A Multimodal Large Language Model Architecture for Civil Aviation Fault Diagnosis
彭焕春 1晓鹏 1周庆钱 1方彦博 2刘青松 3林义闽 3李大中3
作者信息
1. 中国商飞客服
2. 中国联通数智
3. 联通数智
折叠
Abstract
Objective: To develop a multimodal large language model (MLLM) for civil aviation fault diagnosis and to evaluate its performance against its unadapted backbone without overstating comparative evidence.
Methods: The proposed Qwen3-32B-AeroMaint architecture combines domain-adaptive vocabulary construction, Ring Attention with rotary-position extension, a CNN–ViT visual encoder with cross-attention, supervised fine-tuning with Low-Rank Adaptation, Direct Preference Optimization (DPO), and structured chain-of-thought reasoning. Evaluation used general-capability benchmarks, an aviation benchmark, and a self-constructed set of 2,000 diagnostic question–answer pairs. A base-model comparison and leave-one-component-out ablation were conducted. Wilson 95% confidence intervals were calculated for proportions measured on the 2,000-case set.
Results: Within the controlled base-model comparison, the full model achieved a Top-1 root-cause accuracy of 0.73 (95% CI [0.710, 0.749]), compared with 0.23 (95% CI [0.212, 0.249]) for the base model. Top-3 coverage increased from 0.49 to 0.85, and procedure-generation ROUGE-L increased from 0.38 to 0.74. Removing multimodal fusion yielded a Top-1 accuracy of 0.71 (95% CI [0.690, 0.729]). The results therefore distinguish a substantial overall domain-adaptation improvement from a small, statistically unresolved fusion-only contrast; they do not establish superiority over external models.
Limitations: The evaluation lacks external competitive baselines, repeated-run variance, independent-site validation, and item-level paired predictions for McNemar testing. Long-context utilization and explanation faithfulness were not directly measured.
Conclusions: The distinctive contribution is the integration of long-document processing, cross-modal evidence fusion, and step-level preference alignment within one aviation maintenance architecture, together with an evaluation design that separates overall adaptation from component-level evidence. This design may inform other document-intensive, high-risk diagnostic domains. Its generalizability and operational value nevertheless require controlled external comparisons, direct multimodal subgroup analysis, and independent-site validation before deployment as human-supervised decision support.
Key words
civil aviation/ fault diagnosis/ multimodal large language model/ cross-modal fusion/ preference optimization
引用本文复制引用
彭焕春,晓鹏,周庆钱,方彦博,刘青松,林义闽,李大中.A Multimodal Large Language Model Architecture for Civil Aviation Fault Diagnosis[EB/OL].(2026-11-17)[2026-11-18].https://sinoxiv.napstic.cn/article/26500284.