A MoE-LLM-based multisensor flexible fusion fault diagnosis method for rotating machinery

As a core component of industrial systems, rotating machinery requires accurate fault diagnosis to ensure safety and efficient operation. In recent years, although intelligent diagnostic methods based on Large Language Models (LLMs) have made notable progress, most existing approaches are limited to single-type signals and lack effective modeling of the complementary and synergistic information across multiple sensors, rendering them insufficient for complex and dynamic industrial environments. To address this limitation, this paper proposes a MoE-LLM-based multisensor flexible fusion fault diagnosis method for rotating machinery. The proposed method designs a multisensor embedding layer to map various combinations of sensor signals into unified feature embeddings compatible with LLMs. A sparsely activated Mixture of Experts (MoE) mechanism, comprising both uni-signal and fusion-signal experts, is introduced to enable adaptive modeling and fault identification from multisource signals. Additionally, a curriculum learning-based staged training strategy is developed to enhance the model’s transferability across different scenarios. Experiments conducted on three multisensor datasets demonstrate that the proposed method outperforms mainstream approaches in terms of diagnostic accuracy, robustness, and scalability.