Mutual Information Optimization Using Large Language Model for Domain Generalized Remaining Useful Life Prediction in IIoT

The Industrial Internet of Things (IIoT) enables interoperability among networked machines, driving the need for Remaining Useful Life (RUL) prediction to ensure operational safety. However, substantial variations in machine specifications introduce severe data distribution discrepancies. Existing research predominantly handles intra-machine operating conditions, while variations across machines remain largely underexplored. To address this challenge, a novel domain generalized framework is proposed for RUL prediction across machines in IIoT. The core idea is to leverage a renovated large language model, acting as an advisor, to generalize a temporal diffusion model through mutual information optimization. Specifically, the temporal diffusion model is first designed to capture degradation dynamics by forward simulating degradation with stochastic noises and then backward eliminating noises via conditioned denoising over time. Then, the renovated large language model is constructed to provide domain-agnostic knowledge by encoding engineered textual prompts and temporal sensor embeddings via the multimodal knowledge fusion. Finally, a mutual information optimization objective is proposed to integrate individual models by seamlessly amalgamating temporal degradation dynamics with domain-agnostic knowledge, offering theoretical guarantees and enabling generalization to unseen machines. Extensive experiments on four bearing datasets validate the superiority, demonstrating its ability to deliver accurate RUL predictions across machines in IIoT.