loading page

Towards an Interpretable CNN Model for the Classification of Lightning Produced VLF/LF Signals
  • +5
  • Lilang Xiao,
  • Weijiang Chen,
  • Yu Wang,
  • Kai Bian,
  • Zhong Fu,
  • Nianwen Xiang,
  • Hengxin He,
  • Yang Cheng
Lilang Xiao
Huazhong University of Science and Technology
Author Profile
Weijiang Chen
State Grid Corporation of China
Author Profile
Yu Wang
State Grid Electric Power Research Institute, Wuhan 430074, China
Author Profile
Kai Bian
State Grid Corporation of China, Beijing,
Author Profile
Zhong Fu
Electric Power Research Institute, State Grid Anhui Electric Power Company
Author Profile
Nianwen Xiang
School of Electrical Engineering and Automation, Hefei University of Technology
Author Profile
Hengxin He
Huazhong University of Science of Technology

Corresponding Author:[email protected]

Author Profile
Yang Cheng
Electric Power Research Institute, State Grid Anhui Electric Power Company
Author Profile

Abstract

Classification of lightning produced VLF/LF signals plays crucial role in the detection and location of lightning flashes. The machine learning method has potential in the VLF/LF lightning signal classification. Traditional machine learning methods are data-driven and work in a black-box fashion, making the classification accuracy highly dependent on the size and quality of dataset. In this paper, an interpretable convolutional neural network model is proposed for VLF/LF lightning electric field waveform classification. Multi-scale convolutional kernels and shortcut connections are adopted in this model to enhance the ability to capture local waveform features. The CAM method is embedded in our model to open the black-box by visualizing the weight of different waveform features on the classification results. Based on the measured data from five different provinces in China, an accuracy of 98.5% is achieved in a four-type classification task including RS, active stage of IC, PB and NB. The correlation between the weight values of different waveform features and corresponding lightning discharge process are analyzed. It is found that the proposed model can extract decisive features of VLF/LF lightning signals closely related to the physical process of lightning discharges, which is similar to the human expert’s behavior. The proposed model is validated by using an open-source dataset from Argentina. It is indicated that the proposed model can resist the impact of unexpected waveform oscillation and achieve a higher accuracy of 98.39% than that of the support vector method. It is demonstrated that our model is less dependent on the training dataset.
21 Jan 2023Submitted to ESS Open Archive
24 Jan 2023Published in ESS Open Archive