An Alternative Implementation of Multivariate-Teaching-Based Transformer for High-Power Microwave Problems With Plasma Array

High-power microwave (HPM) protection for military weaponry demands exceptional sensitivity and reliability. Columnar plasma array (CPA), reflected by adaptable electromagnetic properties and effective energy absorption, offers a promising solution to these rigorous demands. However, the ionization of argon gas to a plasma state, initiating electron avalanche effects for electromagnetic shielding, is not always achievable with all forms of HPM radiation. Therefore, the ability to perform HPM inversion, forecast plasma electron density, and electric intensity, and deduce missing data points becomes paramount. While time series models (TSMs) have progressed considerably, their application to inversion problems remains largely undeveloped. In addition, the complex interactions of HPM with CPAs have seldom been evaluated utilizing neural networks. This work introduces a novel TSM, the multivariate-teaching-based transformer (MTT), based on an encoder-decoder architecture and incorporating our uniquely developed teaching multivariate attention (TMA). When applied to HPM protection, the MTT exhibits significant improvements over the existing iTransformer, achieving remarkable accuracy and greater stability with a reduced parameter count and lower memory consumption for training.