CN106372402A

CN106372402A - A Parallelization Method of Fuzzy Region Convolutional Neural Network in Big Data Environment

Info

Publication number: CN106372402A
Application number: CN201610762101.1A
Authority: CN
Inventors: 李忠伟; 张卫山; 宋弢; 卢清华; 崔学荣; 刘昕; 赵德海; 何旭
Original assignee: China University of Petroleum East China
Current assignee: China University of Petroleum East China
Priority date: 2016-08-30
Filing date: 2016-08-30
Publication date: 2017-02-01
Anticipated expiration: 2036-08-30
Also published as: CN109800863A; CN109800863B; CN106372402B

Abstract

The invention provides a parallelization method of a fuzzy area convolutional neural network in a big data environment, which comprises the steps of firstly, constructing the fuzzy area convolutional neural network, putting a given target hypothesis area and target identification into the same network, sharing convolutional calculation, and updating the weight of the whole network in a training process; next, the input logging data set is divided into a plurality of small data sets, a plurality of workflows are parallelized and are subjected to convolution and pooling operations through a fuzzy region convolution neural network, and each small data set is trained by means of gradient descent alone. The invention optimizes the network structure and parameters, and realizes better analysis performance and precision; in addition, the invention adjusts the number of layers of FR-CNN fuzzification aiming at different logging data sets, so that the extracted characteristics better reflect the characteristics of the oil and gas reservoir, and the problem of the fuzzification of the logging data can be solved; the invention utilizes the multiple GPUs to perform parallel training and execution of the FR-CNN so as to improve the efficiency of the FR-CNN.

Description

A Parallelization Method of Fuzzy Region Convolutional Neural Network in Big Data Environment

技术领域technical field

本发明涉及石油测井技术领域，特别是涉及大数据测井领域。The invention relates to the technical field of petroleum logging, in particular to the field of big data logging.

背景技术Background technique

测井信息和沉积是地层岩石物理性质的反映和控制因素，因此测井资料一直以来被作为油气储层沉积学研究中基础而重要的信息来源，测井相则是测井信息与储层沉积学特征之间的桥梁。对于大部分的油气井来说，测井资料是仅有的覆盖全井段地层的综合信息来源，因此测井相识别分析方法一直作为油气勘探与开发地质研究中一个最重要的研究手段。Well logging information and deposition are the reflection and controlling factors of formation rock physical properties, so well logging data has always been regarded as the basic and important source of information in the study of oil and gas reservoir sedimentology, and logging facies are logging information and reservoir deposition bridges between academic features. For most oil and gas wells, well logging data is the only comprehensive information source covering the whole well section, so the logging facies identification analysis method has always been the most important research method in the geological research of oil and gas exploration and development.

然而，测井信息具有模糊性的特点，具有地质意义的多解性和模糊性。因此，测井相的识别与分析必须建立在大量已有的沉积特征与测井参数关系(测井响应)综合深度分析基础之上，同时还要参考野外露头、岩心录井和地震分析的结果，选取适合地质特点的建模方法，才能实现测井相的准确识别。However, well logging information has the characteristics of ambiguity, multi-solution and ambiguity in geological significance. Therefore, the identification and analysis of logging facies must be based on a large number of comprehensive in-depth analyzes of the relationship between sedimentary characteristics and logging parameters (logging response), and also refer to the results of field outcrop, core logging and seismic analysis. Only by selecting a modeling method suitable for geological characteristics can the accurate identification of logging facies be realized.

此外，由于缺乏有效的测井相自动识别方法和技术，目前的测井相识别主要是通过地质工作人员的人工识别实现的，并且由于人员经验差异、主观差异、测井数据的系统差异等因素，地质人员面对的数据量大、工作量繁重。不仅如此，地质人员的经验差异、主观因素、不同时期不同仪器测井数据的系统差异等因素，使得传统的测井相识别准确性大打折扣。In addition, due to the lack of effective automatic identification methods and technologies for well logging facies, the current identification of well logging facies is mainly realized through the manual identification of geological staff, and due to factors such as personnel experience differences, subjective differences, and logging data system differences , Geologists are faced with a large amount of data and a heavy workload. Not only that, factors such as differences in experience of geological personnel, subjective factors, and systematic differences in logging data of different tools in different periods have greatly reduced the accuracy of traditional logging facies identification.

将大数据分析、深度学习等先进技术应用于油气地质研究是解决当前石油行业大数据分析资源闲置的探索与尝试。近年来，石油行业建立了大量的云数据中心，但利用率不高，资源被严重浪费。其中一个重要原因就是缺乏大数据处理平台以及相应的大数据技术来充分利用这些计算、存储资源。Applying advanced technologies such as big data analysis and deep learning to oil and gas geological research is an exploration and attempt to solve the current idle big data analysis resources in the petroleum industry. In recent years, the petroleum industry has established a large number of cloud data centers, but the utilization rate is not high, and resources are seriously wasted. One of the important reasons is the lack of big data processing platforms and corresponding big data technologies to make full use of these computing and storage resources.

建立高效、准确的测井相识别方法是现在油气地质研究的迫切需求。Establishing an efficient and accurate logging facies identification method is an urgent need for oil and gas geological research.

