Design and Evaluation of Self-Healing Networks Using Q-Learning and Digital Twin Technology

The expanding complexity of communication networks nowadays requires reliability, resilience, and quick fault recovery as must-have features for future communication infrastructures. Current fault management solutions are largely manual-based or rule-based static systems, and cannot adequately deal with dynamic and massive network faults. Thus, the purpose of this research is to design and test an intelligent self-healing network framework based on the integration of Reinforcement Learning (RL) and the Digital Twin technology for autonomous, adaptive, and real-time self-healing. The main idea of the proposed system is to design an optimal repair strategy learning agent based on Q-learning that continuously interacts with a simulated network environment and to continuously simulate the physical network with the Digital Twin to predict potential failures according to the historical data and real-time data of the network. The architecture proposed consists of several processing modules, including a network simulator that supports various network topologies including small-world, random, adaptive, scale-free, complete, star and grid networks, a node and edge condition type monitoring and prediction module based on Digital Twins, an ε-greedy Q learning algorithm for making adaptive decisions in the repair process, and a composite health score mechanism based on node and edge conditions. The system also utilizes failure prediction models, historical data management, and reward optimization techniques for enhancing learning convergence and network resilience. The criteria used for the evaluation are: network health, mean time to repair (MTTR), break detection accuracy, false positive rate, network availability, and repair success. Experimental results show that the proposed RL–Digital Twin framework significantly outperforms the traditional repair method of doing nothing, random repair, and a rule-based repair method, with an average network health of 89.4%, an average repair time of 3.1 steps, and a failure prediction accuracy of 91.2%. The RL agent can adapt to various network topologies and converge in 65–100 training episodes. Despite these benefits, the study has been restricted to simulated networks up to 30 nodes, and mainly considers operational failures as opposed to cybersecurity attacks. In addition, scalability challenges remain for larger networks due to the limitations of traditional Q-learning methods.