Accurate prediction of rolling force can hardly be maintained by traditional mechanistic models under complex coupled conditions. Meanwhile, although nonlinear rules can be captured by purely data-driven machine learning models, they lack physical knowledge, which easily leads to overfitting and limits their generalization performance and interpretability. A physics-guided neural network (PGNN) model was proposed. By integrating the rolling mechanism into the loss function, the model was guided simultaneously by data features and physical laws. Meanwhile, a whale optimization algorithm (WOA) was introduced to optimize model hyperparameters to improve model performance. Experimental results indicate that the proposed PGNN method is superior to traditional data-driven methods in terms of prediction accuracy, data dependence, and generalization performance, and the prediction results are highly consistent with rolling theory. This method provides a new pathway for high-precision modeling of rolling force and demonstrates its application potential in the field of intelligent manufacturing.
然而,轧制过程机理复杂,对生产环境、轧制工况高度敏感,涉及变量众多,且各变量之间具有强耦合性与显著的非线性特征,这使得轧制力的精确预测极具挑战性[4].当前轧制力建模方法主要可分为基于机理的建模与基于数据的建模两类.基于机理的轧制力预测模型通常基于金属塑性变形的力学分析,其结果具有明确的物理意义[5-6].但由于轧制过程机理复杂、参数众多,分析中常需做出简化与假设,从而导致计算精度下降[7].此外,传统平板轧制理论普遍假定板材横向厚度一致或相近,而对于诸如MAS(mizushima automatic plan view pattern control system)轧制等特殊工艺,由于在特定道次中宽度方向存在显著厚度变化,采用常规平轧模型计算轧制力将产生较大误差[8].因此,在横向变厚轧制场景下,传统理论模型难以满足实际生产精度要求.另一类是基于有限元分析与数值仿真的方法.该类方法通过建立合理的边界条件与网格划分,可较准确地描述变形区应力应变分布.然而其计算耗时较长,难以满足实时在线控制的需求[9].
除常规超参数(如隐含层数与节点数)外,PGNN中物理约束项的权重系数同样需要确定.若通过人工试验逐一搜索最优参数组合,计算代价较高且效率低下.为此,本文采用鲸鱼优化算法(WOA)[29]对模型超参数进行自动优化,以获得最优结构.本文利用WOA同时优化网络结构参数(如隐含层数与隐含层节点数)及物理约束权重.其优化流程如图5所示,图中a为线性递减控制参数; A 为搜索系数向量;C为位置扰动系数;l为螺旋搜索系数;q为[0,1]的随机数.
在模型结构设计中,本文各隐含层均采用ReLU函数作为激活函数,输出层采用线性激活函数.学习率设定为0.01,优化算法选用Adam自适应优化器.为进一步提升模型性能,本文使用WOA对PGNN的超参数进行全局搜索.WOA种群规模设为50,最大迭代次数为100.优化目标为最小化验证集的总损失函数值.WOA优化过程中,模型在验证集上的均方根误差(root mean square error,RMSE)变化如图6所示,随着迭代的进行,模型RMSE显著下降并最终收敛,表明WOA有效提升了模型性能.经优化后得到的最优模型结构与参数如表2所示.
模型性能评估采用以下4个常用指标:平均绝对误差(mean absolute error,MAE)、均方根误差(RMSE)、平均绝对百分比误差(mean absolute percentage error,MAPE)以及决定系数(coefficient of determination,R2).其计算式如下:
WangGuo-dong. Creating a digital innovation infrastructure for iron and steel enterprises to accelerate the digital transformation of the iron and steel industry[J]. Steel Rolling, 2022, 39(6): 2-11.
PengWen, JiYa-feng, ChenXiao-rui, et al. Optimization of rolling force self-learning model in unsteady process of hot rolling[J]. Journal of Northeastern University(Natural Science), 2019, 40(10): 1408-1412.
[5]
CaoY, CaoJ G, WangT, et al. The cold rolling load distribution of the nuclear power zirconium alloy based on the self-adaptive particle swarm optimization algorithm[J]. International Journal of Advanced Manufacturing Technology, 2022, 119(9): 6007-6016.
[6]
YangY B, PengY. Dynamic rolling model based on uniform deformation[J]. Journal of Manufacturing Processes, 2020, 58: 1334-1347.
[7]
SimsR B. The calculation of roll force and torque in hot rolling mills[J]. Proceedings of the Institution of Mechanical Engineers, 1954, 168(1): 191-200.
[8]
LiuY M, MaG S, ZhaoD W, et al. Analysis of hot strip rolling using exponent velocity field and MY criterion[J]. International Journal of Mechanical Sciences, 2015, 98: 126-131.
