Abstract:  Based on production data for 1180 MPa level ultra-high strength cold-rolled dual-phase（DP）steel coils, the chemical principal component extraction method based on principal component analysis, the hyper-parameter optimization method combining grid search and cross-validation were studied, and GBDT prediction models of DP steel mechanics properties were established. The predicted results were compared with those of BP neural network models and generalized additive models(GAM). To improve the prediction accuracy of elongation at break, based on the GBDT prediction model with high prediction accuracy, a prediction correction model of elongation at break considering error compensation was established through the model prediction error classification model and the model prediction correction method considering error compensation, and the prediction accuracy of the elongation correction model reaches 94.63% within an absolute error range of ±0.9%. The DP steel property prediction model performs well with good prediction accuracy during online operation, meeting production requirements, and is helpful for online quality monitoring of mechanics properties.

Key words: cold rolled high strength steel, gradient boosting decision tree（GBDT）, prediction of mechanics property, principal component analysis, error compensation

摘要： 基于1180 MPa级超高强度冷轧双相(DP)钢卷生产数据，研究了基于主成分分析的化学主成分提取方法、网格搜索和交叉验证相结合的超参数寻优方法，建立了DP钢力学性能梯度提升决策树(GBDT)预测模型，并将预测结果与BP神经网络模型和广义可加模型的预测结果进行了比较。为了提高断后伸长率预测精度，以预测精度较高的GBDT预测模型为基础，通过模型预测误差分类模型和考虑误差补偿的模型预测修正方法，建立了考虑误差补偿的断后伸长率预测校正模型，该模型使断后伸长率在绝对误差±0.9%下的预测准确率达到了94.63%。DP钢性能预测模型在线运行时的实际预测精度较高，达到生产要求，有助于力学性能在线质量监控。

关键词: 冷轧高强钢, 梯度提升决策树, 力学性能预测, 主成分分析, 误差补偿