Leveraging large models for domain knowledge-informed architecture search of fault diagnosis

Abstract

To leverage complex domain knowledge in mechanical fault diagnosis for neural architecture search (NAS) effectively, a domain knowledge-informed architecture search approach utilizing large models is proposed. First, a multi-dimensional, standardized architecture data model is constructed. This model is populated with prototype tensors, which are generated via adaptive wavelet packet decomposition. The embedding of these tensors results in a multi-domain representation of prior fault domain knowledge. Subsequently, a dual-channel fusion learning paradigm is designed, where domain knowledge is integrated through the fine-tuning of the objective function. This process involves the parallel execution of low-rank adaptation-based fine-tuning for a large model and topology extraction training for a graph neural network predictor. The resulting semantic and topological vectors are then aligned on Riemannian manifolds and embedded into an attention-gated network to form an expert routing strategy. Finally, dynamic pruning is applied based on the improved Shapley calculation. A differentiable weight-to-performance mapping is created and an adaptive pruning mechanism is applied, which incorporates distribution shift detection and a probability-difference-driven update. Across the unseen datasets, the proposed approach delivered relative improvements of 6.6%–18.6% in accuracy and 6.6%–19.4% in F1 score compared with existing state-of-the-art NAS methods. In addition, dynamic pruning reduced the search time from 10.89 h to 0.83 h. This provides a new paradigm for incorporating fault-domain knowledge into diagnostic model design.