Communication-Efficient Personalized Federated Learning for Digital Twin in Heterogeneous Industrial IoT

As an up-coming digitalization technology, the digital twin (DT) offers a viable implementation for dynamic perception and intelligent decision-making in the industrial IoT (IIoT). For synchronizing the real-time information between the device and its DT, communication is fundamental to the digital twin system. Federated learning (FL) based DT model framework could be seen as an emerging paradigm to avoid large communication loads and high data leakage. However, the current digital twin model constructed scheme based on FL is not suitable for the heterogeneous IIoT scenario. Due to the different data distribution and different tasks among the devices, it leads to severe performance degradation when the personalized requirements of the DT model are ignored. In this paper, we propose a DT model framework based on personalized federated learning (PFL) to perform well for individual devices. Considering the historical personalized knowledge forgetting problem, the personalized federated learning with self-knowledge distillation (DTPFLsd) algorithm is proposed to avoid severe performance degradation and unnecessary time consumption of DT modeling. The numerical results compare with state-of-the-art FL-based DT architecture called DTWN and demonstrate the effectiveness and robustness of the proposed DTPFLsd scheme.