发明内容Contents of the invention

为解决现有技术的不足，本发明提出了一种大数据环境下模糊区域卷积神经网络的并行化方法。In order to solve the deficiencies of the prior art, the present invention proposes a parallelization method of fuzzy region convolutional neural network under the environment of big data.

本发明的技术方案是这样实现的：Technical scheme of the present invention is realized like this:

一种大数据环境下模糊区域卷积神经网络的并行化方法，首先，构建模糊区域卷积神经网络，将给出目标假设区域和目标识别放入同一个网络中，共享卷积计算，一个训练过程更新整个网络的权重；A parallelization method of fuzzy region convolutional neural network in a big data environment. First, construct a fuzzy region convolutional neural network, put the given target hypothetical region and target recognition into the same network, share the convolution calculation, and train The process updates the weights of the entire network;

接下来，把输入的测井数据集分割成若干小数据集，多个工作流并行化经过模糊区域卷积神经网络进行卷积和池化操作，每一小数据集单独利用梯度下降进行训练；训练完成后，把结果输出到等待队列，在一轮训练完成后，读取输出队列，进行共享权重的同步更新操作，更新完成后，进行下一轮训练；在每一轮训练中，对于每个分割的小数据集的计算，都是在分布式基础上异步进行的，每计算出梯度值，就追加到列表当中来，当所有的小数据集都计算完毕后，同步更新模糊区域卷积神经网络的权重和偏置值，然后进行下一轮训练；在并行化识别方面，由Spout收集测井数据，然后将数据分发到各个Bolt节点中并行进行测井相识别，每个Bolt节点将识别结果输入到下一个Bolt节点中，统计其中的物体信息；Next, the input logging data set is divided into several small data sets, and multiple workflows are parallelized through the fuzzy area convolutional neural network for convolution and pooling operations, and each small data set is individually trained using gradient descent; After the training is completed, the result is output to the waiting queue. After a round of training is completed, the output queue is read to perform a synchronous update operation of shared weights. After the update is completed, the next round of training is performed; in each round of training, for each The calculation of each divided small data set is performed asynchronously on a distributed basis. Every time the gradient value is calculated, it is added to the list. When all the small data sets are calculated, the fuzzy area convolution is updated synchronously. The weight and bias value of the neural network, and then the next round of training; in terms of parallel identification, the Spout collects logging data, and then distributes the data to each Bolt node for parallel logging facies identification, and each Bolt node will The recognition result is input to the next Bolt node, and the object information in it is counted;

每一个小数据集经过模糊区域卷积神经网络进行卷积和池化操作的步骤，具体包括：卷积层和池化层交互，在卷积层和池化层进行模糊操作，从模糊区域卷积神经网络的第一层开始，逐渐增加模糊化的层数，针对不同的数据集调整模糊化层数，模糊区域卷积神经网络的最后一层得到特征向量，该特征向量通过一个滑动窗口将特征映射到一个低维向量中，然后将特征输入到两个全连接层，一个全连接层用来定位，另一个全连接层用来分类。The steps of convolution and pooling operation of each small data set through the fuzzy area convolution neural network include: the interaction between the convolutional layer and the pooling layer, the fuzzy operation in the convolutional layer and the pooling layer, and the convolution from the fuzzy area Starting from the first layer of convolutional neural network, the number of fuzzy layers is gradually increased, and the number of fuzzy layers is adjusted for different data sets. The last layer of convolutional neural network in the fuzzy area obtains a feature vector, and the feature vector is passed through a sliding window. The features are mapped into a low-dimensional vector, and then the features are input into two fully connected layers, one for localization and the other for classification.

可选地，所述卷积层公式表达为：Optionally, the convolution layer formula is expressed as:

${v v}_{i i j j}^{x x y the y} = = f f (({\overset{&OverBar; &OverBar;}{b b}}_{i i j j} + + {Σ Σ}_{m m} {Σ Σ}_{p p = = 00}^{{P P}_{i i} - - 11} {Σ Σ}_{q q = = 00}^{{Q Q}_{i i} - - 11} {\overset{&OverBar; &OverBar;}{W W}}_{i i j j m m}^{p p q q} {v v}_{((i i - - 11)) m m}^{((x x + + p p)) ((y the y + + q q))}))$

池化层公式表达为：The pooling layer formula is expressed as:

${x x}_{j j} = = f f (({\overset{&OverBar; &OverBar;}{β β}}_{i i j j} d d o o w w n no (({x x}_{i i - - 11 j j})) + + {\overset{&OverBar; &OverBar;}{b b}}_{i i j j}))$

其中，偏置和权重均为模糊数，这里使用对称三角模糊数，为模糊数组成的向量，第j个模糊数的隶属函数为：Among them, bias and weight Both are fuzzy numbers, here we use symmetric triangular fuzzy numbers, is a vector composed of fuzzy numbers, the jth fuzzy number The membership function of is:

${\overset{&OverBar; &OverBar;}{W W}}_{j j} ((w w)) = = max max {{11 - - \frac{| | w w - - {w w}_{j j} | |}{{\overset{^^}{w w}}_{j j}},, 00}} . .$

