量子神经网络模型及在储层识别中的应用

发布时间：2018-05-06 13:46

本文选题：量子神经网络 + 混合量子衍生神经网络　；参考：《东北石油大学》2017年硕士论文

【摘要】：油气层的准确识别是勘探开发的重中之重,大多数识别单纯依靠着专家经验,然而许多的人为因素导致了储层识别的精准度不高。在浪费大量人力物力的同时,却没有达到人们期待的效果。由于储层识别中的影响因素之间是存在着非线性的映射关系,很难用公式来描述其关系,而神经网络能够表达这种映射关系。但是传统神经网络具有一些已经被发现的缺陷,如逼近能力差、易于陷入局部极小值等。在近年来的研究中,有实验可证,将量子计算与神经计算进行融合而出现的新的量子神经网络这一方法对改善神经网络缺陷具有明显作用,因此本文中提出将量子神经网络应用到储层识别中的新方法,以期大幅度提高识别储层识别准确率,并为储层识别问题提供一条新途径。量子神经网络涌现且研究的时间并不是很长,目前可以说是一个全新的领域,且并未完全成熟,进一步深入研究它与其他算法的融合,以期能够进一步提高其性能是十分必要的。基于这一目的,本文拟提出一种基于量子比特在Bloch球面的绕轴旋转构造神经网络模型的新思想,进而研究一种新的量子衍生神经网络模型及算法,把其应用到油气储集层识别问题中。该模型可显著提高其逼近和预测能力。本文主要研究内容如下。首先,设计了一种新型量子神经网络模型。该模型为三层结构,输入层和输出层为普通神经元,隐层为量子神经元。量子神经元的输入为量子比特,关于其映射机制,首先使输入比特绕坐标轴旋转,然后采用泡利矩阵计算旋转后的坐标值,再用Sigmoid函数将坐标值映射为量子神经元的输出。第二,在量子神经网络的训练方面,本文设计了L-M算法和量子蜂群算法。然而这两种算法都有局限性,首先L-M算法收敛较快,但易于陷入局部极小值,量子蜂群算法虽然具有较好的全局寻优能力,但由于采用种群寻优,因此计算效率较低。所以本文提出了一种将两种算法融合的两阶段训练算法。具体研究方案为:首先采用量子蜂群算法实施网络权值的全局探索,然后采用L-M算法实施网络权值的局部开发。第三,针对储层识别问题,研究基于混合量子衍生神经网络的识别方法。首先研究储层分类、储层识别的影响因素,然后结合了矿场实际的测井解释数据,提出了基于混合量子衍生神经网络的储层识别方法。该方法为储层识别问题开辟了新道路。
[Abstract]:The accurate identification of oil and gas reservoir is the most important in exploration and development. Most of the identification is based on the experience of experts. However, many human factors lead to the low accuracy of reservoir identification. Waste a lot of manpower and material resources at the same time, but did not achieve the desired results. Because there is a nonlinear mapping relationship between the influencing factors in reservoir identification, it is difficult to describe the relationship by formula, and the neural network can express this mapping relationship. However, the traditional neural network has some defects that have been found, such as poor approximation ability, easy to fall into local minima and so on. In recent years, it has been proved by experiments that the new method of quantum neural network, which combines quantum computing with neural computing, has obvious effect on improving the defect of neural network. Therefore, a new method of applying quantum neural network to reservoir recognition is proposed in this paper, in order to improve the accuracy of reservoir recognition and provide a new way for reservoir recognition. Quantum neural networks appear and research is not very long, it can be said to be a new field, and not fully mature, further in-depth study of its fusion with other algorithms, It is necessary to further improve its performance. For this purpose, this paper proposes a new idea of constructing a neural network model based on the rotation of quantum bits around the Bloch sphere, and then studies a new quantum derivative neural network model and its algorithm. It is applied to the problem of oil and gas reservoir identification. The model can significantly improve its ability of approximation and prediction. The main contents of this paper are as follows. Firstly, a new quantum neural network model is designed. The model is composed of three layers: the input layer and the output layer are ordinary neurons, and the hidden layer is quantum neurons. The input of a quantum neuron is a quantum bit. As to its mapping mechanism, the input bit is first rotated around the coordinate axis, then the rotated coordinate value is calculated by using Pauli matrix, and then the coordinate value is mapped to the output of the quantum neuron by using the Sigmoid function. Secondly, in the training of quantum neural network, L-M algorithm and quantum bee colony algorithm are designed in this paper. However, both algorithms have their limitations. Firstly, L-M algorithm converges fast, but it is easy to fall into local minimum. Quantum bee colony algorithm has better global optimization ability, but because of population optimization, the computational efficiency is low. So this paper proposes a two-stage training algorithm which combines the two algorithms. The specific schemes are as follows: firstly, the quantum bee colony algorithm is used to implement the global exploration of network weights, and then the L-M algorithm is used to implement the local development of network weights. Thirdly, the recognition method based on hybrid quantum derivative neural network is studied for reservoir identification. Firstly, the reservoir classification and the influencing factors of reservoir identification are studied. Then, a method of reservoir identification based on hybrid quantum derivative neural network is proposed by combining the actual logging interpretation data in the field. This method opens up a new way for reservoir identification.
【学位授予单位】：东北石油大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：P618.13;TP183

【参考文献】