基于PCA-BP神经网络模型的采场底板破坏深度预测方法

Failure depth prediction method of stope floor based on PCA-BP neural network model

为了准确预测煤矿开采过程中底板岩层破坏深度,在综合考虑影响底板岩层变形破坏因素和数学描述可能性的基础上,选取煤层采深、煤层倾角、煤层采厚、工作面斜长、岩层抗破坏能力,以及是否有切穿型断层或破碎带这6个最主要影响因子作为分析预测破坏深度的判别指标。根据我国各煤矿采煤工作面实测资料,选取其中具有代表意义的实测数据组成6×32的数据集。首先利用主成分分析(principal component analysis,PCA)法降低数据集间的多重相关性,并将降维后的主成分重新命名为Z 1、Z 2、Z 3;其次将新的数据集划分为训练集与测试集,并代入基于MATLAB平台构建的BP神经网络模型进行训练测试,通过计算训练集、测试集预测值与实际值的误差和决定系数,评判所构建网络的可靠程度;最后将孙村煤矿十一煤层11121工作面实测数据分别代入规程计算公式,结合已经训练好的PCA-BP神经网络模型进行预测,发现PCA-BP网络模型的预测值比规程计算公式的更接近实测值。

In order to solve the problem of floor damage depth in the process of coal mining,based on the comprehensive consideration of the factors that affecting the deformation and failure of floor rocks of coal seam and the possibility of a mathematical description,six main factors affecting the mining depth of coal seam,mining depth,coal seam dip,mining height,facing length,floor damage resistance ability and the type of faults and crushed zone passed were selected as the discriminant indexes for the analysis and prediction of the failure depth.According to the measured data of each coal mining face in China,the measured data with representative significance are selected to form a 6×32 data set.Firstly,principal component analysis(PCA)was used to reduce multiple correlations among data sets,and the reduced principal components were renamed as Z 1,Z 2 and Z 3.Secondly,the new data set was divided into training set and test set,and the BP neural network model based on MATLAB platform was substituted for the training and testing,and the reliability of the constructed network was evaluated by calculating the error and determination coefficient of the actual value and the predicted value of the training set sample and the predicted value of the test set sample respectively.Finally,the measured data of working face 11121 of No.11 coal seam in Suncun coal mine was respectively substituted into the regulation calculation formula and the trained PCA-BP neural network model simulation.It is proved that the predicted value of the composite model is closer to the measured value than that of the formula calculated by the regulation.