基于机器学习与卷积神经网络的放电声音识别研究

Study of Discharge Sound Diagnosis Based on Machine Learning and Convolutional Neural Networks

为了实现对电气设备放电声音的精准检测,文中筛选比较了多种经典的机器学习算法和新兴的卷积神经网络算法,以期得到识别效果最优的选择。首先对音频进行预处理,再通过将放电声与环境噪声和变电站正常工况背景声混合来模拟变电站真实工作环境,并使用梅尔频率倒谱系数提取特征,最后采用支持向量机等机器学习算法与卷积神经网络算法进行识别,选取识别效果最佳的算法并考察不同采样频率、采样时长等因素对识别效果的影响。实验结果表明,使用梅尔频率倒谱系数提取特征可以良好区分放电与环境噪声,支持向量机在一系列算法中识别放电声音能力最强,采样频率、标准化方式等因素对识别效果影响较小。

In order to realize more accurate detection of the discharge sound of electrical equipment,a variety of classical machine learning algorithms and the emerging convolution neural network algorithm were compared and screened to get the best choice.In this study,audio preprocessing was implemented firstly,then the discharge sound was mixed with environment noise and the background sound of substation on normal working condition to simulate the real working environment of substation.The feature vector and spectrum were extracted by Mel frequency cepstrum coefficients,and the recognition performances of convolution neural network and machine learning algorithms such as support vector machine(SVM) were compared.Selected the best recognition algorithm,the influence of different sampling rate or other factors were also studied.The experimental results show that the method of extracting eigenvector or spectrogram by Mel frequency cepstrum coefficient can distinguish discharge from environment noise,and SVM has the best performance in a series of algorithms.Sampling rate,methods of standardization and other factors do not have a significant influence on recognition accuracy.