From Lab to Line: Deployment-Aware NMR–Text Expert Routing for Real-Time Apple Moldy Core Disease Screening and Explanation

Yang Lv, Yin Li, Yehui Han

Keywords: Nuclear Magnetic Resonance, Multimodal Dataset, Apple Quality Assessment, Lightweight Diagnostic Model, Real-time Inference

TL;DR: NMR–Text Expert Routing for Real-Time Apple Moldy Core Disease Screening and Explanation

Abstract: Real-time, interpretable diagnosis of Apple Moldy Core Disease (AMCD) under industrial sorting constraints is addressed. AppleNMR-MM V1.0, an LF-NMR–centric dataset an expert textual descriptions (n = 237), is introduced to enable multi-modal learning in a small-sample regime. A Task-Aware Mixture-of-Experts (T-MoE) fusion is proposed to route among NMR and text experts conditioned on predictive uncertainty and compute budget, while a Multi-agent Collaborative Chain-of-Thought (MACCT) with retrieval-augmented generation (RAG) coordinates triage→diagnose→explain agents using a domain corpus of SOPs, pathology notes, and batch logs for evidence-grounded reasoning. To align modeling with production constraints, a new metric, TAAPM, is introduced as the primary deployment criterion. Unlike prior evaluation schemes, TAAPM is explicitly derived from real factory export-trade regulations and uniquely optimized for overall economic benefit. On AppleNMR-MM V1.0, T-MoE attains AUC = 0.863 and F1 = 0.750, exceeding the strongest single-modality baselines by +5.8–6\% AUC; TAAPM reaches 972.84, indicating favorable accuracy–latency Pareto efficiency for in-line screening. The RAG-enabled explainer achieves a 92\% expert-check pass rate and 4.07 / 5 explanation quality; Collectively, AppleNMR-MM V1.0, T-MoE, TAAPM, and RAG-driven multi-agent reasoning establish a practical foundation for trustworthy AMCD screening and explanation on production lines.

Primary Area: datasets and benchmarks

Submission Number: 11343