针对现有机器人接触性操作任务中作用力建模方法表征能力不足、通用性差的问题，以玻璃基板卸片这一实际生产过程为例，对复杂接触动力学的建模方法进行研究。考虑到玻璃基板铲起过程中的作用力受多个界面接触动力学的影响，表现出多模态、非线性、非平稳性的特性，将物理先验知识以不同形式融合在深度学习模型的设计以及训练过程中，提出一种结合深度学习与机理模型的铲起过程作用力预测方法——针对玻璃基板铲起过程的受力特点，提出结合多尺度卷积核、注意力机制以及长短时记忆网络的深度学习模型结构；提出动力学参数随机化方法与基于材料力学、断裂力学的接触力补偿措施，使仿真训练数据更鲁棒地反映真实接触情况；在均方误差损失函数基础上，针对不合理的物理“穿透”行为引入附加损失函数进行网络训练。实验结果表明，所提方法在作用力单步预测中的均方根误差为0.286，可以准确地预测水平与竖直两个方向的作用力，多步预测结果也可以满足应用需求，预测性能优于现有的主流模型。消融实验表明，文中方法的优良预测性能是局部特征提取模块、注意力机制模块与时序特征提取模块3个组件共同作用的结果，同时，所提出的改进损失函数提高了模型训练的稳定性。文中方法可以应用于类似场景中对机器人与环境的接触力预测。

In view of the problem of insufficient characterization ability and poor versatility of action force modeling methods in existing robot contact operation tasks, this paper studied on the modeling methods of complex contact dynamics by taking the actual production process of glass substrate shoveling as an example. Considering the fact that the forces during the glass substrate shoveling are affected by the contact dynamics of multiple interfaces, exhibiting the multimodal, nonlinear, and non-stationarity properties, this paper proposed a method for the prediction of the contact force by integrating physical prior knowledge in different forms in the design and training process of deep learning models. According to the stress characteristics of the glass substrate’s shoveling up process, a deep learning model structure combining multi-scale convolutional kernel, attention mechanism and long and short-term memory network was proposed; the kinetic parameter randomization method and the contact force compensation measures based on material mechanics and fracture mechanics were proposed to make the simulation training data more robust to reflect the real contact situation; based on the mean square error loss function, the additional loss function was introduced for network training for the unreasonable physical “penetration” behavior. The experimental results show that the proposed model can accurately predict the horizontal and vertical forces with the root mean square error of 0.286 in the single-step prediction, and the multi-step prediction results are also good enough to meet the application requirements. The prediction performance of the model is superior to the existing mainstream models. The ablation experiment shows that the excellent prediction performance of the model proposed is the result of the joint contribution of the local feature extraction module, the attention mechanism module and the temporal feature extraction module. At the same time, the improved loss function improves the stability of the model training. Our method can be used in similar applications for the prediction of robotic contact force with the environment.