可选地，在模糊区域卷积神经网络的训练过程中，定义一个联合损失函数：Optionally, a joint loss function is defined during the training of the convolutional neural network in blurred regions:

$L L (({{{p p}_{i i}}},, {{{t t}_{i i}}})) = = \frac{11}{{N N}_{c c l l s the s}} {Σ Σ}_{i i} {L L}_{c c l l s the s} (({p p}_{i i},, {p p}_{i i}^{* *})) + + λ λ \frac{11}{{N N}_{r r e e g g}} {Σ Σ}_{i i} {p p}_{i i}^{* *} {L L}_{r r e e g g} (({t t}_{i i},, {t t}_{i i}^{* *}))$

其中，p_i是此样本为测井曲线形态的预测概率，是样本的标签，如果是相应的测井曲线形态，为1，否则为0，N_cls是二分类逻辑损失；t_i是预测物体边界的四个参数组成的向量，为标注区域参数组成的向量，它们分别为：Among them, p _i is the predicted probability that this sample is the shape of the well log curve, is the label of the sample, if it is the corresponding log shape, is 1, otherwise is 0, N _cls is the binary classification logic loss; t _i is a vector composed of four parameters for predicting the boundary of the object, is a vector of parameters for the label area, they are:

t_x＝(x-x_a)/w_a t_h＝(y-y_a)/h_a t _x ＝(xx _a )/w _a t _h ＝(yy _a )/h _a

t_w＝log(w/w_a) t_h＝log(h/h_a)t _w =log(w/w _a ) t _h =log(h/h _a )

$\begin{matrix} {t t}_{w w}^{* *} = = (({w w}^{* *} / / {w w}_{a a})) / / {w w}_{a a} & {t t}_{y the y}^{* *} = = ((y the y * * - - {y the y}_{a a})) / / {h h}_{a a} \end{matrix}$

$\begin{matrix} {t t}_{w w}^{* *} = = l l o o g g (({w w}^{* *} / / {w w}_{a a})) & {t t}_{h h}^{* *} = = l l o o g g ((h h * * / / {h h}_{a a})) \end{matrix}$

其中，x、y、w和h分别代表物体的中心坐标、宽度和长度，x，x_a，x^*分别代表预测区域，锚定区域和标注区域，回归损失R为平滑损失函数 Among them, x, y, w, and h represent the center coordinates, width, and length of the object, respectively, x, x _a , x ^* represent the prediction area, anchor area, and label area, respectively, and the regression loss R is the smoothing loss function

${smooth smooth}_{{L L}_{11}} ((x x)) = = \{\begin{matrix} 0.5 0.5 {x x}^{22} & i i f f | | x x | | < < 11 \\ | | x x | | - - 0.5 0.5 & o o t t h h e e r r w w i i s the s e e \end{matrix}$

表示只有当锚定区域为正样本时，才计算回归损失，否则不计算，归一化参数N_cls和N_reg分别代表从特征向量映射的低维向量的长度和锚定区域的数量。 Indicates that only when the anchor region is a positive sample When , the regression loss is calculated, otherwise Not calculated, the normalization parameters N _cls and N _reg represent the length of the low-dimensional vector mapped from the feature vector and the number of anchor regions, respectively.

可选地，首先进行测井数据的规范化，将原始数据均转换为无量纲化指标测评值，各指标测评值都处于同一个数量级别上，再进行综合测评分析。Optionally, the logging data is normalized first, and the original data are converted into dimensionless index evaluation values. The evaluation values of each index are at the same quantitative level, and then comprehensive evaluation and analysis are performed.

可选地，进行测井数据的规范化采用如下的规范化方法：Optionally, the normalization of logging data adopts the following normalization method:

Sx＝(x-M)/S,x∈{GR，AC，DEN，CNL，SDN，...}Sx=(x-M)/S, x∈{GR, AC, DEN, CNL, SDN,...}

其中，x表示每条测井曲线的数据，Sx表示规范化后的测井曲线数据，M为相应测井曲线数据的均值，S为每条测井曲线数据的标准差。Among them, x represents the data of each well log curve, Sx represents the normalized well log data, M is the mean value of the corresponding well log data, and S is the standard deviation of each well log data.

本发明的有益效果是：The beneficial effects of the present invention are:

(1)根据测井大数据中数据模糊的特点，融入模糊理论，提出模糊区域卷积神经网络FR-CNN，并提出渐进模糊的方法，从模糊区域卷积神经网络的第一层开始，逐渐增加模糊化的层数，从而优化网络结构和参数，实现更好的分析性能和精度；(1) According to the characteristics of fuzzy data in well logging big data, the fuzzy area convolutional neural network FR-CNN is proposed by incorporating the fuzzy theory, and a progressive fuzzy method is proposed, starting from the first layer of the fuzzy area convolutional neural network, gradually Increase the number of fuzzy layers to optimize the network structure and parameters to achieve better analysis performance and accuracy;

(2)针对不同的测井数据集调整FR-CNN模糊化的层数，使提取的特征更好的反映油气储层本身的特性，可以解决测井数据模糊性问题；(2) Adjust the number of fuzzy layers of FR-CNN for different logging data sets, so that the extracted features can better reflect the characteristics of the oil and gas reservoir itself, and can solve the fuzzy problem of logging data;

(3)本发明利用多GPU进行FR-CNN的并行训练和执行，以提高FR-CNN的效率。(3) The present invention utilizes multiple GPUs to perform parallel training and execution of FR-CNN to improve the efficiency of FR-CNN.

