Innovation and Application Development of Signal Processing Technology Based on Artificial Intelligence (Deep Learning / Reinforcement Learning)

Abstract

This paper systematically reviews the technological innovation and application development of artificial intelligence (deep learning/reinforcement learning) in signal processing. Firstly, it sorts out the evolutionary path of deep learning models (such as CNN and Transformer) in signal feature extraction. It also analyzes the technological breakthroughs and limitations from local perception to global modeling. Secondly, it discusses the collaborative application of few-shot and transfer learning in medical signal processing and AI-driven industrial radar and communication signal processing cases. Finally, it focuses on three core issues: the real-time bottleneck of AI models in edge devices, improving signal classification accuracy in few-shot scenarios, and designing signal processing models with strong interpretability. It also proposes solutions and future research directions. Studies have shown that different models must be selected according to signal characteristics (e.g., CNN is suitable for locally stationary signals, and Transformer is ideal for long-time series signals). Integrating physical perception and data-driven approaches can improve the system's robustness. This paper provides a theoretical reference for the interdisciplinary integration of signal processing and AI and helps the technical implementation in fields such as 5G/6G and innovative healthcare.