Application research of convolutional neural network and its optimization in lightning electric field waveform...

Area of Science:

• Atmospheric Physics
• Computer Science
• Signal Processing

Background:

• Accurate identification and classification of lightning waveforms are crucial for lightning forecasting and early warning systems.
• Existing methods may be limited by noise and require optimization for improved performance.

Purpose of the Study:

• To design and implement a Convolutional Neural Network (CNN) algorithm for recognizing and classifying lightning electric field (LEF) waveforms.
• To optimize the CNN model from dataset, parameter, and network structure perspectives to enhance recognition accuracy.
• To investigate the impact of optimization strategies on model training time and efficiency.

Main Methods:

• Utilized electric field observation data from the Beijing lightning location website.
• Developed and implemented a CNN model for pulse signal waveform recognition.
• Optimized the CNN network by adjusting dataset, model parameters, and network structure.
• Studied the effects of various optimization terms and their order on training time.

Main Results:

• Achieved a lightning waveform recognition rate exceeding 90% with the optimized CNN model.
• Demonstrated that noise preprocessing, specifically fitting idealized waveforms, significantly improves recognition accuracy.
• Found that optimization substantially impacts training efficiency, increasing training time by approximately 51%, with optimization order having a negligible effect.

Conclusions:

• The CNN algorithm is well-suited for classifying and recognizing lightning electric field (LEF) waveforms.
• Optimization techniques are essential for significantly improving recognition rates.
• Noise preprocessing is a critical step for enhancing the performance of waveform recognition systems.