附图说明Description of drawings

为了更清楚地说明本发明实施例或现有技术中的技术方案，下面将对实施例或现有技术描述中所需要使用的附图作简单地介绍，显而易见地，下面描述中的附图仅仅是本发明的一些实施例，对于本领域普通技术人员来讲，在不付出创造性劳动的前提下，还可以根据这些附图获得其他的附图。In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings that need to be used in the description of the embodiments or the prior art. Obviously, the accompanying drawings in the following description are only These are some embodiments of the present invention. Those skilled in the art can also obtain other drawings based on these drawings without creative work.

图1为本发明模糊区域卷积神经网络的结构示意图；Fig. 1 is the structural representation of fuzzy area convolutional neural network of the present invention;

图2为对称三角模糊数坐标示意图；Fig. 2 is a schematic diagram of symmetrical triangular fuzzy number coordinates;

图3为本发明模糊区域卷积神经网络并行化处理实时数据的原理图。Fig. 3 is a principle diagram of the parallel processing of real-time data by the fuzzy region convolutional neural network of the present invention.

具体实施方式detailed description

下面将结合本发明实施例中的附图，对本发明实施例中的技术方案进行清楚、完整地描述，显然，所描述的实施例仅仅是本发明一部分实施例，而不是全部的实施例。基于本发明中的实施例，本领域普通技术人员在没有做出创造性劳动前提下所获得的所有其他实施例，都属于本发明保护的范围。The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only some, not all, embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by persons of ordinary skill in the art without making creative efforts belong to the protection scope of the present invention.

测井数据具有模糊性的特点，造成这种模糊性的原因是多方面的，包括噪音、不一致性、不完整性等造成的测井数据的数据空间污染，也包括不同时期、不同仪器测井带来的系统性数据差异，这些问题带来的测井数据模糊性都制约了测井相的准确识别。Well logging data is characterized by ambiguity, and there are many reasons for this ambiguity, including the data space pollution of well logging data caused by noise, inconsistency, incompleteness, etc. The systematic data differences brought about by these problems and the ambiguity of logging data caused by these problems all restrict the accurate identification of logging facies.

本发明提出了一种大数据环境下模糊区域卷积神经网络的并行化方法，对测井数据构建出多维度的数据空间，将模糊理论与深度学习网络R-CNN融合，提出解决模糊数据情况下测井相的识别方法，根据测井大数据中数据模糊的特点，融入模糊理论，提出模糊区域卷积神经网络FR-CNN(Fuzzy R-CNN)，进一步提出渐进模糊方法，从卷积神经网络第一层开始，逐渐增加模糊化的层数，优化网络结构和参数，最终建立FR-CNN的理论和方法，实现更好的分析性能和精度，同时，本发明利用多GPU进行FR-CNN的并行训练和执行，以提高FR-CNN的效率。The present invention proposes a parallelization method of fuzzy area convolutional neural network in a big data environment, constructs a multi-dimensional data space for well logging data, integrates fuzzy theory with deep learning network R-CNN, and proposes to solve the fuzzy data situation The identification method of down-logging facies, according to the fuzzy characteristics of data in well-logging big data, integrates fuzzy theory, proposes fuzzy region convolutional neural network FR-CNN (Fuzzy R-CNN), and further proposes a progressive fuzzy method. Starting from the first layer of the network, gradually increase the number of fuzzy layers, optimize the network structure and parameters, and finally establish the theory and method of FR-CNN to achieve better analysis performance and accuracy. At the same time, the present invention uses multiple GPUs to perform FR-CNN Parallel training and execution of FR-CNN to improve the efficiency of FR-CNN.

设计合适的模糊区域卷积神经网络是本发明的重点，下面对本发明模糊区域卷积神经网络的构建进行详细说明。Designing a suitable fuzzy region convolutional neural network is the key point of the present invention, and the construction of the fuzzy region convolutional neural network of the present invention will be described in detail below.

模糊区域卷积神经网络FR-CNN建立于深度学习网络R-CNN的基础之上，如图1所示，FR-CNN将给出目标假设区域和目标识别放入同一个网络中，共享卷积计算，避免复杂的计算步骤，只需要一个训练过程便可以更新整个网络的权重，同时也加快检测速度，达到快速处理的目的。The fuzzy area convolutional neural network FR-CNN is built on the basis of the deep learning network R-CNN. As shown in Figure 1, FR-CNN puts the target hypothetical area and target recognition into the same network, sharing the convolution Calculation, avoiding complex calculation steps, only one training process is required to update the weight of the entire network, and at the same time speed up the detection speed to achieve the purpose of fast processing.

