On-Device Large Language Models: Architectures, Compression, and Personalization for AI-Empowered Intelligent Engineering Systems

Abstract

The Giant Language Model (LLM) on devices becomes very important for intelligent engineering systems (IES) using AI. In these systems, the network physical system (CPS) needs to make fast, private and reliable decisions at the edge. This review brings together the progress of architecture, compression technology and customization strategy, using system engineering (modularization and closed-loop control), CPS theory (close relationship between physics and network) and information theory (efficient semantic coding with low resources). The architecture from transformer uses parameter-saving design and hardware optimization to run in the memory and energy of edge devices. Using knowledge decomposition, post-training quantization and structured preprocessing to compress can reduce the size of the model, while keeping the importance of information flow for real-time control and diagnosis as much as possible. Through lightweight customization method, the model can be adjusted according to engineering rules and operators’ working methods without collecting all the data in one place. Real examples show how to use digital twins for predictive maintenance, reinforcement learning for dynamic planning of smart factories, computer vision for online fault detection and edge computing for closed-loop operation. Comparative advantages, such as lower delay, more energy saving and better durability, risks, such as lower accuracy, security issues in key CPS, and lower robustness when data changes. The original idea is to extend the information bottleneck framework to increase the dynamics of physical systems and the importance of decision-making. Finally, the review points out the ways to realize verifiability and adaptive intelligence on devices, which will make the upcoming IES more autonomous and sustainable.