A Novel LLM-Enabled Neural Architecture Search Method for Compositional Generalization Fault Diagnosis

Abstract

Neural architecture search (NAS) is increasingly important for reducing manual trial-and-error in deep fault diagnosis model design. This demand is even more pronounced in system-level compound fault diagnosis, where heterogeneous subsystems, cross-component coupling, and unseen fault compositions make architecture design particularly difficult. To address this challenge, we propose a large language model (LLM) guided multi-model collaborative NAS framework for fault diagnosis with compositional generalization. The framework adopts a bilevel optimization strategy: the inner loop learns subsystem-specific representations with dedicated encoders and a global fusion module, while the outer loop uses a LLM to refine architecture configurations based on structured validation feedback under constrained one-step updates. Experiments on the BJTU-RAO metro bogie system-level dataset show that the proposed method outperforms representative baselines and improves compositional generalization accuracy (CGA) over stochastic search, achieving 87.42% CGA versus 83.24% for random-search NAS.