用多层前馈网络进行三维储层参数反演的方法

地震反演的主要任务是依据地震资料并综合利用地质、测井等资料得到地下地层的详细信息。三维地震反演需要处理庞大的地震数据体 ,同时在反演过程中既要考虑模型和测井的约束 ,又要考虑地震在横向上的连续性。将地震反演看作是地震数据到储层参数的模糊映射 ,并利用神经网络建立了这种映射关系。针对网络收敛速度慢、学习时间长等缺陷 ,提出了一种学习率自适应调整算法。该算法使每个权都有自己的学习率 ,使网络的训练速度大幅度提高。利用该方法进行地震反演 ,抛开了褶积模型的限制 ,也无须已知地震子波。外推过程是在三维空间内进行的 ,所得的储层参数数据体保持了横向上合理自然的连续性。对该数据体进行三维可视化解释 ,可以直接描述储层的空间展布。

Three-dimensional seismic inversion with a fast multi-layer feed-forward neural network

The main task of seismic inversion is to get the detailed subsurface stratigraphic information by the use of seismic, geological and logging data. A huge data cube has to be dealt with in three-dimensional seismic inversion. At the same time, both the constraint of model and well log data and the lateral continuity of seismic data have to be considered in the process of seismic inversion. The seismic inversion was taken as a fuzzy mapping from seismic data to reservoir parameters. This mapping relation can be realized with neural network. In order to solve the drawbacks of standard back-propagation algorithm such as slow converging speed and long time expenditure, an algorithm for learning rate adaptive adjustment was presented. In this algorithm, each weight has its own learning rate. Consequently, the training time was shortened obviously. The new seismic inversion method is independent to the convolution model and wavelet information. Since extrapolation was done in the whole three-dimensional space, the inverted parameter data cube could keep its natural lateral continuity. The spatial distribution of reservoirs could be directly described by three-dimensional visualization interpretation of the data cube.