机械装配技术的智能化发展是人工智能时代下的必然趋势。作为综合国力的象征之一，深刻把握人工智能技术对机械装配技术的内在驱动性，是加速发展智能装配技术的重要前提。对于重型车辆（如装甲车）运动总成装配生产线而言，由于螺栓连接众多、装备件体积庞大，拧紧工位通常是该类装备件装配质量和装配效率的瓶颈工位。影响拧紧质量和效率的原因体现在多个方面：在拧紧前，螺栓初始位置姿态的倾斜或偏心等问题会给拧紧过程带来不确定性，进而引发质量问题；在拧紧中，同一装备件上多螺栓依次拧紧产生的交互影响使得最终螺栓预紧力松弛，进而引起质量问题；在拧紧后，拧紧质量通过人工检验或传统自动检验方法均无法准确判断质量原因而造成产品使用安全和可靠性问题。本文以大型轮毂类、多螺栓装备件拧紧系统为研究对象，对拧紧工艺中出现的上述三个关键技术问题展开研究，分析传统拧紧方法在解决上述关键问题的局限性，结合人工智能技术提出新模型，并搭建工业级的智能拧紧装配系统将新模型封装集成。围绕课题研究的目标分别对系统核心技术开展深入研究，具体包括：（1）针对拧紧过程中由于多螺栓交互作用造成螺栓预紧力松弛的问题，论文对影响预紧力松弛的因素开展研究，提出预紧力松弛预测的参数模型，将复杂多变的环境因素用待拟合参数来表达，提高模型适应不同拧紧场合的柔性，为拧紧策略优化提供依据；（2）针对传统拧紧装配质检方法主观性强、问题反馈能力不足等问题，论文设计智能诊断模型，通过输入拧紧过程曲线可以实现对曲线按故障模式进行分类。具体地，提出基于对抗迁移学习的注意力门控模型对拧紧曲线进行故障诊断，并通过对比实验证明该模型在诊断小样本、长时序特征的拧紧曲线的优势；（3）针对螺栓错位或偏心影响装配质量的问题，论文提出自主决策方案，依托Unity仿真平台建立拧紧系统数字世界，并提出深度马尔可夫决策模型启发拧紧轴在数字世界自主调整螺栓位姿，紧接着通过建立感知系统捕捉真实世界螺栓位姿坐标，该坐标信息传入训练成熟的决策模型即可生成真实的螺栓位姿调整方案。依托项目设计开发了智能拧紧装配系统，验证了本文所提理论研究成果的鲁棒性，在提高实际工厂拧紧装配工艺质量和效率的同时，也提升了其智能化水平。

The intelligent development of the mechanical assembly technology is an inevitable trend in the era of artificial intelligence (AI). As a symbol of comprehensive national strength, profoundly grasping the inherent driving force of the AI technology for assembly process is an important prerequisite for accelerating the development of the intelligent assembly. For the production line of the large-scale armored vehicle assembly, due to the large number of bolt connections and the large size of the spare parts, the tightening station is usually the bottleneck station for the assembly quality and efficiency. The reasons that affect the quality and efficiency of tightening process are reflected in many aspects: Before tightening, the inclination and eccentricity of the initial position and posture of the bolt will bring uncertainty to the tightening process and cause quality problems; During tightening, the interaction effects of the sequential tightening of multiple bolts on the same equipment part make the final bolt tightening force loose and cause quality problems; After tightening, tightening quality can not be accurately diagnosed by manual inspection or traditional automatic inspection methods, resulting in product safety and reliability issues. This paper takes the tightening system of large wheel hubs and multi-bolt equipment parts as the research object, studies the above three key technical problems in the tightening process, analyzes the limitations of traditional tightening methods in solving the above key problems, proposes new methods based on artificial intelligence technology, and build an industrial-level intelligent tightening assembly system to integrate the new models. In-depth researches on the core technology of the system around the objectives of the subject are carried out, including: (1) Aiming at the problem of bolt pretightening force relaxation caused by the interaction of multiple bolts during the tightening process, the paper studies the factors affecting pretightening force relaxation, and proposes a parametric model to predict the pretightening force relaxation. In detail, complex and variable environmental factors are expressed by the parameters, which could improve the flexibility of the model to adapt to different tightening situations, and provide a basis for tightening strategy optimization;(2) Aiming at the problems of high subjectivity and insufficient problem feedback ability of traditional tightening assembly quality inspection methods, this paper designs an intelligent diagnosis model, which can classify the tightening curves according to the failure mode. Specifically, an attention-gated model based on adversarial transfer learning is proposed for fault diagnosis of tightening curves, and comparative experiments are conducted to prove that the model has the advantages of learning small samples and long-temporal tightening curves and has high diagnostic accuracy;(3) Aiming at the problem that bolt misalignment or eccentricity affects assembly quality, the paper proposes an autonomous decision-making scheme, a digital world of the tightening system is established based on Unity simulation platform and a deep Markov decision model is proposed to inspire the tightening axis to adjust the bolt position and posture independently in the digital world. Then, a perception system is established to capture the bolt pose coordinates in the real world, and the coordinate information can be passed into a well-trained decision-making model to generate the bolt pose optimization strategy.Relying on the project, an intelligent tightening assembly system is developed, which verifies the robustness of the theoretical research results presented in this paper, and effectively improves the quality and efficiency of the tightening process in the real factory, while improving its intelligence level.