CAJ | 학술논문

提出一种采用人工神经网络判断扬声器是否存在异常音的方法。首先简单介绍了获取扬声器异常音曲线的方法和人工神经网络中的BP模型及其训练方法，并比较了基本BP算法和共轭梯度法两种训练方法的差异。再将所获得的异常音曲线作为人工神经网络的输入向量，将听音员的听测结果作为目标向量，并使用共轭梯度法进行网络的训练。最后通过已训练好的人工神经网络判断扬声器是否存在异常音。实验结果表明，该方法可替代传统的人工设置门限的方法，并可大幅降低扬声器异常音检测的虚警率。
제출일충채용인공신경망락판단양성기시부존재이상음적방법。수선간단개소료획취양성기이상음곡선적방법화인공신경망락중적BP모형급기훈련방법，병비교료기본BP산법화공액제도법량충훈련방법적차이。재장소획득적이상음곡선작위인공신경망락적수입향량，장은음원적은측결과작위목표향량，병사용공액제도법진행망락적훈련。최후통과이훈련호적인공신경망락판단양성기시부존재이상음。실험결과표명，해방법가체대전통적인공설치문한적방법，병가대폭강저양성기이상음검측적허경솔。
This paper proposes a method of using neural network to judge whether a loudspeaker is good or not. First, the method of how to obtain the Rub&Buzz curve and the BP model including its training methods are simply introduced. Besides, the comparison between the basic BP algorithm and the conjugate gradient algorithm is also made. Then the Rub&Buzz curve is used as the BP network’s input vector and the judgment result of experienced worker is used as the BP network’s output vector and use the conjugate gradient algorithm to train the network. Finally, the trained BP net-work can judge whether the measured loudspeaker is good or not. The experimental results show that judging a louds-peaker is good or not by a threshold, which is set up by engineer, can be replaced by artificial neural network, and the false alarm rate is greatly reduced.