图1中，测井数据经过模糊区域卷积神经网络，进行卷积和池化操作。模糊区域卷积神经网络训练的核心在于卷积层和池化层的交互，因此在卷积层和池化层进行模糊操作。为了避免模糊过度导致的信息损失过多，并且考虑到模糊区域卷积神经网络提取特征的精细化程度逐层降低，这里改变传统模糊神经网络对每一层的模糊化，本发明提出渐进模糊的方法，即从模糊区域卷积神经网络的第一层开始，逐渐增加模糊化的层数，针对不同的数据集调整模糊化层数，使提取的特征更好的反映测井曲线的特性，从而得到最佳识别结果，并提高识别效率。In Fig. 1, the well logging data is subjected to convolution and pooling operations through the convolutional neural network in the fuzzy region. The core of the convolutional neural network training in the fuzzy area lies in the interaction between the convolutional layer and the pooling layer, so the fuzzy operation is performed on the convolutional layer and the pooling layer. In order to avoid excessive information loss caused by excessive fuzziness, and considering that the refinement degree of feature extraction by convolutional neural network in the fuzzy area decreases layer by layer, the fuzzification of each layer by the traditional fuzzy neural network is changed here, and the present invention proposes a progressive fuzzy method. method, starting from the first layer of the convolutional neural network in the fuzzy area, gradually increasing the number of fuzzy layers, adjusting the number of fuzzy layers for different data sets, so that the extracted features can better reflect the characteristics of the logging curve, so that Get the best recognition result and improve the recognition efficiency.

模糊区域卷积神经网络的最后一层得到特征向量，该特征向量通过一个小的滑动窗口将特征映射到一个低维向量中，然后将特征输入到两个全连接层，一个全连接层用来定位，另一个全连接层用来分类。在每一个滑动窗口处同时给出几个目标假设区域，可称之为锚定区域，这个区域以滑动窗口为中心，拥有不同的横纵比和缩放比例。The last layer of the convolutional neural network in the fuzzy area obtains the feature vector, which maps the feature into a low-dimensional vector through a small sliding window, and then inputs the feature into two fully connected layers, one fully connected layer is used for localization, and another fully connected layer for classification. At each sliding window, several target hypothetical areas are given at the same time, which can be called the anchor area. This area is centered on the sliding window and has different aspect ratios and zoom ratios.

卷积神经网络R-CNN的卷积层公式可以表达为：The convolutional layer formula of the convolutional neural network R-CNN can be expressed as:

${v v}_{i i j j}^{x x y the y} = = f f (({b b}_{i i j j} + + {Σ Σ}_{m m} {Σ Σ}_{p p = = 00}^{{P P}_{i i} - - 11} {Σ Σ}_{q q = = 00}^{{Q Q}_{i i} - - 11} {W W}_{i i j j m m}^{p p q q} {v v}_{((i i - - 11)) m m}^{((x x + + p p)) ((y the y + + q q))})) - - - - - - ((11))$

其中，表示的是在第i层神经元的第j个特征向量的(x,y)位置处的值，表示连接到第m个特征向量的卷积核在位置(p,q)上的权值。P_i和Q_i分别表示卷积核的高度和宽度，b_ij为偏置项，f(x)表示神经元的激活函数。in, Represents the value at the (x, y) position of the jth eigenvector of the i-th layer neuron, express The weight of the filter at position (p,q) connected to the mth eigenvector. P _i and Q _i represent the height and width of the convolution kernel, respectively, _bij is the bias item, and f(x) represents the activation function of the neuron.

R-CNN池化层公式表达为：The R-CNN pooling layer formula is expressed as:

x_ij＝f(β_ijdown(x_i-1j)+b_ij) (2)x _ij ＝f(β _ij down(x _i-1j )+b _ij ) (2)

down(.)表示一个下采样函数，典型的操作一般是对输入数据的不同n*n块的所有信息进行求和，这样输出数据在两个维度上都缩小了n倍，每个输出map都对应一个属于自己的乘性偏置β和一个加性偏置b。down(.) represents a downsampling function. A typical operation is to sum all the information of different n*n blocks of input data, so that the output data is reduced by n times in both dimensions, and each output map is Corresponding to its own multiplicative bias β and an additive bias b.

卷积神经网络的输入和计算过程都是实数，得到的结果都是确定性的，而对于数据缺失等数据模糊的情况，本发明的模糊区域卷积神经网络中引入模糊理论，改进的公式如下：The input and calculation process of the convolutional neural network are all real numbers, and the results obtained are all deterministic, and for data ambiguity such as missing data, fuzzy theory is introduced in the fuzzy area convolutional neural network of the present invention, and the improved formula is as follows :

卷积层公式表达为：The convolutional layer formula is expressed as:

${v v}_{i i j j}^{x x y the y} = = f f (({\overset{&OverBar; &OverBar;}{b b}}_{i i j j} + + {Σ Σ}_{m m} {Σ Σ}_{p p = = 00}^{{P P}_{i i} - - 11} {Σ Σ}_{q q = = 00}^{{Q Q}_{i i} - - 11} {\overset{&OverBar; &OverBar;}{W W}}_{i i j j m m}^{p p q q} {v v}_{((i i - - 11)) m m}^{((x x + + p p)) ((y the y + + q q))})) - - - - - - ((33))$

池化层公式表达为：The pooling layer formula is expressed as:

