面向MaaS模式的AI模型快速部署框架研究

doi:10.3901/JME.260401

摘要/Abstract

摘要： 以具身智能、大模型为代表的新一轮人工智能技术正在加速与制造业深度融合，推动制造业向高效、智能、柔性的方向转型升级。面对传统AI模型部署效率低、服务响应慢、资源利用率不高等问题，聚焦“AI驱动的工业应用快速构建部署”主题，提出一种面向“模型即服务”(Model as a service,MaaS)模式的AI模型快速部署框架。该框架采用云-边-端三层架构，针对AI模型标准化定义、封装解耦、边缘快速加载与部署、智能卸载调度，提出了基于容器的AI模型标准封装方法和基于深度强化学习的模型推理任务边缘卸载策略IF-DQN。实验与工程实践结果表明，所提出的框架模型封装率提升40%，镜像体积减小近60%，在不同网络条件下均能保持稳定的分发时延，智能卸载策略在多类任务分布下平均性能提升超过40%。本研究为制造互联与工业智能中的AI部署能力建设提供了可复制、可推广的工程实践方案。

关键词: 模型即服务(MaaS), 模型封装, 模型分发, AI推理任务, 任务卸载, 深度强化学习

Abstract: A new wave of AI technologies, represented by embodied intelligence and large models, is accelerating integration with manufacturing and driving transformation toward greater efficiency, intelligence, and flexibility. To address challenges such as low deployment efficiency, slow response, and poor resource utilization, a rapid AI model deployment framework based on the model-as-a-service (MaaS) paradigm is proposed. Built on a cloud–edge–device architecture, the framework supports standardized model definition, decoupled packaging, fast edge deployment, and intelligent task offloading. MaaSContainer is introduced, a Docker-based packaging format. And IF-DQN is introduced, a deep reinforcement learning–based offloading strategy. Experiments and engineering practice show that the framework achieves up to 40% encapsulation rate, reduces image size by nearly 60%, maintains stable distribution latency under varying networks, and improves task scheduling performance by over 40%. The proposed framework offers a replicable, scalable solution for AI deployment in industrial intelligence, supporting the digital and intelligent transformation of manufacturing.

Key words: model-as-a-service(MaaS), model encapsulation, model distribution, AI inference task, task offloading, deep reinforcement learning

中图分类号:

TP18
TH166

柳先辉, 赵俊龙, 朱城林, 余德水, 盛磊, 赵卫东. 面向MaaS模式的AI模型快速部署框架研究[J]. 机械工程学报, 2026, 62(11): 14-29.

LIU Xianhui, ZHAO Junlong, ZHU Chenglin, YU Deshui, SHENG Lei, ZHAO Weidong. Research on Rapid Deployment Framework of AI Models for MaaS Patterns[J]. Journal of Mechanical Engineering, 2026, 62(11): 14-29.

