In most cases, designers must manually specify geometric tolerance types and values when designing mechanical products. For the same nominal geometry, different designers may specify different types and values of geometric tolerances. To reduce the uncertainty and realize the tolerance specification automatically, a tolerance specification method based on machine learning is proposed. The innovation of this paper is to find out the information that affects geometric tolerances selection and use machine learning methods to generate tolerance specifications. The realization of tolerance specifications is changed from rule-driven to data-driven. In this paper, feature engineering is performed on the data for the application scenarios of tolerance specifications, which improves the performance of the machine learning model. This approach firstly considers the past tolerance specification schemes as cases and sets up the cases to the tolerance specification database which contains information such as datum reference frame, positional relationship, spatial relationship, and product cost. Then perform feature engineering on the data and established machine learning algorithm to convert the tolerance specification problem into an optimization problem. Finally, a gear reducer as a case study is given to verify the method. The results are evaluated with three different machine learning evaluation indicators and made a comparison with the tolerance specification method in the industry. The final results show that the machine learning algorithm can automatically generate tolerance specifications, and after feature engineering, the accuracy of the tolerance specification results is improved.

在大多数情况下，设计人员在设计机械产品时必须手动指定几何公差类型和值。对于相同的标称几何形状，不同的设计人员可以指定不同的几何公差类型和值。为了减少不确定性并自动实现公差规格，提出了一种基于机器学习的公差规格方法。本文的创新之处在于找出影响几何公差选择的信息，并使用机器学习方法生成公差规格。公差规格的实现已从规则驱动变为数据驱动。本文针对公差规格的应用场景对数据进行了特征工程，从而提高了机器学习模型的性能。该方法首先将过去的公差规格方案视为案例，并将案例建立到公差规格数据库中，该数据库包含诸如基准参考系，位置关系，空间关系和产品成本之类的信息。然后对数据进行特征工程并建立机器学习算法，以将公差规格问题转换为优化问题。最后，以齿轮减速器为例进行了验证。使用三个不同的机器学习评估指标对结果进行评估，并与行业中的公差指定方法进行比较。最终结果表明，机器学习算法可以自动生成公差规格，并且经过特征工程后，