[9]
DongZ S, LiX, LuanF, et al. Prediction and analysis of key parameters of head deformation of hot-rolled plates based on artificial neural networks[J]. Journal of Manufacturing Processes, 2022, 77: 282-300.
[10]
DingJ G, WangG Q, HeY H C, et al. Controllable points setting method for plan view pattern control in plate rolling process[J]. Steel Research International, 2020, 91(1): 1900345.
YanSong-tao, YangBin, JiangHai-tao. Finite element analysis of hot continuous rolling process of 6016 aluminum alloy based on secondary development of ABAQUS[J]. Journal of Plasticity Engineering, 2024, 31(6): 87-94.
[13]
JiY F, SongL B, YuanH, et al. Prediction of strip section shape for hot-rolled based on mechanism fusion data model[J]. Applied Soft Computing, 2023, 146: 110670.
[14]
DongZ S, LiX, LuanF, et al. Point and interval prediction of the effective length of hot-rolled plates based on IBES-XGBoost[J]. Measurement, 2023, 214: 112857.
WangZhen-hua, GongDian-yao, LiGuang-tao, et al. Bending force prediction model in hot strip rolling based on artificial neural network optimize by genetic algorithm[J]. Journal of Northeastern University(Natural Science), 2018, 39(12): 1717-1722.
[17]
JiY J, LiuS X, ZhouM C, et al. A machine learning and genetic algorithm-based method for predicting width deviation of hot-rolled strip in steel production systems[J]. Information Sciences, 2022, 589: 360-375.
[18]
WuY, NiH C, LiX, et al. Prediction of bending force in the hot strip rolling process using multilayer extreme learning machine[J]. Mathematical Problems in Engineering, 2021, 2021(1): 6682660.
[19]
WangQ N, SongL B, ZhaoJ W, et al. Application of the gradient boosting decision tree in the online prediction of rolling force in hot rolling[J]. The International Journal of Advanced Manufacturing Technology, 2023, 125(1): 387-397.
[20]
HuangX L, GuoX J, LiY L, et al. A novel data enhancement approach to DAG learning with small data samples[J]. Applied Intelligence, 2023, 53(22):27589-27607.
[21]
WangN Z, ZhangD X, ChangH B, et al. Deep learning of subsurface flow via theory-guided neural network[J]. Journal of Hydrology, 2020, 584: 124700.
[22]
LeiserF, RankS, Schmidt-KraepelinM, et al. Medical informed machine learning: a scoping review and future research directions[J]. Artificial Intelligence in Medicine, 2023, 145: 102676.
[23]
ZhangZ Y, ZhangH, LiuY, et al. Generalized conditional symmetry enhanced physics-informed neural network and application to the forward and inverse problems of nonlinear diffusion equations[J]. Chaos Solitons & Fractals, 2023, 168: 113169
[24]
ArzaniA, CasselK W, D'souzaR M. Theory-guided physics-informed neural networks for boundary layer problems with singular perturbation[J]. Journal of Computational Physics, 2023, 473: 111768.
[25]
MengL M, DingJ G, LiX J, et al. Novel shape control system of hot-rolled strip based on machine learning fused mechanism model[J]. Expert Systems with Applications, 2024, 255: 124789.
[26]
WenY, JiY F, WuB R, et al. A mechanism and data fusion step-wise prediction framework for hot-rolled strip thickness[J]. Journal of Manufacturing Processes, 2025, 134: 680-696.
[27]
ZhangS H, DengL, CheL Z, et al. An integrated model of rolling force for extra-thick plate by combining theoretical model and neural network model[J]. Journal of Manufacturing Processes, 2022, 75:100-109.
[28]
CaoL, LiX, LiX H, et al. Variable speed rolling force prediction with theoretical and data-driven models[J]. International Journal of Mechanical Sciences, 2024, 264: 108833.
[29]
RoutM, PalS K, SinghS B. Finite element analysis of cross rolling on AISI 304 stainless steel: prediction of stress and strain fields[J]. Journal of the Institution of Engineers (India): Series C, 2017, 98(1): 27-35.
[30]
ZhaoY, YangQ, HeA R, et al. Precision plate plan view pattern predictive model[J]. Journal of Iron and Steel Research, International, 2011, 18(11): 26-30.
JiaoZhi-jie, HeChun-yu, DingJing-guo, et al. Industrial popularization and application of plan view pattern control technology for plate mill[J]. Iron and Steel, 2019, 54(1): 49-55.
[33]
WangX L, ZhangW D, AiY B. A novel model based on quartic velocity field for hot rolling force of aluminum alloy thick plate[J]. The International Journal of Advanced Manufacturing Technology, 2023, 127(11): 5119-5125.
[34]
MirjaliliS, LewisA. The whale optimization algorithm[J]. Advances in Engineering Software, 2016, 95: 51-67.