Knowledge-grounded multimodal cognitive decision-making for aerospace machining: A physics-aware reasoning framework for generative process planning of special-shaped features

Abstract

Complex aerospace products encompass numerous special-shaped features (SSFs), whose machining processes rely on highly customized process planning. Traditional data-driven discriminative models predominantly operate as physics-agnostic "black boxes." Moreover, applying general-purpose Multimodal Large Language Models (MLLMs) to manufacturing often causes "physical hallucinations", generating linguistically plausible but mechanically infeasible process plans. To bridge this gap, this paper proposes a Knowledge-Grounded Multimodal Cognitive Framework (KGMCF), which extends discriminative feature recognition with physics-aware generative reasoning for machining decision-making. First, an Aerospace Multimodal Process Knowledge Graph (Aero-MPKG) is constructed to formalize manufacturing constraints. A cross-modal visual anchoring mechanism is introduced, enabling the MLLM to precisely map unstructured engineering drawings to discrete symbolic nodes via topological embedding retrieval. Second, a KG-guided Symbolic-Neural Synthesis (SNS) strategy is developed to construct a multimodal instruction-tuning dataset, thereby driving a multi-stage physics-aware Chain-of-Thought (CoT). Finally, a symbolic verification loop with semantic traceback detects constraint-violating parameters and triggers self-correction through natural language feedback. Extensive experiments on an aerospace vortex shell benchmark demonstrate that the KGMCF effectively suppresses the Graph-based Physics Violation Rate (G-PvR) from 24.52% (vanilla MLLM) to 9.62%, while simultaneously achieving a Tool Selection Accuracy (TSA) of 93.12% and a Machining Parameter Selection Accuracy (MPSA) of 91.44%.