PREDICTION OF DEM PARAMETERS OF COATED FERTILIZER PARTICLES BASED ON GA-BP NEURAL NETWORK

ABSTRACT To provide an efficient and reliable calibration method with reduced time cost and increased accuracy, the angle of repose (AoR) in the simulation is batch-processed based on Python and the GA-BP neural network is used to improve the prediction accuracy of the DEM parameters of coated fertilizer particles. The single-factor test data were firstly interpolated to obtain sufficient training samples, thus avoiding the drawback that the BP network tends to fall into the local minimum during the training process. Then the GA-BP neural network was trained in combination with the orthogonal combination test, and the fitted correlation coefficients were all greater than 0.975, indicating that the algorithm has strong generalization performance and good stability. The predicted values matched the expected output values, indicating that the GA-BP neural network can accurately predict the nonlinear function output, and the network predicted output can be approximated as the actual output of the function. With the actual AoR as the output value, the simulation value of the AoR was obtained as 24.457° when the coefficient of restitution (CoR), coefficient of static friction (CoSF), and coefficient of rolling friction (CoRF) were 0.509, 0.176, and 0.0332, respectively, and the relative error with the actual value was 0.068%, indicating that the well-fitted GA-BP neural network could accurately predict the DEM parameters of fertilizer particles.摘要 为了提供一种高效可靠、可降低时间成本并提升校准精度的标定方法，本研究基于Python对离散元法模拟中的休止角（angle of repose，AoR）进行批量处理，并采用遗传算法-反向传播（GA-BP）神经网络，以提升包膜肥料颗粒离散元参数的预测精度。首先，对单因素试验数据开展插值处理以获取足量训练样本，规避了反向传播神经网络在训练过程中易陷入局部极小值的缺陷。随后结合正交组合试验完成GA-BP神经网络的训练，拟合得到的相关系数均大于0.975，表明该算法具备较强的泛化性能与良好的稳定性。预测值与预期输出值匹配度优异，证明GA-BP神经网络可精准预测非线性函数输出，且该网络的预测输出可近似等效为函数的实际输出。以实际休止角作为输出值，当恢复系数（coefficient of restitution，CoR）、静摩擦系数（coefficient of static friction，CoSF）以及滚动摩擦系数（coefficient of rolling friction，CoRF）分别为0.509、0.176与0.0332时，得到的休止角模拟值为24.457°，与实际值的相对误差仅为0.068%，这表明拟合效果优异的GA-BP神经网络可精准预测肥料颗粒的离散元参数。