导出引用管理器 EndNote|Reference Manager|ProCite|BibTeX|RefWorks

链接本文: http://www.cjmenet.com.cn/CN/10.3901/JME.260401

http://www.cjmenet.com.cn/CN/Y2026/V62/I11/14

参考文献

[1] GAN W，WAN S，YU P S. Model-as-a-service (MaaS)：A survey[Z]. 2023 IEEE International Conference on Big Data (BigData). IEEE Computer Society. 2023： 4636-45.
[2] ABOU-FOUL M，RUIZ-ALBA J L，LÓPEZ-TENORIO P J. The impact of artificial intelligence capabilities on servitization：The moderating role of absorptive capacity-A dynamic capabilities perspective[J]. Journal of Business Research，2023，157：113609.
[3] SJöDIN D，PARIDA V，KOHTAMÄKI M. Artificial intelligence enabling circular business model innovation in digital servitization：Conceptualizing dynamic capabilities，AI capacities，business models and effects[J]. Technological Forecasting and Social Change，2023，197：122903.
[4] PALEYES A，URMA R G，LAWRENCE N D. Challenges in deploying machine learning：A survey of case studies[J]. ACM Computing Surveys，2023，55(6)：114.
[5] ZHAO S，TALASILA M，JACOBSON G，et al. Packaging and sharing machine learning models via the acumos AI open platform[C]// Proceedings of the 201817th IEEE International Conference on Machine Learning and Applications (ICMLA)，2018-12-01，2018.
[6] SALMANI M，GHAFOURI S，SANAEE A，et al. Reconciling high accuracy，cost-efficiency，and low latency of inference serving systems [C]// Proceedings of the 3rd Workshop on Machine Learning and Systems. Rome，Italy; Association for Computing Machinery. 2023：78-86.
[7] LIN Q R，WU S，ZHAO J B，et al. SmartLite：A DBMS-based serving system for DNN inference in resource-constrained environments[J]. Proceedings of the Vldb Endowment，2023，17(3)：278-291.
[8] MCDOUGAL R A，MORSE T M，CARNEVALE T，et al. Twenty years of modelDB and beyond：Building essential modeling tools for the future of neuroscience[J]. Journal of Computational Neuroscience，2017，42(1)：1-10.
[9] VAN RIJN J N，BISCHL B，TORGO L，et al. OpenML：A collaborative science platform[M]. Heidelberg：Springer，2013.
[10] WANG Y J，WANG Q，WANG N，et al. End-users’ acceptance of ’X as a service’：Evidence from agriculture 4.0[J]. Computers & Industrial Engineering，2024，197：110524.
[11] HU S，XIAO Y. Design of cloud computing task offloading algorithm based on dynamic multi-objective evolution[J]. Future Generation Computer Systems，2021，122：144-148.
[12] ISLAM A，DEBNATH A，GHOSE M，et al. A survey on task offloading in multi-access edge computing[J]. Journal of Systems Architecture，2021，118：102225.
[13] KUMARI N，YADAV A，JANA P K. Task offloading in fog computing：A survey of algorithms and optimization techniques[J]. Computer Networks，2022，214：109137.
[14] LIN L，LIAO X，JIN H，et al. Computation offloading toward edge computing[J]. Proceedings of the IEEE，2019，107(8)：1584-607.
[15] MA X，ZHAO Y，ZHANG L，et al. When mobile terminals meet the cloud：computation offloading as the bridge[J]. IEEE Network，2013，27(5)：28-33.
[16] WANG J，PAN J，ESPOSITO F，et al. Edge cloud offloading algorithms：Issues，methods，and perspectives[J]. ACM Comput Surv，2019，52(1)：Article 2.
[17] ANANTHANARAYANAN G，BAHL P，BODÍK P，et al. Real-time video analytics：The killer App for edge computing[J]. Computer，2017，50(10)：58-67.
[18] WANG X，NING Z，WANG L. Offloading in Internet of vehicles：A fog-enabled real-time traffic management system[J]. IEEE Transactions on Industrial Informatics，2018，14(10)：4568-4578.
[19] YANG S，HE Y，ZHENG X. FoVR：Attention-based VR streaming through bandwidth-limited wireless networks[C]// Proceedings of the 201916th Annual IEEE International Conference on Sensing，Communication，and Networking (SECON)，10-13 June 2019，2019.
[20] TRAN T X，POMPILI D. Joint task offloading and resource allocation for multi-server mobile-edge computing networks[J]. IEEE Transactions on Vehicular Technology，2019，68(1)：856-868.
[21] ZHANG T，CHIANG Y H，BORCEA C，et al. Learning-based offloading of tasks with diverse delay sensitivities for mobile edge computing[C]// Proceedings of the 2019 IEEE Global Communications Conference (GLOBECOM)，2019-12-01，2019. IEEE.
