采用BP神经网络和Burgers模型的细观参数标定

Calibration method of mesoscopic parameters using BP neural network and Burgers model

PFC软件作为一款成熟的离散元分析软件,由于在处理连续与非连续介质方面的出色表现,得到了广泛的应用。但PFC软件所需要的细观参数均需要采用室内试验数据通过试错法反复调试才能获得,效率低、盲目性高,严重影响后续试验数据,因此需要细观参数校准方法标定PFC。该研究以玉米秸秆颗粒的单轴蠕变试验为基础,结合离散元软件PFC 2D,通过正交试验多因素方差分析方法分析了Burgers模型宏细观参数之间的影响关系,从而证明宏细观参数之间存在着复杂关系,不宜采用通过回归分析获得宏细观参数之间的关系式的方式标定细观参数,适合利用BP神经网络进行参数标定,利用创建的BP神经网络对细观参数进行标定,根据测试组的标定结果分析得出Burgers模型各细观参数的标定精度均在92%以上,且误差较为稳定,而且训练好的神经网络相关系数R>0.96,从而证明BP神经网络的细观参数标定性能较为可靠。将玉米秸秆单轴蠕变试验的宏观参数带入训练好的BP神经网络中进行细观参数标定,比对模拟蠕变试验与物理蠕变试验发现,两者的蠕变曲线基本一致,应变量的最大误差为2%,证明了BP神经网络具有良好的参数标定能力,方法可为PFC参数标定提供一定的参考价值。

Particle flow code(PFC)software has been widely used as the general discrete-element modeling(DEM),due to the excellent performance to deal with continuous and discontinuous media.Among them,the mesoscopic parameters can only be acquired to repeatedly debug the experimental data using trial-and-error method,leading to the low efficiency with the high blindness.A set of usable parameters can be inevitable in the dozens of trial and error during calibration,even though the sound experience of experts.Therefore,it is highly urgent to accurately and rapidly calibrate the mesoscopic parameters for the promotion of PFC software and the follow-up test,particularly beyond the manual operation.In this study,the uniaxial creep test model of corn stalk particles was established to combine with the built-in Burgers model of the PFC 2D.An orthogonal experiment was also carried out to verify the improved model.The multivariate analysis of variance was then made to analyze the complex relationship between the macroscopic and mesoscopic parameters of the Burgers model.There was a quite difference in the significance of the influence of each mesoscopic parameter on the macroscopic one.A highly nonlinear relationship was also found between the macroscopic and mesoscopic parameters.Therefore,the regression analysis was inappropriate to obtain the relationship between the macroscopic and mesoscopic parameters for the calibration of the mesoscopic parameters.Fortunately,BP neural network can be expected to serve as these complex relationships,just suitable for the parameter calibration.As such,the BP neural network was established with the 4,9 and 5 nodes in the input,hidden,and output layer,respectively,according to the number and characteristics of macroscopic and mesoscopic parameters.Then,the resulting BP neural network was trained and calibrated using 150 sets of macroscopic and mesoscopic parameters.It was found that above 92%was achieved in the calibration accuracy of all mesoscopic parameters in the Burgers model,especially with the relatively stable errors.Moreover,the correlation coefficient(R)was greater than 0.96 in the trained BP neural network,indicating the more reliable performance of inversion.The improved calibration of parameters can also be popularized for the mesoscopic parameters.Furthermore,the macroscopic parameters after the uniaxial creep test of corn stalk were introduced into the trained BP neural network for the calibration of the mesoscopic parameters.A better consistence was found in the simulated and measured creep curves with the maximum error of the dependent variable of 2%,indicating the excellent calibration ability of parameters.The finding can also provide a strong reference for the PFC parameter calibration.