Efficient Signal Representation Learning for Communication and Sensing Systems

Abstract

The convergence of communication and sensing paradigms within modern wireless networks demands unprecedented capabilities in processing high-dimensional data streams. As future networks transition toward integrated sensing and communication architectures, traditional signal processing algorithms face severe limitations regarding computational complexity, adaptability, and resource utilization. This paper explores the development and application of efficient signal representation learning frameworks designed to extract robust, low-dimensional features from complex radio frequency signals. By leveraging advanced machine learning techniques, specifically tailored for resource-constrained environments, the proposed methodologies significantly reduce the overhead associated with signal reconstruction, modulation classification, and target detection. We provide a comprehensive review of the theoretical foundations governing signal representation, followed by a detailed methodological framework that eliminates the need for hand-crafted feature extraction. Through extensive simulations and theoretical analysis, we demonstrate that optimized representation learning models can achieve superior bit error rates in communication tasks while simultaneously enhancing range and Doppler resolution in sensing applications. The findings indicate that implementing structured sparsity and quantization within the learning pipeline yields substantial gains in inference speed without compromising accuracy. This research paves the way for scalable, intelligent transceivers capable of supporting the stringent latency and reliability requirements of next-generation autonomous systems and smart infrastructures.