[22] QIU X，LIU L，CHEN W，et al. Online deep reinforcement learning for computation offloading in blockchain-empowered mobile edge computing[J]. IEEE Transactions on Vehicular Technology，2019，68(8)：8050-8062.
[23] ZHU Z，LIU T，YANG Y，et al. BLOT：Bandit learning-based offloading of tasks in fog-enabled networks[J]. IEEE Transactions on Parallel and Distributed Systems，2019，30(12)：2636-2649.
[24] BAEK J，KADDOUM G. FLoadNet：Load balancing in fog networks with cooperative multiagent using actor–critic method[J]. IEEE Transactions on Network and Service Management，2023，20(1)：400-414.
[25] BAEK J Y，KADDOUM G，GARG S，et al. Managing fog networks using reinforcement learning based load balancing algorithm[C]// Proceedings of the 2019 IEEE Wireless Communications and Networking Conference (WCNC)，15-18 April 2019，2019.
[26] YU S，CHEN X，ZHOU Z，et al. When deep reinforcement learning meets federated learning：Intelligent multitimescale resource management for multiaccess edge computing in 5G ultradense network[J]. IEEE Internet of Things Journal，2021，8(4)：2238-2251.
[27] DAI F，LIU G，MO Q，et al. Task offloading for vehicular edge computing with edge-cloud cooperation[J]. World Wide Web，2022，25(5)：1999-2017.
[28] LE D V，THAM C K. A deep reinforcement learning based offloading scheme in ad-hoc mobile clouds [C]// Proceedings of the IEEE INFOCOM 2018- IEEE Conference on Computer Communications Workshops (INFOCOM WKSHPS)，15-19 April 2018，2018.
[29] TONG Z，WANG J，MEI J，et al. Multi-type task offloading for wireless Internet of Things by federated deep reinforcement learning[J]. Future Generation Computer Systems，2023，145：536-549.
[30] BAEK J，KADDOUM G. Heterogeneous task offloading and resource allocations via deep recurrent reinforcement learning in partial observable multifog networks[J]. IEEE Internet of Things Journal，2021，8(2)：1041-56.
[31] SILVA D，RAFAEL J，FONTE A. Toward optimal virtualization：An updated comparative analysis of docker and LXD container technologies[J]. Computers，2024，13(4).
[32] GIRMA R. Evaluation of container virtualization systems supporting Open container initiative images[J]. Proceedings Girma 2018 Evaluation OC，2018，https：//api.semanticscholar.org/CorpusID：117230605.
[33] GUPTA A，PACCHIANO A，ZHAI Y，et al. Unpacking reward shaping：Understanding the benefits of reward engineering on sample complexity[J]. Advances in Neural Information Processing Systems，2022，35：15281-95.
[34] DOSOVITSKIY A，BEYER L，KOLESNIKOV A，et al. An image is worth 16x16 words：Transformers for image recognition at scale[J]. arXiv preprint arXiv：201011929，2020.
[35] CARION N，MASSA F，SYNNAEVE G，et al. End-to-end object detection with transformers[C]// Proceedings of the European Conference on Computer vision，2020. Springer.
[36] CONNEAU A，KHANDELWAL K，GOYAL N，et al. Unsupervised cross-lingual representation learning at scale[J]. arXiv preprint arXiv：191102116，2019.
[37] LEWIS M，LIU Y，GOYAL N，et al. Bart：Denoising sequence-to-sequence pre-training for natural language generation，translation，and comprehension[J]. arXiv preprint arXiv：191013461，2019.
[38] RADFORD A，WU J，CHILD R，et al. Language models are unsupervised multitask learners[J]. OpenAI blog，2019，1(8)：9.
[39] YANG A，LI A，YANG B，et al. Qwen3 technical report[J]. arXiv preprint arXiv：250509388，2025.
[40] FAN J，WANG Z，XIE Y，et al. A theoretical analysis of deep Q-learning[C]// Proceedings of the Learning for Dynamics and Control，2020. PMLR.
[41] LU J，YANG J，LI S，et al. A2C-DRL：Dynamic scheduling for stochastic edge–cloud environments using A2C and deep reinforcement learning[J]. IEEE Internet of Things Journal，2024，11(9)：16915-16927.
[42] YU C，VELU A，VINITSKY E，et al. The surprising effectiveness of ppo in cooperative multi-agent games[J]. Advances in Neural Information Processing Systems，2022，35：24611-24624.
[43] KöK İ，OKAY F Y，ÖZDEMIR S. FogAI：An AI-supported fog controller for next generation IoT[J]. Internet Things-Neth，2022，19：100572.