${x x}_{j j} = = f f (({\overset{&OverBar; &OverBar;}{β β}}_{i i j j} d d o o w w n no (({x x}_{i i - - 11 j j})) + + {\overset{&OverBar; &OverBar;}{b b}}_{i i j j})) - - - - - - ((44))$

其中偏置和权重均为模糊数，这里使用对称三角模糊数，为模糊数组成的向量，第j个模糊数的隶属函数为which biases and weight Both are fuzzy numbers, here we use symmetric triangular fuzzy numbers, is a vector composed of fuzzy numbers, the jth fuzzy number The membership function of is

${\overset{&OverBar; &OverBar;}{W W}}_{j j} ((w w)) = = max max {{11 - - \frac{| | w w - - {w w}_{j j} | |}{{\overset{^^}{w w}}_{j j}},, 00}} - - - - - - ((55))$

如图2所示，w_j是模糊数的对称中心，是模糊数的半长，代表w处的隶属度。As shown in Figure 2, w _j is the symmetry center of the fuzzy number, is the half length of the fuzzy number, represents the degree of membership at w.

在模糊区域卷积神经网络的训练过程中，定义一个联合损失函数：During the training process of the convolutional neural network in the fuzzy area, a joint loss function is defined:

$L L (({{{p p}_{i i}}},, {{{t t}_{i i}}})) = = \frac{11}{{N N}_{c c l l s the s}} {Σ Σ}_{i i} {L L}_{c c l l s the s} (({p p}_{i i},, {p p}_{i i}^{* *})) + + λ λ \frac{11}{{N N}_{r r e e g g}} {Σ Σ}_{i i} {p p}_{i i}^{* *} {L L}_{r r e e g g} (({t t}_{i i},, {t t}_{i i}^{* *})) - - - - - - ((66))$

其中p_i是此样本为测井曲线形态的预测概率，是样本的标签，如果是相应的测井曲线形态，为1，否则为0，N_cls是二分类(0或1)逻辑损失。where p _i is the predicted probability that this sample is the shape of the log curve, is the label of the sample, if it is the corresponding log shape, is 1, otherwise is 0, N _cls is the binary classification (0 or 1) logistic loss.

t_i是预测物体边界的四个参数组成的向量，为标注区域参数组成的向量，它们分别为：t _i is a vector composed of four parameters to predict the boundary of the object, is a vector composed of parameters of the label area, which are:

t_x＝(x-x_a)/w_a t_h＝(y-y_a)/h_a (7)t _x ＝(xx _a )/w _a t _h ＝(yy _a )/h _a (7)

t_w＝log(w/w_a)t_h＝log(h/h_a)t _w =log(w/w _a )t _h =log(h/h _a )

$\begin{matrix} {t t}_{x x}^{* *} = = (({x x}^{* *} - - {x x}_{a a})) / / {w w}_{a a} & {t t}_{y the y}^{* *} = = (({y the y}^{* *} - - {y the y}_{a a})) / / {h h}_{a a} \end{matrix}$

其中x、y、w和h分别代表物体的中心坐标、宽度和长度，x，x_a，x^*分别代表预测区域，锚定区域和标注区域(y，w，h同理)。回归损失R为平滑损失函数 Where x, y, w, and h represent the center coordinates, width, and length of the object, respectively, and x, x _a , and x ^* represent the prediction area, anchor area, and label area respectively (y, w, h are the same). regression loss R is the smoothing loss function

${smooth smooth}_{{L L}_{11}} ((x x)) = = \{\begin{matrix} 0.5 0.5 {x x}^{22} & i i f f | | x x | | < < 11 \\ | | x x | | - - 0.5 0.5 & o o t t h h e e r r w w i i s the s e e \end{matrix} - - - - - - ((88))$

表示只有当锚定区域为正样本时才计算回归损失，否则不计算。归一化参数N_cls和N_reg分别代表从特征向量映射的低维向量的长度和锚定区域的数量。 Indicates that only when the anchor region is a positive sample Only calculate the regression loss, otherwise Not counted. The normalization parameters N _cls and N _reg represent the length of the low-dimensional vector mapped from the feature vector and the number of anchor regions, respectively.

采用不同的测井手段会产生不同的数据。如采用自然伽马(GR)、补偿声波(AC)、补偿密度(DEN)、补偿中子(CNL)及中子视孔隙度与密度视孔隙度差(SDN)等具有不同的量纲，数据之间不具有可比性，因此，本发明需要首先进行测井数据的规范化，将原始数据均转换为无量纲化指标测评值，即各指标测评值都处于同一个数量级别上，再进行综合测评分析。Using different logging methods will produce different data. For example, natural gamma ray (GR), compensated acoustic wave (AC), compensated density (DEN), compensated neutron (CNL) and neutron apparent porosity and density apparent porosity difference (SDN) have different dimensions, and the data There is no comparability between them, therefore, the present invention needs to standardize the logging data first, and convert the original data into dimensionless index evaluation values, that is, the evaluation values of each index are on the same magnitude level, and then carry out comprehensive evaluation analyze.

