基于机器学习的整体穿刺加压参数预测方法

针对整体穿刺加压密实过程中碳布回弹导致平均层高波动范围较大,影响立体织物性能的问题,提出一种基于机器学习理论的加压参数预测方法,将平均层高与加压参数之间复杂建模转换为多元回归问题,使用适合计算机运算的无约束优化迭代方法求解。基于scikit-learn类库对特征变量进行特征选择,对比6种回归模型的预测性能得分后选择K近邻回归作为基学习器,使用集成算法提升模型的预测性能。预测模型部署到生产环境后的实验结果表明:使用机器学习预测后,加压参数对整体穿刺过程平均层高均值变化的响应速度得到提高,且均值变化幅度得到降低,实验样本平均层高波动范围均值从12.0%降低至6.8%,标准差从0.008 3降低至0.006 6。

Prediction method of integrated piercing pressure parameters based on machine learning

In view of the problem that the spring back of carbon cloth leads to a large fluctuation range of average layer height and thus affects the performance of three-dimensional fabric during pressurized compaction process of integrated piercing, a real-time prediction method of pressure parameters based on machine learning theory was proposed. The complex modeling of the relationship between the average layer height and the pressurized parameters was transformed into multiple regression problems, and an unconstrained optimization iteration method suitable for computer operation was adopted to solve the problem. Based on the scikit-learn class library, the feature variables were selected. After comparing the predictive performance scores of the six regression models, the K nearest neighbor regression were selected as the base learners, and the prediction performance of the model was improved by using the integration algorithm. The experimental results after the prediction model was deployed to the production environment show that owing to the use of the machine learning prediction, the response speed of the pressure parameters to the average level in the integrated piercing process is improved, and the mean change amplitude is reduced. The average height fluctuation range of the experimental sample product is reduced from 12.0% to 6.8%, and the standard deviation from 0.008 3 to 0.006 6.