基于自适应神经网络的火炮身管结构优化研究

针对火炮多学科优化设计存在计算量大、收敛慢和易陷入局部最优的问题,提出一种基于自适应径向基函数(RBF)神经网络的结构优化方法。编程计算火炮高低温压力曲线,并对ABAQUS有限元软件二次开发将其加载进有限元模型以获取身管的优化目标值,构建其与设计变量间自适应RBF神经网络模型。引入罚函数法处理约束条件,采用遗传算法在模型中求解寻优。每次优化迭代时利用建立的局部和全局分析模型分别选取更新点,增加样本点来更新神经网络,以提高神经网络的局部和全局预测能力。采用典型函数算例和某火炮身管结构多目标优化,实例验证了所提出优化策略的有效性。研究结果表明:身管优化后质量减小了6.63%,结构刚度提高了5.60%,最大等效应力减小了6.34%;与仅使用遗传算法相比,该方法所需的有限元模型调用次数降低了86.5%,运行时间减少了83.3%,为火炮结构设计和优化提供了参考。

Multi-objective Optimization of Gun Barrel Structure Based on Adaptive Neural Network

A structure optimization strategy using adaptive radial basis function artificial neural network (RBF ANN) is proposed for the large computational cost, slow convergence and easy to fall into local optimum in the multidisciplinary optimization design of gun. The high and low temperature-pressure curves of gun barrel are calculated based on the interior ballistic theory, as the load of numerical calculation of the finite element analysis model, by the secondary development of ABAQUS, which is used to obtain the optimization objectives. Then a RBF ANN is built to approximate the surrogate model for understanding the nonlinear relationships among the design variables and the optimization objectives. Penalty function method is used to solve the constraint problem, and the genetic algorithm is used to obtain current optimal solution. In the process of optimization, new sampling points are added, and the surrogate model is updated according to all the samples and their responses to improve the approximation accuracy around the local and global optimal solution. The multi-objective optimization strategy is validated by numerical test and the problem of optimization of the gun barrel structure performance to prove the efficiency of this optimization strategy. The results show that , compared to the initial design, the mass of optimized gun barrel is decreased by 6.63%, the structural stiffness is increased by 5.60%, and the maximum Von Mises stress is decreased by 6.34%. Furthermore, compared to GA without surrogate model, the number of function evaluation is decreased by 86.5%, and the total runtime is decreased by 83.3%.