基于迁移学习LSSVM的模拟电路故障诊断

Analog Circuit Fault Diagnosis Using Transfer Learning and LSSVM

采用数据驱动方法进行模拟电路故障诊断时,在目标故障数据较少的条件下,诊断效果显著下降。针对该问题,提出一种基于TL-LSSVM的模拟电路故障诊断方法。该方法将相关的源域数据迁移至目标故障训练集,首先提取输出信号的小波系数作为特征数据,然后在LSSVM分类器的目标函数中增加源域辅助数据的误差惩罚项,构建出新的诊断模型。以滤波电路为诊断实例,实验结果表明,该方法使单、双故障诊断正确率分别达到97.2%和95.7%,显著提高了诊断正确率。

When diagnosing analog circuit faults using data-driven approaches,if the target fault set is insufficient,the diagnostic performance will decrease significantly.To solve the above problem,a novel fault diagnosis method based on transfer learning and least square support vector machine(TL-LSSVM)was proposed.This method transfers related sample set into the target fault training set effectively.In the diagnosis,wavelet coefficients of the output signals are firstly extracted as features.Then,error penalty term of the source domain auxiliary data is added to the objective function of the LSSVM classifier,and constructs a new diagnosis model.Finally,testing samples are imported to the new model for classification.The performance of the proposed approach was validated with a filter circuit.In the experiment,the proposed approach made the accuracy of single and double fault diagnosis reach 97.2% and 95.7%,respectively.and show a significant improvement of diagnostic accuracy in situation of insufficient target fault samples.