正交试验协同BP神经网络模型预测充填体强度

Strength Forecasting of Backfilling Materials by BP Neural Network Model Collaborated with Orthogonal Experiment

为准确预测充填体强度,基于18组混合水平的正交试验样本,以水泥熟料、脱硫灰渣、芒硝和钢渣的掺量作为4个输入因子,以充填体的7d和28d抗压强度作为输出因子,建立4×Y×2的BP神经网络充填体强度预测模型,并通过训练误差和预测强度误差的对比获得当隐含层神经元的个数Y取值为9时,模型的预测强度误差最小,其平均误差为0.7%。研究表明,该预测模型拟合的相关系数R高达0.999 89,7d和28d预测强度的最大相对误差分别为4.33%和0.84%,通过正交试验协同BP神经网络模型预测充填体强度可行性较强、准确度较高。该方法具有输入数据均匀分散、齐整可比和非线性优化的优点,为充填体强度的准确预测提供了新思路。

In order to forecast the strength of backfilling materials accurately,based on the 18 mixed samples of orthogonal test,the 4×Y×2 BP neural network model was established,in which cement clinker,desulphurized ash,Glauber’s salt and steel slag were utilized as four input factors and the compressive strength of 7 d and 28 d were used as output factors.Then,the training error curve and the predicted intensity error were analyzed.When the number of neurons in the hidden layer was 9,the prediction strength error of the model is minimized and the average error was 0.7%.The results revealed that the fitting correlation coefficient was 0.999 89 and the maximum relative error of 7 d and 28 d prediction strength were 4.33%and 0.84%,respectively.The method of BP neural network model collaborated with orthogonal experiment to forecast the compressive strength of backfilling materials are more feasibility and accurate.This method has advantages of high uniformity and non-linear optimization of input data,which provides a new idea for accurate prediction of backfilling materials strength.