Tao Zhang
School of Integrated Circuits and Electronic Information, Zibo Polytechnic University, Zibo 255300, Shandong, China.
*Corresponding author:Tao Zhang
Abstract
Abnormal perception in edge IoT environments faces triple constraints of limited resources, dynamic and variable data, and high real-time response requirements. Traditional methods are difficult to balance detection accuracy and inference efficiency. This article systematically studies the method system of empowering edge IoT anomaly perception and lightweight reasoning with large models. Firstly, the intrinsic mechanism of edge adaptation was analyzed, revealing the technical bottlenecks in model capacity, computational energy consumption, and dynamic adaptability when the inference advantages of large models migrate to the edge, as well as the dialectical unity logic between accuracy and real-time performance under resource constraints. Furthermore, from the perspective of perceptual empowerment, lightweight translation based on knowledge distillation, anomaly representation learning under multimodal alignment, context-aware driven threshold dynamic adjustment, and edge cloud collaborative perception task offloading criteria were proposed. In terms of a lightweight inference mechanism, model structure pruning and attention sparsity, mixed precision energy consumption control, dynamic selection of inference path, and update forget balance strategy under continuous learning have been designed. Through the collaborative optimization of architecture design and operation mechanism, this method aims to achieve effective unity of anomaly perception accuracy and real-time performance on resource-constrained edge nodes, providing theoretical support for intelligent IoT systems.
References
[1] Sun DJ. Design of optimized deployment scheme for edge collaborative sensing nodes in water conservancy IoT. China Inf. 2026;(5):169-170.
[2] Lin JQ. IoT-based multi-heterogeneous network security intrusion detection technology. Inf Rec Mater. 2026;27(1):151-153,156. doi:10.16009/j.issn.1009-5624.2026.01.049.
[3] Gao Y, Dong W, Xiao KJ, et al. Large-model-driven automated IoT application development technology. Computing. 2025;1(9):31-37.
[4] Han Y. Intelligent monitoring and fault diagnosis of industrial equipment based on large-model and IoT integration. Intell IoT Technol. 2025;57(6):81-85.
doi:10.26921/j.cnki.2096-6059.2025.06.018.
[5] Liu TW. Lightweight transmission technology for industrial Internet data security based on IoT. China Broadband. 2025;21(11):7-9.
doi:10.20167/j.cnki.ISSN1673-7911.2025.11.03.
[6] Yue W. Design study of lightweight IoT management system based on microservice architecture. Inf Comput. 2025;37(15):188-190.
[7] Xue FY, Zhou YL, Li JK, et al. Review on large-model technology applications in smart aquaculture IoT and edge computing. Trans Chin Soc Agric Mach. 2025;56(9):291-311.
[8] Li Q. Edge computing and lightweight AI model deployment in power IoT. Guangxi Electr Power. 2025;48(3):51-58. doi:10.16427/j.cnki.issn1671-8380.2025.03.005.
[9] Yu XB, Zhang SH, Zhang HL. Cloud-edge collaborative evolution model and communication paradigm for low-altitude IoT. J Internet Things. 2024;8(3):76-90.
[10] Xu XH, Zhou CB, Hu ZX, et al. Review of lightweight deep neural network model adaptation for edge intelligence. Comput Sci. 2024;51(7):257-271.
[11] Wu D, Zhao L, Kim B, et al. Dynamic energy-aware fixed-point linear mapping multiplier for internet of things edge devices. Sci Rep. 2026. doi:10.1038/s41598-026-53462-y.
[12] Upadhyaya B, Panda CA. Matrix-based blockchain-integrated certificateless signature scheme for secure industrial internet of things edge device authentication. Comput Electr Eng. 2026;136:111227. doi:10.1016/j.compeleceng.2026.111227.
How to cite this paper
Research on Empowering Edge IoT Anomaly Perception and Lightweight Reasoning Methods with Large Models
How to cite this paper: Tao Zhang. (2026) Research on Empowering Edge IoT Anomaly Perception and Lightweight Reasoning Methods with Large Models. Journal of Applied Mathematics and Computation, 10(2), 98-102.
DOI: http://dx.doi.org/10.26855/jamc.2026.06.004