CAJ | 학술논문

在模拟电路故障诊断中，由于标准的BP神经网络算法在训练样本时存在着收敛速度慢、分布不均匀、效率不高等缺点，导致电路的整体诊断性能下降。提出了一种将Levenberg-Marquardt （LM）算法与神经网络相结合的方法，对电路的脉冲信号进行多尺度分解，提取故障特征作为神经网络的输入对网络进行训练。实验仿真表明， Pspice与Matlab相结合的样本训练方法的稳定性高于传统方法，证明了该方法的实用性与可行性。
재모의전로고장진단중，유우표준적BP신경망락산법재훈련양본시존재착수렴속도만、분포불균균、효솔불고등결점，도치전로적정체진단성능하강。제출료일충장Levenberg-Marquardt （LM）산법여신경망락상결합적방법，대전로적맥충신호진행다척도분해，제취고장특정작위신경망락적수입대망락진행훈련。실험방진표명， Pspice여Matlab상결합적양본훈련방법적은정성고우전통방법，증명료해방법적실용성여가행성。
When analog circuit faults are diagnosed, the disadvantages concerning the standard BP neural network algokithm such as slow convergence, uneven distribution and lower efficiency when training sample result in de-creased overall diagnostic performance of the circuit.This paper proposed a method combining neural network algo-rithm and the Levenberg-marquardt ( LM ) algorithm to decompose the pulse signal of the circuit from multiple scales and extract fault features as the input of neural network to train the network.The simulation results showed that the stability of sample training method combining Pspice and Matlab was higher than traditional methods, and that the practicability and feasibility of this method were verified.