基于改进RBF神经网络的变形场插值算法研究

Research of Interpolation Algorithm for Deformation Field Based on Improved RBF Neural Network

摘要:

传统的物体表面力学变形场计算方法存在计算量大,无法计算边缘点变形等问题；提出一种改进的萤火虫算法优化RBF神经网络的变形插值方法,利用阈值约束RBF神经网络隐含层结点数,运用可变步长萤火虫算法优化RBF神经网络隐含层节点的中心和宽度,采用递推最小二乘法计算隐含层到输出层之间的权值,建立物体表面位移神经网络插值模型；为提高位移插值精度,在训练和测试的输入中增加坐标组合数据；应用于混凝土梁三点弯实验,仿真结果表明,该算法比常用的神经网络算法有更快的仿真速度和更高的预测精度,可用于土工材料表面变形场的快速、准确的计算。

Abstract:

The traditional method of calculating the transformation field on the object surface has difficulties in enormous computation, and the displacements near the boundary of an object cannot be calculated, etc. This paper proposes the method, which adopts the threshold to constrain the number of the nodes in the hidden layer, uses glowworm swarm optimization algorithm with the variable step to optimize the central vectors and extension width of nodes in the hidden layer, and employs the recursive least squares method to evaluate the weights between the hidden layer and the output layer, to build the interpolation model based on the neural network for the displacement field. In order to improve the precision of displacement interpolation, the coordinate dependencies as a part of input are added to the training and testing data. The proposed method is applied to the three-point bending tests of the concrete, and experimental results show that it has a faster convergence speed and higher precision than the normal neural networks method, is conducive to rapidly and accurately compute the displacement field.