltzmann machines in DBN and further improve the feature extraction performance.Compared with other commonly used feature extraction methods,the proposed DBN feature extraction method can extract the deep and essential features of incipient faults.The proposed method also has the features,such as the same high fault aggregation degree and obvious different fault separation capacity.Two-stage four-op-amp biquad lowpass filter simulation circuit and Sallen-Key bandpass filter circuit board were used to carried out incipient fault diagnosis experiments,and the obtained fault diagnosis accuracies are 98.13%and 100%,respectively.