超参数优化神经网络驱动的弹炮碰撞动力学建模OA

To address the modeling challenges of contact/impact phenomena caused by the complex contacting surface between barrel and bourrelet,a modeling method for complex surface contact/impact process between barrel and projectile was developed based on artificial neural networks,accompanied by optimization strategies for the enhancement of model performance.Due to the difficulty of conducting practical tests during extreme launching environment,a barrel-projectile contact/impact model was constructed using finite element software.Simulation was employed to obtain the sample dataset required for neural network training under various initial normal contacting velocity conditions.A hybrid optimization framework for hyperparameter tuning was developed,wherein the activation function and the distribution range of hidden layer nodes were determined through trial-and-error screening.The number of hidden layer nodes and the initial weight-bias configuration were fine-tuned using the genetic algorithm.To comprehensively assess the model performance,a dual evaluation mechanism combining global statistical analysis and local detailed assessment was introduced.The overall predictive fluctuation characteristics of the data-driven model were quantified using box plots,while the Taylor diagram was employed to analyze local predictive accuracy of the representative median model,systematically validating the stability and generalization capability of model.Simulation results confirm the critical role of well-optimized hyperparameters in enhancing the extrapolation performance of model.The proposed approach not only establishes an accurate predictive model for barrel-projectile contact/impact responses but also provides a transferable methodological framework for constitutive modeling in other complex systems.