Abstract:  Due to the influences of environmental instability factors such as temperature and air pressure during the injection molding processes， the processing parameters were changed during the molding processes， resulting in a decrease in product accuracy， product degradation or scrap. Aiming at the problems of similar environmental instability factors， using the data in the molding processes to predict the sizes of injection molding was helpful for the timely detection of unqualified products and reducing the occurrence of unqualified products. Based on the LightGBM framework， an injection molding product size prediction model was designed based on processing data and parameters. Through feature extraction， abnormal data processing， data set division， model training， model verification， and other steps， a product size prediction model with asymmetric risk characteristics was established. Because of the asymmetric risk of product size exceeding the specification， a weighted correction method was introduced to improve the prediction accuracy of the prediction model for the abnormal size based on the size range in the model training processes. Finally， the Foxconn injection molding size prediction data set was used to verify the prediction model， the results show that the model has higher prediction accuracy for out-of-specification dimensions. The average error of the verification set size prediction results is as 0.015 mm， and the weighted average error considering the asymmetry risk is as 5×10-6 mm. 

Key words:  , injection molding, asymmetric risk, machine learning, product size prediction, light gradient boosting machine （LightGBM）

摘要： 受温度、气压等环境不稳定因素的影响，注塑成形加工过程中工艺参数发生变化，从而导致产品精度下降，产品降级或报废。针对类似环境不稳定因素影响问题，利用加工过程中的数据进行注塑成形尺寸预测，有助于不合格产品的及时发现，减少不合格品的产生。基于轻量级梯度提升机（LightGBM）框架设计了基于加工过程数据及参数的注塑成形产品尺寸预测模型，通过特征提取、异常数据处理、数据集划分、模型训练、模型验证等步骤，建立了具有非对称风险特征的产品尺寸预测模型。针对产品尺寸超规的非对称风险问题，在模型训练过程中引入了基于尺寸范围的加权修正方法，以提高预测模型对超规尺寸的预测精度。最后利用富士康注塑成形尺寸预测数据集进行了验证，结果表明，该模型对超规尺寸具有更高的预测精度，尺寸预测结果平均误差为0.015 mm，考虑非对称风险的加权平均误差为5×10-6 mm。

关键词: 注塑成形, 非对称风险, 机器学习, 尺寸预测, 轻量级梯度提升机