.png){kind=icon}
.png){kind=icon}
Return
Prediction of rolling force based on finite element data enhancement and physical information neural network
.png){kind=icon}
PRE
AI
Z
J
Z
DOI：10.1007/s42243-026-01858-5.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Aiming at the problems of insufficient accuracy of traditional theoretical models and weak physical interpretability of artificial neural networks, a rolling force prediction method combining finite element data enhancement and physical information neural network (PINN) was proposed. A finite element model of a six-high cold rolling mill was constructed, and the scaled expansion of sample data was achieved by simulating multi-parameter coupling working conditions (the error between simulation values and real values was not above 5%). Two types of PINN frameworks were constructed, which incorporated process-sensitive physical features (such as the partial differential relationships of reduction and tension difference) and the Hill formula, respectively, realizing the in-depth integration of data-driven approaches and physical mechanisms. The Bayesian optimization algorithm was used for global optimization of model hyperparameters (such as the number of hidden layers, learning rate, and weight of physical loss), ensuring that each model achieved the optimal configuration. Combined with Spearman correlation coefficient and SHapley Additive exPlanations value analysis, quantitative research was conducted on the monotonicity and influence degree of each feature. The research showed that the determination coefficient of the process-sensitive physical feature model was 0.9711, and the root mean square error was 0.0159. Additionally, this model exhibited the fewest data points exceeding the 5% error threshold in the test set, showcasing the best prediction accuracy.
Keywords:
Physical information neural network
Rolling force
Finite element
Data-driven
Physical constraint
Journal
IF:
3.6
Papers:
3.6K
Citations:
6.1K
Organization
Cited Papers