Deep learning aided intelligent signal recognition for backscatter based metamaterial passive Internet of Things system

In this paper, a single point-frequency sensing mechanism-based metamaterial passive Internet of Things (IoT) system is proposed, where the sensing of the environmental physical quantities is realized by the reflection coefficients at fixed frequency points. Firstly, the time-domain pilot structure of the unit basis vector is designed, so that each pilot time slot can directly instantiate the corresponding column of the channel matrix after eliminating the link gain ambiguity, and provide a known operator for the linear inversion of the data time slot. Secondly, an end-to-end intelligent recognizer is proposed, where the robustness can be improved with residual compensation and gating fusion. Finally, data consistency and physical feasibility are ensured by lightweight residual refinement and interval projection. Simulations show that the proposed approach can reconstruct the reflection coefficients more accurately than the classical least squares method. In the practical spatial temperature field reconstruction task, the proposed method also demonstrates highly competitive reconstruction performance. Compared with the traditional broadband sweeping method, it exhibits stronger real-time performance and is more suitable for environments with rapidly changing temperature.