TimeGAN and Coordinated Attention Prototype Network Based Prediction Model for Infrasound Signal

Downloads

Authors

  • Quanbo Lu Chongqing Polytechnic University of Electronic Technology, China
  • Xiaojuan Huang Chongqing Three Gorges University, China
  • Rao Li Chongqing Polytechnic University of Electronic Technology, China
  • Mei Li China University of Geosciences Beijing, China
  • Dong Zhu Sevnce Robotics Co, Ltd, China

Abstract

Due to the complexity of the infrasound environment and the high costs associated with data collection, frequent acquisition of infrasound data is often impractical, resulting in a limited amount of labeled data. To address the challenge of low classification prediction accuracy caused by data scarcity, this paper proposes an infrasound prediction model based on a Time-series Generative Adversarial Network (TimeGAN) and Coordinated Attention Prototype Network (CAPN) (TimeGAN-CAPN). The model begins by introducing TimeGAN, where the generative network is trained using a combination of unsupervised and supervised learning. This approach enables the network to operate within the latent space of temporal features and generate time-series data that closely aligns with the distribution of the original data. These generated samples are then combined with the original data to form an augmented dataset. Subsequently, the augmented data is input into the CAPN, which enhances the sample size per class, allowing for more precise class prototypes and improving the prediction accuracy of the model. Furthermore, the quality and diversity of the data generated by TimeGAN are quantitatively and qualitatively assessed using Maximum Mean Discrepancy (MMD) and t-Distributed Stochastic Neighbor Embedding (t-SNE), facilitating a comparison and verification of the generated data's performance. Experimental results show that TimeGAN-CAPN significantly outperforms the CAPN model in classification tasks with limited infrasound data, achieving a 7.15% increase in accuracy. This demonstrates that the proposed method is highly effective for predicting infrasound-related disasters, particularly in scenarios with small sample sizes.

Keywords:

infrasound signal, time-series generative adversarial network, coordinated attention prototype network, maximum mean discrepancy