An Efficient Lightning Classifier Using a Self‐Supervised Learning Neural Network

Abstract The automatic classification of lightning discharge processes is a critical challenge in both lightning physics research and protection. Despite recent advancements in artificial intelligence‐based classification models, rely on supervised learning, which demands extensive manually labeled samples and data preparation. To overcome these limitations, this paper proposes a self‐supervised neural network based on the masked autoencoder framework. The model first undergoes self‐supervised pretraining with a large amount of unlabeled lightning waveform to capture general signal features by reconstructing the masked parts, followed by supervised finetuning with minimal labeled data to further optimize its performance on classification task. The model achieves 98.30% accuracy on the Beijing Broadband Lightning NETwork data set. By applying the model to two public data sets, it demonstrates competitive performance (97.94% and 98.29%) with significantly less labeled data. T‐distributed stochastic neighbor embedding visualizations confirm that the combination of the pretraining and finetuning stages is critical for achieving optimal performance.