针对机载系统部署大模型实时性差、资源占用高及语义结构信息易丢失等问题，提出一套轻量化意图理解部署优化框架。该框架从三个层面协同解决问题：（1）算法层面，融合知识蒸馏与量化感知训练，引入意图-槽位一致性损失与跨层语义对齐损失，在压缩过程中保留结构化语义；（2）部署层面，针对嵌入式平台特性，实现从ONNX模型转换、图优化、算子适配到动态推理引擎的全链路部署优化；（3）自适应层面，引入基于MAML的在线元学习，支持小样本场景下快速适应语义漂移。在CMNEE和UAV-Inst数据集上的实验结果表明，语义框架准确率分别达到90.59%和88.19%，相比通用压缩方法DistilBERT分别提升5.25和3.56个百分点；在Orange Pi 5 Plus边缘平台上，模型体积压缩至127.62MB，较BERT-base减小69.4%，推理延迟降至9.7ms，能耗降低62.7%。消融实验验证了各结构感知损失模块的有效性，元学习实验验证了少样本快速自适应能力，部署实验验证了全链路优化的工程可行性。研究结果表明，所提方法在保持高精度的同时，有效提升了机载环境下的运行效率，为空天智能系统的人机协同提供了可行方案。
[1]CHEN Q, ZHUO Z, WANG W. Bert for joint intent classification and slot filling[J]. arXiv prep.[J].rXiv:1902.10909, 2019., rint, :-
[2] DEVLIN J, CHANG M W, LEE K, et al.BERT: Pre-training of deep bidirectional transformers for language understanding[C]//Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. 2019: 4171-4186.
[3]WANG X, LI Q, JIA W. Cognitive Edge Computing: A Comprehensive Survey on Optimizing Large Models and AI Agents for Pervasive Deployment[J]. arXiv prep.[J].rXiv:2501.03265, 2025., rint, :-
[4]PATTERSON D, GONZALEZ J, LE Q, et al. Carbon emissions and large neural network training[J]. arXiv prep.[J].rXiv:2104.10350, 2021., rint, :-
[5]HINTON G, VINYALS O, DEAN J. Distilling the knowledge in a neural network[J]. arXiv prep.[J].rXiv:1503.02531, 2015., rint, :-
[6] JACOB B, KLIGYS S, CHEN B, et al.Quantization and training of neural networks for efficient integer-arithmetic-only inference[C]//Proceedings of the IEEE conference on computer vision and pattern recognition. 2018: 2704-2713.
[7]ZHU M, GUPTA S. To prune, or not to prune: exploring the efficacy of pruning for model compression[J]. arXiv prep.[J].rXiv:1710.01878, 2017., rint, :-
[8]Ganesh P, Chen Y, Lou X, et al.Compressing large-scale transformer-based models: A case study on bert[J].Transactions of the Association for Computational Linguistics, 2021, 9(无):1061-1080
[9] RAYMOND C, RICCARDI G.Generative and discriminative algorithms for spoken language understanding[C]//Interspeech 2007-8th Annual Conference of the International Speech Communication Association. 2007.
[10]LIU B, LANE I. Attention-based recurrent neural network models for joint intent detection and slot filling[J]. arXiv prep.[J].rXiv:1609.01454, 2016., rint, :-
[11]SANH V, DEBUT L, CHAUMOND J, et al. DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter[J]. arXiv prep.[J].rXiv:1910.01108, 2019., rint, :-
[12] JIAO X, YIN Y, SHANG L, et al.Tinybert: Distilling bert for natural language understanding[C]//Findings of the association for computational linguistics: EMNLP 2020. 2020: 4163-4174.
[13]FRANTAR E, ASHKBOOS S, HOEFLER T, et al. Gptq: Accurate post-training quantization for generative pre-trained transformers[J]. arXiv prep.[J].rXiv:2210.17323, 2022., rint, :-
[14]LIN J, TANG J, TANG H, et al.Awq: Activation-aware weight quantization for on-device llm compression and acceleration[J].Proceedings of machine learning and systems, 2024, 6(无):87-100
[15]LAGUNAS F, CHARLAIX E, SANH V, et al. Block pruning for faster transformers[J]. arXiv prep.[J].rXiv:2109.04838, 2021., rint, :-
[16]DUBIEL M, BARGHOUTI Y, KUDRYAVTSEVA K, et al.On-device query intent prediction with lightweight LLMs to support ubiquitous conversations[J].Scientific reports, 2024, 14(1):12731-12731
[17] SARKAR S, BABAR M F, HASSAN M M, et al.Processing Natural Language on Embedded Devices: How Well Do Transformer Models Perform?[C]//Proceedings of the 15th ACM/SPEC International Conference on Performance Engineering. 2024: 211-222.
[18] WU Z, LIU Z, LIN J, et al.Lite Transformer with Long-Short Range Attention[C]//International Conference on Learning Representations. 2020.
[19] SARKAR S, BABAR M F, HASSAN M M, et al.Processing Natural Language on Embedded Devices: How Well Do Transformer Models Perform?[C]//Proceedings of the 15th ACM/SPEC International Conference on Performance Engineering. 2024: 211-222.
[20]WANG R, ZHOU Y, CHEN L.Adaptive deployment of lightweight language models in dynamic edge environments[J].IEEE Transactions on Mobile Computing, 2024, 23(4):3125-3138
[21]Johansson S, Riihonen T. On the Military Applications of Large Language Models[J]. arXiv prep.[J].rXiv:2511.10093, 2025., rint, :-
[22] Zhu M, Xu Z, Zeng K, et al.CMNEE: a large-scale document-level event extraction dataset based on open-source Chinese military news[C]//Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024). 2024: 3367-3379.
[23]Chenghai LI, Ke W, Yafei S, et al.Air target intent recognition method combining graphing time series and diffusion models[J].Chinese Journal of Aeronautics, 2025, 38(1):103177-103177
[24]Hou Y, Mao J, Lai Y, et al. Fewjoint: A few-shot learning benchmark for joint language understanding[J]. arXiv prep.[J].rXiv:2009.08138, 2020., rint, :-
CJA