Customized mechanical products based on orders have characteristics such as small batch sizes and multi-variety in production, which results in less availability of sample data related to assembly. Therefore, a problem of prediction of assembly precision exists due to the small number of samples. This paper studies the prediction method of mechanical product assembly precision based on the fusion of measured samples and feature fidelity samples. First, an assembly-feature fidelity sample (AFFS) generation method based on measured data is proposed, which expands the amount of mechanical product assembly feature samples through the meta-model concept. Then, a fusion method of measured samples and AFFS of mechanical products based on a double-layer learning network is proposed, which improves the under-fitting of a few samples and enhances the generalization ability of the model. Finally, case studies of gear-shaft assembly structure that are common in mechanical products, such as engines, gearboxes, and traction machines, were examined to verify the method we proposed and compare them with a tolerance analysis method and a machine learning method. The average errors of the assembly precision predicted by the method in this paper are 1.2% and 5.0%, respectively, which are better than the outcomes of SVR and NNs. The average errors of SVR and NNs are 3.2% and 2.8% in case 1 and 19% and 17% in case 2, respectively. The results show that the method in this paper has advantages of smaller error fluctuations and the best accuracy stability.

基于订单的定制机械产品具有诸如小批量和多品种生产的特征，这导致与组装相关的样本数据的可用性降低。因此，由于样本数量少，存在预测组装精度的问题。本文研究了基于实测样本与特征保真度样本融合的机械产品装配精度预测方法。首先，提出了一种基于实测数据的装配特征保真度样本生成方法，通过元模型概念扩展了机械产品装配特征样本的数量。然后，提出了一种基于双层学习网络的机械产品实测样本与AFFS的融合方法，这改善了一些样本的拟合不足，并增强了模型的泛化能力。最后，研究了机械产品（例如发动机，变速箱和牵引机）中常见的齿轮轴总成结构的案例研究，以验证我们提出的方法，并将其与公差分析方法和机器学习方法进行比较。该方法预测的装配精度平均误差分别为1.2％和5.0％，优于SVR和NN的结果。SVR和NN的平均误差在案例1中分别为3.2％和2.8％，在案例2中分别为19％和17％。结果表明，该方法具有误差波动较小，精度稳定性最好的优点。研究了机械产品（例如发动机，变速箱和牵引机）中常见的齿轮轴总成结构的案例研究，以验证我们提出的方法，并将其与公差分析方法和机器学习方法进行比较。该方法预测的装配精度平均误差分别为1.2％和5.0％，优于SVR和NN的结果。SVR和NN的平均误差在案例1中分别为3.2％和2.8％，在案例2中分别为19％和17％。结果表明，该方法具有误差波动较小，精度稳定性最好的优点。研究了机械产品（例如发动机，变速箱和牵引机）中常见的齿轮轴总成结构的案例研究，以验证我们提出的方法，并将其与公差分析方法和机器学习方法进行比较。该方法预测的装配精度平均误差分别为1.2％和5.0％，优于SVR和NN的结果。SVR和NN的平均误差在案例1中分别为3.2％和2.8％，在案例2中分别为19％和17％。结果表明，该方法具有误差波动较小，精度稳定性最好的优点。被检查以验证我们提出的方法，并将其与公差分析方法和机器学习方法进行比较。该方法预测的装配精度平均误差分别为1.2％和5.0％，优于SVR和NN的结果。SVR和NN的平均误差在案例1中分别为3.2％和2.8％，在案例2中分别为19％和17％。结果表明，该方法具有误差波动较小，精度稳定性最好的优点。被检查以验证我们提出的方法，并将其与公差分析方法和机器学习方法进行比较。该方法预测的装配精度平均误差分别为1.2％和5.0％，优于SVR和NN的结果。SVR和NN的平均误差在案例1中分别为3.2％和2.8％，在案例2中分别为19％和17％。结果表明，该方法具有误差波动较小，精度稳定性最好的优点。