Generative digital twin: A Sparse-xLSTM based self-evolution hybrid modeling approach for complex equipment

The integration of generative artificial intelligence (GenAI) into digital twins (DTs) offers transformative potential for characterizing the non-linear dynamics of complex equipment. However, realizing high-fidelity, self-evolution generative DTs remains challenging due to the computational prohibitiveness of Transformer based architectures and the insufficient long-term memory of traditional recurrent networks. To address these limitations, this paper proposes a novel generative digital twin (GDT) method driven by a Physics-Guided Sparse-xLSTM. First, we construct a Sparse-xLSTM architecture designed to model high-dimensional temporal dynamics efficiently. By leveraging exponential gating and matrix memory, this architecture achieves linear computational complexity while effectively capturing long-range dependencies, overcoming the quadratic bottleneck of conventional attention mechanisms. A sparse attention strategy is further incorporated to extract distinct state features from noisy sensor data, enhancing representation robustness. Second, to ensure physical consistency and generalization, a hybrid modeling mechanism is devised. Physical priors are embedded as input feature, and governing equations are integrated into the optimization objective as physical constraints, effectively confining the generative solution space to physically valid trajectories and mitigating model hallucinations. Third, addressing the imperative for continuous adaptation, a low-rank adaptation (LoRA)-based self-evolution engine is developed. This mechanism enables the GDT to perform online updates using limited streaming data by fine-tuning low-rank matrices, ensuring rapid synchronization with the time-varying states of the physical entity without catastrophic forgetting. Experimental validation on CNC machine tool wear datasets demonstrates that the proposed approach significantly outperforms advanced baselines in both prediction accuracy and adaptation efficiency, providing a robust paradigm for intelligent predictive maintenance of complex equipment.