采用如下的规范化方法：Normalize as follows:

Sx＝(x-M)/S,x∈{GR，AC，DEN，CNL，SDN，...}Sx=(x-M)/S, x ∈ {GR, AC, DEN, CNL, SDN, ...}

其中，x表示每条测井曲线的数据，Sx表示规范化后的测井曲线数据；M为相应测井曲线数据的均值，S为每条测井曲线数据的标准差。Among them, x represents the data of each well log curve, Sx represents the normalized well log data; M is the mean value of the corresponding well log data, and S is the standard deviation of each well log data.

本发明从已有的测井数据中进行标定，建立FR-CNN的训练数据集，在此基础上，由于不同的测井方法所揭示的信息不尽相同，所以选择不同测井数据的组合作为FR-CNN的输入，从而确定FR-CNN最优的测井数据组合，并优化FR-CNN的进行测井相识别时的网络参数和结构。The present invention calibrates from the existing logging data and establishes the training data set of FR-CNN. On this basis, because the information revealed by different logging methods is not the same, the combination of different logging data is selected as the FR-CNN's input, so as to determine the optimal logging data combination of FR-CNN, and optimize the network parameters and structure of FR-CNN for logging facies identification.

FR-CNN比传统卷积神经网络多了两类全连接层，还多了区域坐标的计算等操作，这些操作计算量都很大。在模糊神经网络中模糊操作存在于网络的每一层，也就是说网络越深所增加的计算量就越多，这就使本来需要繁重计算的网络显得笨重。计算量的增加导致网络的训练时间大幅增长，延长了网络模型更新的周期，削弱系统的灵活性，同时检测时间也会加长。Compared with the traditional convolutional neural network, FR-CNN has two more types of fully connected layers, and more operations such as the calculation of regional coordinates. These operations are computationally intensive. In the fuzzy neural network, the fuzzy operation exists in each layer of the network, that is to say, the deeper the network, the more the amount of calculation will be increased, which makes the network that originally needs heavy calculation appear cumbersome. The increase in the amount of calculation leads to a substantial increase in the training time of the network, prolonging the update cycle of the network model, weakening the flexibility of the system, and prolonging the detection time.

本发明通过并行化提高FR-CNN训练和运行效率，首先把输入的测井数据集分割成若干小数据集，多个工作流同时运行，每一部分单独利用梯度下降进行训练。训练完成后，把结果输出到等待队列，在一轮训练完成后，读取输出队列，进行共享权重的同步更新操作。更新完成后，进行下一轮训练。The invention improves the training and operation efficiency of FR-CNN through parallelization. Firstly, the input logging data set is divided into several small data sets, and multiple workflows run simultaneously, and each part is independently trained by gradient descent. After the training is completed, the result is output to the waiting queue. After a round of training is completed, the output queue is read to perform a synchronous update operation of the shared weights. After the update is completed, proceed to the next round of training.

在每一轮训练中，对于每个分割的小数据集的计算，都是在分布式基础上异步进行的，每计算出梯度值，就追加到列表当中来，当所有的小数据集都计算完毕后，同步更新网络的权重和偏置值，然后进行下一轮训练。In each round of training, the calculation of each divided small data set is performed asynchronously on a distributed basis. Every time the gradient value is calculated, it is added to the list. When all the small data sets are calculated After completion, update the weights and bias values of the network synchronously, and then proceed to the next round of training.

如图3所示，在并行化识别方面，采取的解决方案为：由Spout收集测井数据，然后将数据分发到各个Bolt节点中并行进行测井相识别，每个Bolt节点将识别结果输入到下一个Bolt节点中，统计其中的物体信息。As shown in Fig. 3, in terms of parallel recognition, the solution adopted is: the spout collects logging data, and then distributes the data to each Bolt node for parallel logging facies recognition, and each Bolt node inputs the recognition results to In the next Bolt node, the object information in it is counted.

本发明根据测井大数据中数据模糊的特点，融入模糊理论，提出模糊区域卷积神经网络FR-CNN，并提出渐进模糊的方法，从模糊区域卷积神经网络的第一层开始，逐渐增加模糊化的层数，从而优化网络结构和参数，实现更好的分析性能和精度；而且，本发明针对不同的测井数据集调整FR-CNN模糊化的层数，使提取的特征更好的反映油气储层本身的特性，可以解决测井数据模糊性问题；针对操作计算量大的问题，本发明利用多GPU进行FR-CNN的并行训练和执行，以提高FR-CNN的效率。According to the characteristics of fuzzy data in well logging big data, the present invention integrates fuzzy theory, proposes fuzzy area convolutional neural network FR-CNN, and proposes a progressive fuzzy method, starting from the first layer of fuzzy area convolutional neural network, gradually increasing The number of fuzzy layers, thereby optimizing the network structure and parameters, to achieve better analysis performance and accuracy; moreover, the present invention adjusts the number of fuzzy layers of FR-CNN for different logging data sets, so that the extracted features are better Reflecting the characteristics of the oil and gas reservoir itself, it can solve the problem of ambiguity of logging data; for the problem of large amount of operation and calculation, the present invention uses multiple GPUs to perform parallel training and execution of FR-CNN to improve the efficiency of FR-CNN.

以上所述仅为本发明的较佳实施例而已，并不用以限制本发明，凡在本发明的精神和原则之内，所作的任何修改、等同替换、改进等，均应包含在本发明的保护范围之内。The above descriptions are only preferred embodiments of the present invention, and are not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present invention shall be included in the scope of the present invention. within the scope of protection.

