Self-evolutionary control architecture based on distributed brain-inspired decision making and privacy-preserving PID

Abstract

Traditional PID control suffers from limited parameter adaptability in complex dynamic systems and poor robustness against nonlinearities and uncertain disturbances. Meanwhile, in distributed power control scenarios, the risk of leakage regarding multi-node sensor data and AI model parameters hinders the practical implementation of brain-inspired intelligent control. Therefore, this paper proposes a privacy-preserving, distributed, brain-inspired DQN-PID self-evolving control architecture; it establishes an integrated operational logic for perception and regulation, which can enable dynamic cognition of control objectives and the prioritization of decision-making. A PID parameter self-evolution mechanism is designed, which can enable online learning and optimization of parameters through brain-like synaptic plasticity simulation. A multi-level feedback loop is constructed. Furthermore, a federated learning training framework and parameter differential privacy protection strategy for the distributed brain-inspired DQN model are designed. While achieving "data remains stationary while the model moves," lightweight encryption and privacy perturbation mechanisms ensure secure cross-node information interaction, which can achieve the dual goals of optimizing control performance and ensure the security of distributed AI model data. A distributed simulation platform based on Matlab/Simulink is built, and a distributed nonlinear coupled system and a distributed industrial temperature control object are selected for verification. The engineering practicality and data security protection capabilities of the architecture are verified through distributed hardware-in-the-loop experiments. Experimental results show that the proposed architecture outperforms the comparative algorithms in dynamic response speed, steady-state accuracy, and anti-interference capability. Compared to traditional PID, the rise time is reduced by 35.6%-42.5% and the settling time by 36.7%-42.1%. Compared to fuzzy PID, the rise time is reduced by 23.7%-28.1% and the settling time by 26.9%-30.4%. The convergence time for parameter self-evolution is shortened by 37.5% relative to the original brain-inspired PID architecture, and the steady-state error is reduced by more than 45%. Following the implementation of privacy protection strategies, the loss in control performance for the distributed model is less than 5%, while the risk of raw data leakage is reduced by over 90%.  This approach achieves a balance among control performance, computational overhead, and data security in industrial real-time control scenarios, which can offer a new solution for the secure, high-precision adaptive control of complex power systems.