基于可解释极端随机树模型的 DCT 液压响应预测

为解决传统湿式双离合器变速器 （Dual Clutch Transmission, DCT） 控制策略在硬件误差以及复杂工况下液压响应预测精度不完全可控的问题,提出了一种基于 SHAP 图可解释极端随机树预测模型,使用机器学习方法结合某汽车公司 DCT 实验室采集的真实离合器数据对 DCT 液压响应进行预测。模型利用 SHAP 算法对于重要特征选择的可解释性,筛选并保留对液压响应影响较大的特征,将时间切片和升降压判定作为特征加入训练数据,训练预测模型。结果表明,该模型训练结果的均方误差 MSE 为 0.670 3,可决系数 R2为 1.000 0,并且在测试集上预测值与实际值之间的平均误差为12.99 kPa,远低于设计误差 25 kPa,具有较高的预测精度,特征选择较准确,可以很好地解决传统物理模型无法计算不同工况下液压响应的问题,为下阶段基于数据和物理双驱动的DCT控制策略优化提供较准确的预测结果。

Interpretable Extremely Randomized Trees Model for Predicting DCT Hydraulic Response

Traditional wet dual-clutch transmissions(DCT) control strategies face challenges in accurately predicting hydraulic responses, especially under hardware errors and complex working conditions. Therefore this paper proposes an Interpretable Extremely Randomized Trees prediction model based on the SHAP graph. By employing machine learning techniques and utilizing the actual clutch data collected from the DCT laboratory of a certain automobile company, it predicts the hydraulic response of DCT. This model uses the interpretability of the SHAP algorithm to select essential features that greatly impact hydraulic response, and adds time slices and buck-boost determinations as features into the training data to train the prediction model. The results show that the mean square error(MSE) of the model''s training results is 0.670 3, with a coefficient of determination(R2) of 1.000 0.Furthermore, the average error between the predicted and actual values on the test set is 0.129 9 bar, which is much lower than the design error of 0.25 bar. This proves high prediction accuracy and precise feature selection of the model, which can effectively address the limitations of traditional physical models in calculating hydraulic responses under different working conditions. The proposed model provides more accurate prediction results for the next stage of DCT control strategy optimization based on a dual data and physical drive approach.