神经网络反演双侧向电阻率测井曲线的物理约束

Physical Restriction on Model Selection of Neural Network in Dual-Lateral Log Inversion

以一种新型高分辨率双侧向测井仪器在二维轴对称地层模型中的模拟响应为训练集训练BP神经网络 ,得到了针对该双侧向测井仪的反演网络模型 .在训练中 ,神经网络结构的确定一般采用交叉验证法 ,但这种验证法经验性的成分偏重 ,不能完全解决网络结构的范化问题 .为此 ,本文在模型的训练中 ,在交叉验证的基础上 ,根据双侧向测井的原理和仪器响应特性 ,提出了一种新的物理约束方法 :反演地层电阻率的误差应随着侵入半径的增加而加大 ,违反此规律的模型不予采纳 .理论研究结果表明 ,由此得到的神经网络模型具有很好的范化能力 .

Neural network is a intelligent system which is widely used in pattern recognition, data processing and function fitting. We adopted this method in the inversion of formation resistivity from dual lateral logs in this paper. In order to overcome the limition of overfitting and generalization in the application of nueral network which cannot be resovlved based on the available neural network theory, the variation of the response of dual lateral logs with the invasion radius is analyzed so that the law behind the change can be applied to help the selection of neural network structure during the learning. The prediction error should be enlarged with the increasing of invasion radius. This principle is used flowing the conventional cross validation method. The results show that this new restriction on the model selection of neural network will make the neural network method practicable in the inversion of formation resistivity.