基于MATLAB的BP神经网络的淀粉离散元接触参数标定  

Contact Parameter Calibration of Starch Discrete Element Method Based on BP Neural Network of MATLAB

作　　者：白玉菱 谢文影 赵孟涛 周康明 范仁宇 管天冰 孙会敏[2] 戴传云[1] BAI Yu-ling;XIE Wen-ying;ZHAO Meng-tao;ZHOU Kang-ming;FAN Ren-yu;GUAN Tian-bing;SUN Hui-min;DAI Chuan-yun(Chongqing Key Laboratory of Industrial Fermentation Microorganisms,College of Chemistry and Chemical Engineering,Chongqing University of Science and Technology,Chongqing 401331,China;National Institu-tes for Food and Drug Control,NMPA Key Laboratory for Quality Research and Evaluation for Pharmaceutical Excipients,Beijing 100050,China)

机构地区：[1]重庆科技学院化学化工学院,工业发酵微生物重庆市重点实验室,重庆401331 [2]中国食品药品检定研究院,国家药品监督管理局药用辅料质量研究与评价重点实验室,北京100050

出　　处：《中国药学杂志》2022年第15期1268-1277,共10页Chinese Pharmaceutical Journal

基　　金：国家科技重大新药创制专项资助(2017ZX09101-001-006);重庆市技术创新与应用发展专项面上项目资助(cstc2020jscx-msxmX0048)。

摘　　要：目的建立基于MATLAB的BP神经网络模型的淀粉离散元接触参数标定方法。方法建立单球型颗粒,并通过颗粒缩放分析将粒径放大至0.8 mm,以待标定的5种型号淀粉(高预胶化淀粉:S-1、低预胶化淀粉:S-2、直压型预胶化淀粉:S-3、水溶性淀粉:S-4、玉米淀粉:S-5)离散元接触参数(颗粒-颗粒恢复系数:A,颗粒-颗粒静摩擦系数:B,颗粒-颗粒滚动摩擦系数:C,颗粒-不锈钢恢复系数:D,颗粒-不锈钢静摩擦系数:E,颗粒-不锈钢滚动摩擦系数:F,表面能(J·m^(-2)):G)为输入层,两种测量方法(提升缸法和剪切盒法)休止角为输出层,MATLAB随机抽样50组进行离散元仿真模拟。BP神经网络算法对模拟结果进行训练,得到理想的神经网络模型后分别对以上淀粉的接触参数进行预测,并进行物理实验验证。结果当BP神经网络中隐含层中神经元个数为11个时,训练样本与测试样本的决定系数R^(2)分别为0.9999和0.9409,拟合较好,所建立神经网络的预测输出可以达到期望输出。获得的参数组合休止角模拟值与实测值相对误差均小于2.5%,表明预测准确。结论BP神经网络预测药用辅料淀粉的离散元接触参数可靠,可为后续固体制剂的仿真模拟提供可靠的数据支持。OBJECTIVE To build a calibration method for the discrete element method contact parameters of starch based on BP neural network model.METHODS Single spherical granules were created and the particle size was scaled up to 0.8 mm by particle scaling analysis.The discrete element contact parameters(particle-particle recovery coefficient A,particle-particle static friction coefficient B,particle-particle rolling friction coefficient C,particle-stainless steel recovery coefficient D,particle-stainless steel static friction coefficient E,particle-stainless steel rolling friction coefficient F,particle-particle surface energy(J·m^(-2))G)of the five starch(high pregelatinized starch:S-1,low pregelatinized starch:S-2,direct compression pregelatinized starch S-3,water soluble starch S-4,maize starch S-5)to be calibrated are used as the input layer,and the angle of repose obtained by the two measurement methods(hoist cylinder method and shear box method)is used as the output layer to build a BP neural network.The contact parameters were randomly sampled by MATLAB for 50 groups,and then the discrete element simulation was carried out according to the combination of parameters.The BP neural network algorithm was used to train the simulation results,and then the contact parameters of the starch were predicted,which were verified by physical experiments.RESULTS When the number of neurons in the hidden layer in the BP neural network is 11,the coefficient of determination R^(2)for the training and test samples was 0.9999 and 0.9409,respectively,which are a good fit,so the established neural network can be proved to be reliable.The fractional error between the simulated and measured values of the obtained parameter combinations of rest angle is less than 2.5%,indicating that the predictions are accurate.CONCLUSION The discrete element contact parameters of starch obtained using BP neural network predictions are reliable and can provide reliable data support for subsequent simulations of solid formulations.

关 键 词：淀粉 离散元 BP神经网络 颗粒缩放 参数标定 休止角 

分 类 号：R94[医药卫生—药剂学]