Parameter Inversion Algorithm for Complex Systems Integrating Deep Physical Networks and Optoelectronic Information Sensing

The research background of this study is the need for parameter inversion in complex systems deeply coupled with deep physical networks and optoelectronic information sensing technologies. As a theoretical innovation, this study constructs a unified mathematical representation framework and proposes a Deep Physical-Optoelectronic Reinforcement Parameter Inversion (DPORPI) algorithm. First, this study abstracts the system as a multi-layered heterogeneous graph composed of physical nodes, optoelectronic nodes, heterogeneous links, security levels, and computing resources. This innovation can uniformly characterize the system's interconnected structure across scales, modalities, and topologies. Based on this, a Joint-PGCN (Physical-Optoelectronic Joint Graph Convolutional Network) is designed. Simultaneously, a dual-domain convolution operator and a physical coupling attention term are introduced to simultaneously extract multi-domain interconnected features. A cross-domain mapping path is constructed through a Service Function Chain (SFC) constraint mechanism, implementing structured constraints for parameter transfer under five conditions: node consistency, computing resources, optoelectronic bandwidth, security level, and physical-optoelectronic dual-domain consistency. Finally, this study constructs a reinforcement learning optimization module. This module can jointly model the inverted state space, policy actions, and reward function. This optimization enables the inversion path to maintain stable convergence in dynamic and complex environments. This work used a complicated system with multi-source sensor fusion to build a test set in the experimental portion. According to the findings, DPORPI has an inversion root mean square error of 0.029 at$\text{SNR}=20 ~\text{dB}$, which is 61.3%, 50.0%, and 29.3% less than GCN (0.075), PINN (0.058), and Joint-PGCN (0.041), in that order. DPORPI maintains an RMSE of 0.062 under the low-quality observation condition of$\text{SNR}=5$dB, which is 61.7% less than GCN (0.162). Furthermore, DPORPI outperforms Joint-PGCN (84.7%) and PINN (79.2%) in cross-scenario testing. This maintains an average inversion stability of almost 92.1% under the three types of the system structure modifications.