欢迎访问林业科学,今天是

林业科学 ›› 2020, Vol. 56 ›› Issue (6): 76-82.doi: 10.11707/j.1001-7488.20200608

• 论文与研究报告 • 上一篇    下一篇

基于人工神经网络模型的木材干燥应变模拟预测

付宗营1,蔡英春2,高鑫1,周凡1,江京辉1,周永东1,*   

  1. 1. 中国林业科学研究院木材工业研究所 国家林业和草原局木材科学与技术重点实验室 北京 100091
    2. 东北林业大学材料科学与工程学院 哈尔滨 150040
  • 收稿日期:2018-03-20 出版日期:2020-06-25 发布日期:2020-07-17
  • 通讯作者: 周永东
  • 基金资助:
    国家自然科学基金青年科学基金项目(31800478)

Simulation of Drying Strain Based on Artificial Neural Network Model

Zongying Fu1,Yingchun Cai2,Xin Gao1,Fan Zhou1,Jinghui Jiang1,Yongdong Zhou1,*   

  1. 1. Key Laboratory of Wood Science and Technology of National Forestry and Grassland Administration Research Institute of Wood Industry, CAF Beijing 100091
    2. College of Material Science and Engineering, Northeast Forestry University Harbin 150040
  • Received:2018-03-20 Online:2020-06-25 Published:2020-07-17
  • Contact: Yongdong Zhou

摘要:

目的: 研究常规干燥过程中干燥基准、预处理条件、含水率对木材干燥应力的影响,探讨干燥应力沿髓心至树皮方向的分布情况,以实现干燥应变的模拟预测。方法: 整合分析采用图像解析法测算得到的弹性应变和机械吸附蠕变相关数据,基于人工神经网络模型,以干燥温度、含水率、相对湿度、距髓心距离为输入变量对弹性应变进行模拟预测,以预处理温度、干燥温度、含水率、相对湿度、距髓心距离为输入变量对机械吸附蠕变进行模拟预测。通过网络训练和验证,得到合理的人工神经网络预测模型,并对模型进行测试,探讨分析所建立模型的预测能力。结果: 弹性应变预测模型中,各数据集均呈现出较好的相关性,训练集、验证集和测试集的相关系数(R)分别为0.988、0.983和0.978,所有数据集的决定系数(R2)均高于0.95,验证集达到最优时的均方差(MSE)为1.21×10-6。机械吸附蠕变预测模型中,利用含水率为28%和12%的数据集进行模型训练和验证,训练集和验证集的相关系数(R)分别为0.981、0.977,验证集达到最优时的均方差(MSE)为1.26×10-6;利用含水率20%的数据集进行模型测试,测试集的相关系数(R)为0.969,所有数据集的决定系数(R2)均高于0.94,网络模型能够解释94%以上的试验数据,表现出较好的预测能力。结论: 所建立模型的预测值和试验值吻合较好,预测成功率较高,能够为人工神经网络在干燥应力、应变方面的应用提供可行性依据。

关键词: 人工神经网络, 弹性应变, 机械吸附蠕变, 模拟预测

Abstract:

Objective: The effects of drying schedule, pretreatment condition, moisture content on wood drying stress and its distribution from pith to bark were investigated in this study to achieve the simulation and prediction of drying strain. Methods: The data sets of elastic strain and mechano-sorptive creep obtained with image analysis method were analyzed. The elastic strain and mechano-sorptive creep were modelled by artificial neural network, for the model of elastic strain with four inputs, i.e., drying temperatures, wood moisture content, relative humidity and the distance from pith, and for mechano-sorptive creep added pre-steaming temperature as an additional input. According to the training and validation processes, two reasonable prediction models were obtained, and then the predictive ability of the models were discussed with testing processes. Results: In elastic strain model, there were significant correlations between experimental and predicted values in all date sets. The R-values for training, validation and test sets were 0.988, 0.983 and 0.978, respectively. The R2 values were greater than 0.95 in all data sets, and the best validation performance of the mean square error was 1.21×10-6. In mechano-sorptive creep model, the R-values for training and validation sets were 0.981 and 0.977 with the data sets at moisture content of 28% and 12%, respectively. The best validation performance of the mean square error was 1.26×10-6. The R-value for test set was 0.969 with the data set at moisture content of 20%. Furthermore, the R2 values were greater than 0.94 in all data sets, indicating that the network model was capable to explain more than 94% experimental values. Conclusion: In the two established models, the predicted values were in good agreement with the experimental values, showing a high prediction accuracy, which provided a feasible basis for the application of artificial neural network in exploring drying stress and strain.

Key words: artificial neural network, elastic strain, mechano-sorptive creep, simulation and prediction

中图分类号: