Abstract:

In order to solve the problem of the lack of accurate and reliable discrete element simulation parameters for the design of agricultural implements in Yunnan forest Panax pseudoginseng cultivation， the red soil of Yunnan forest Panax pseudoginseng cultivation was taken as the research object， and the Hertz-Mindlin with JKR cohesion model in EDEM was selected to calibrate the relevant simulation parameters. On the basis of the previous experiments， the Plackett-Burman test and the steepest-climbing test were used to screen the significance factors and their optimal value intervals， the soil simulation stacking angle was used as the response value， and the response surface methodology （RSM） and machine learning were used to optimize and compare the significance parameters， respectively. The results showed that the soil JKR （Johnson-Kendall-Roberts） surface energy was 5.597 J·m^-2， soil-soil collision recovery coefficient was 0.314， soil-soil rolling friction factor was 0.132， and soil-65Mn steel collision recovery coefficient was 0.264 after using RSM optimized， the simulated stacking angle was 38.16°， and the relative error with the actual stacking angle was 2.03%. While using genetic algorithm （GA）-back propagation （BP）-GA model optimized， the soil JKR surface energy was 5.245 J·m^-2， soil-soil collision recovery coefficient was 0.404， soil-soil rolling friction factor was 0.171， soil-65Mn steel collision recovery coefficient was 0.318， simulated stacking angle was 36.81°， and the relative error with the actual stacking angle was 1.57%， which was superior to that of the RSM. Above results showed that the GA-BP-GA algorithm was superior to the RSM method in parameter optimization， and the red soil parameter calibration results obtained could be used in discrete element simulation.

Key words: red soil, discrete element, parameter calibration, response surface, BP neural network, genetic algorithm

摘要：

为解决云南林下三七种植中农机具设计缺乏准确、可靠的离散元仿真参数的问题，以云南林下三七种植红壤为研究对象，选取EDEM中的Hertz-Mindlin with JKR cohesion模型，对相关仿真参数进行标定。在前期试验的基础上，采用Plackett-Burman试验和最陡爬坡试验筛选显著性因素及其最优值区间；以土壤仿真堆积角为响应值，分别采用响应面法（response surface methodology，RSM）和机器学习对显著性参数进行优化和对比。结果显示，采用RSM优化得到土壤JKR（Johnson-Kendall-Roberts）表面能为5.597 J·m^-2、土壤-土壤碰撞恢复系数为0.314、土壤-土壤滚动摩擦因数为0.132、土壤-65Mn钢碰撞恢复系数为0.264，测得仿真堆积角为38.16°，与实际堆积角相对误差为2.03%；采用GA-BP（genetic algorithm-back propagation）-GA模型优化的土壤JKR表面能为5.245 J·m^-2、土壤-土壤碰撞恢复系数为0.404、土壤-土壤滚动摩擦因数为0.171、土壤-65Mn钢碰撞恢复系数为0.318，仿真堆积角为36.81°，与实际堆积角相对误差为1.57%。综上表明，GA-BP-GA算法在参数优化方面优于RSM方法，获得的红壤参数标定结果可以用于离散元仿真。

关键词: 红壤, 离散元, 参数标定, 响应面, BP神经网络, 遗传算法