基于双通道时频卷积神经网络的故障电弧检测

Arc fault detection based on dual-channel time-frequency convolutional neural network

交流故障电弧产生的高温极易点燃周围的可燃材料,是引发电线火灾的重要原因之一.准确检测不同类型的故障电弧对于预防重大火灾事故的发生具有重要意义.然而故障电弧的复杂性与隐蔽性给检测方法带来了极大挑战.基于阈值和电流特征提取的技术难以全面概括故障电弧的特征,而大多数基于深度神经网络的方法直接对电流信号进行特征学习,忽略了信号中的频率信息,从而导致泛化能力差的问题.对此,本文提出了基于时频特征学习的双通道时频卷积神经网络的故障电弧识别方法,设计了可学习的自适应离散小波变换,用于提取一维信号中的多尺度特征,同时通过短时傅里叶变换获取二维的时频图像特征,分别在这2种特征信号上进行卷积,最后将2个通道中学习的特征进行融合,用于分类预测.通过对故障电弧发生器采集到的3种工况下电弧电流信号进行性能评估,验证所提方法的有效性.实验结果表明,该方法与其他同类方法相比具有更高的电弧识别准确率,达到了97.91%.

The high temperature generated by the AC fault arc can easily ignite the surrounding combustible materials,which is one of the important causes of wire fires.Accurate detection of different types of fault arcs is of great significance to avoid major fire accidents.However,the complexity and concealment of arc faults bring great challenges to detection methods.Techniques based on threshold and electric current feature extraction are difficult to comprehensively generalize the characteristics of arc faults,while most methods based on deep neural networks directly perform feature learning on current signals and ignore the frequency information in the signal,resulting in poor generalization.In this regard,this paper proposes a fault arc identification method with a two-channel time-frequency convolutional neural network based on time-frequency feature learning,and designs a learnable discrete wavelet transform to extract multi-scale features in one-dimensional signals.Meanwhile,time fourier transform is employed to obtain two-dimensional time-frequency feature.Then convolutions are performed on those two channels’features,respectively for further feature fusion and prediction.Experiments are carried out on the arc current signals obtained under the three types of working conditions to verify the effectiveness of the proposed method.The experimental results show that the method has higher arc recognition accuracy compared with several competitive methods,reaching 97.91%.