The identification of lightning return stroke waveform plays an important role in the accurate location of lightning. Very low frequency/low frequency (VLF/LF) electromagnetic signals are widely used in lightning detection systems. The signals in this frequency band have a long propagation distance and are prone to attenuation distortion, which makes it more possible to misclassify the return stroke waveforms when using methods based on some specific waveform characteristic. The convolutional neural network is robust enough and performs well at discovering hidden patterns in images. In this paper, the residual convolutional neural network model is trained to obtain the waveform classifier. The lightning waveform data dataset is collected by lightning electric field measuring meters deployed in various provinces. After training, the classification accuracy of this classifier reaches 97.2% in the test set, and the accuracy reaches 86.75% using the traditional method based on waveform characteristics. The result proves the superior performance of the residual convolutional neural network model. By exploring the interpretability of the model, it is also proved that the better performance of the convolution network model comes from making full use of waveform information.

convolutional neural network;VLF/LF lightning waveform;model interpretability