论文部分内容阅读
神经网络可被看成是映射理论的具体应用,即输入空间到输出空间的映射。为了模拟出期望的映射,网络必须具有在对许多输入模式及与它们对应的输出模式的感知基础上进行学习的过程,这一过程指内部参数的迭代变化。对于所有用来训练的样本集而言,当计算出的输出结果与期望输出模式间的误差达到最小时,则完成了网络的训练过程。这样一来,网络在限定的训练样本域中能模拟出期望的映射。本文所介绍的实验是用合成的共炮点道集(即由一个炮点所获得的地震记录)作为网络的输入模式,并将计算出与其对应的一维大尺度速度模型作为输出。所有的地质模型均由8个位于均匀的半空间上的等厚地层组成。用了450个样本来训练网络。经训练的网络虽然不可能计算出与期望输出完全吻合的地质模型,但得到的近似结果完全可用来作为下步地震成像算法中的初始模型。经训练的网络对新的地震道集(即不包括在训练集以内)而言,计算出的速度剖面的精度达80％。虽然输入剖面中有白噪干扰,但对输出结果并无影响,不过在有强的相干噪声情况下,影响很大。本文的应用说明,我们可以使用神经网络对较重要的反问题进行求解。
Neural network can be regarded as the concrete application of mapping theory, that is, mapping from input space to output space. In order to simulate the desired mapping, the network must have a learning process based on the perception of many input modes and the output modes that correspond to them. This process refers to iterative changes in internal parameters. For all sample sets used for training, the network training process is completed when the calculated error between the output result and the expected output pattern is minimized. In this way, the network can simulate the desired mapping in a limited training sample domain. The experiment described in this paper uses the synthesized common shot gathers (ie, the seismic records obtained from a shot) as the input model of the network and calculates the corresponding one-dimensional large-scale velocity model as the output. All geological models consist of eight equal-thickness formations located in a homogeneous half-space. 450 samples were used to train the network. Although it is not possible to calculate a geologic model that exactly matches the expected output, the trained network can be used as an initial model in the next seismic imaging algorithm. Trained Network For new seismic gathers (ie, not included in the training set), the calculated velocity profile has an accuracy of 80%. Although there is white noise in the input profile, it has no effect on the output. However, in the case of strong coherent noise, it has a great impact. The application of this article shows that we can use neural networks to solve the more important inverse problems.