基于深度置信网络的煤层含气量测井解释研究

为了解决煤层含气量定量解释问题,将煤层测井数据与煤心解吸数据作为输入和输出参数,构建深度置信网络(DBN),进而预测煤层含气量。研究以甘肃合水地区测井数据为例,筛选出该地区120组煤层样品作为DBN样本分析数据。选择短源距自然伽马、自然伽马、密度、长源距自然伽马和浅侧向5条测井曲线,作为DBN的输入参数,煤层气含量作为DBN的输出参数,研究RBM数量和隐藏神经元数量对计算结果的影响。并通过概率统计法、BPNN、DBN和SVM计算了30组煤层的煤层气含量,比较不同方法的预测效果。结果表明:①受限玻尔兹曼机(RBM)对DBN计算结果的精度有一定的影响,RBM数量达到7层时,预测结果准确性更高;②选择合适的隐藏层神经元数量,可以保证计算结果的精度和稳定性,神经元数量为20时,预测结果精度更高,稳定性更好;③RBM使得DBN的准确性高于BPNN,此外,DBN的计算准确性和稳定性高于概率统计法和SVM。

Research for Coal Seam Gas Content Well Logging Interpretation Based on Deep Belief Network

To solve coal seam gas content quantitative interpretation issue,taking coal seam well logging data and coal core desorption data as input and output parameters have established deep belief network(DBN),and then predicted coal seam gas content.Taking the well logging data from the Heshui area,Gansu as subject investigated has screened out 120 groups of coal sample from the area as DBN sample analysis data.The short spacing natural gamma,natural gamma,density,long spacing gamma and shallow lateral 5 logging traces have been selected as DBN input parameters,coal seam gas content as DBN output parameters,studied impacts from numbers of RBM and hidden neuron on computed results.Through probabilistic method,BPNN,DBN and SVM have computed 30 groups of coal seam gas content,and then compared predicted effects from different methods.The results have shown that:①restricted Boltzmann machine(RBM)has certain impact on accuracy of DBN computed results,when RBM number achieved 7 layers,accuracy of predicted results will be higher;②to select suitable hidden layer neuron number can ensure computed results accuracy and stability,when neuron number is 20,predicted result will be higher,stability better;③RBM makes DBN accuracy higher than BPNN,besides,DBN computation accuracy and stability higher than that from probabilistic method and SVM.