Claims

1. the parallelization method of fuzzy area convolutional neural network under a kind of big data environment, it is characterized in that,

First, construct a fuzzy area convolutional neural network, put the given target hypothetical area and target recognition into the same network, share the convolution calculation, and update the weight of the entire network in one training process;

Next, the input logging data set is divided into several small data sets, and multiple workflows are parallelized through the fuzzy area convolutional neural network for convolution and pooling operations, and each small data set is individually trained using gradient descent; After the training is completed, the result is output to the waiting queue. After a round of training is completed, the output queue is read to perform a synchronous update operation of shared weights. After the update is completed, the next round of training is performed; in each round of training, for each The calculation of each divided small data set is performed asynchronously on a distributed basis. Every time the gradient value is calculated, it is added to the list. When all the small data sets are calculated, the fuzzy area convolution is updated synchronously. The weight and bias value of the neural network, and then the next round of training; in terms of parallel identification, the Spout collects logging data, and then distributes the data to each Bolt node for parallel logging facies identification, and each Bolt node will The recognition result is input to the next Bolt node, and the object information in it is counted;

The steps of convolution and pooling operation of each small data set through the fuzzy area convolution neural network include: the interaction between the convolutional layer and the pooling layer, the fuzzy operation in the convolutional layer and the pooling layer, and the convolution from the fuzzy area Starting from the first layer of convolutional neural network, the number of fuzzy layers is gradually increased, and the number of fuzzy layers is adjusted for different data sets. The last layer of convolutional neural network in the fuzzy area obtains a feature vector, and the feature vector is passed through a sliding window. The features are mapped into a low-dimensional vector, and then the features are input into two fully connected layers, one for localization and the other for classification.

2. the parallelization method of fuzzy region convolutional neural network under a kind of big data environment as claimed in claim 1, is characterized in that, described convolution layer formula is expressed as:

{v v}_{i i j j}^{x x y the y} = = f f (({\overset{&OverBar; &OverBar;}{b b}}_{i i j j} + + {Σ Σ}_{m m} {Σ Σ}_{p p = = 00}^{{P P}_{i i} - - 11} {Σ Σ}_{q q = = 00}^{{Q Q}_{i i} - - 11} {\overset{&OverBar; &OverBar;}{W W}}_{i i j j m m}^{p p q q} {v v}_{((i i - - 11)) m m}^{((x x + + p p)) ((y the y + + q q))}))

The pooling layer formula is expressed as:

{x x}_{j j} = = f f (({\overset{&OverBar; &OverBar;}{β β}}_{i i j j} d d o o w w n no (({x x}_{i i - - 11 j j})) + + {\overset{&OverBar; &OverBar;}{b b}}_{i i j j}))

Among them, bias and weight Both are fuzzy numbers, and symmetric triangular fuzzy numbers are used here, is a vector composed of fuzzy numbers, the jth fuzzy number The membership function of is:

{\overset{&OverBar; &OverBar;}{W W}}_{j j} ((w w)) = = max max {{11 - - \frac{| | w w - - {w w}_{j j} | |}{{\overset{^^}{w w}}_{j j}},, 00}} . .

3. the parallelization method of fuzzy region convolution neural network under a kind of big data environment as claimed in claim 2, is characterized in that, in the training process of fuzzy region convolution neural network, define a joint loss function:

L L (({{{p p}_{i i}}},, {{{t t}_{i i}}})) = = \frac{11}{{N N}_{c c l l s the s}} {Σ Σ}_{i i} {L L}_{c c l l s the s} (({p p}_{i i},, {p p}_{i i}^{* *})) + + λ λ \frac{11}{{N N}_{r r e e g g}} {Σ Σ}_{i i} {p p}_{i i}^{* *} {L L}_{r r e e g g} (({t t}_{i i},, {t t}_{i i}^{* *}))

Among them, p _i is the predicted probability that this sample is the shape of the well log curve, is the label of the sample, if it is the corresponding log shape, is 1, otherwise is 0, N _cls is the binary classification logic loss; t _i is a vector composed of four parameters for predicting the boundary of the object, is a vector composed of parameters of the label area, which are:

t _x ＝(xx _a )/w _a t _h ＝(yy _a )/h _a

t _w =log(w/w _a ) t _h =log(h/h _a )

{t t}_{x x}^{* *} = = (({x x}^{* *} - - {x x}_{a a})) / / {w w}_{a a},, {t t}_{y the y}^{* *} = = (({y the y}^{* *} - - {y the y}_{a a})) / / {h h}_{a a}

{t t}_{w w}^{* *} = = log log (({w w}^{* *} / / {w w}_{a a})),, {t t}_{h h}^{* *} = = log log (({h h}^{* *} / / {h h}_{a a}))

Among them, x, y, w, and h represent the center coordinates, width, and length of the object, respectively, x, x _a , x ^* represent the prediction area, anchor area, and label area, respectively, and the regression loss R is the smoothing loss function

{smooth smooth}_{{L L}_{11}} ((x x)) = = \{\begin{matrix} 0.5 0.5 {x x}^{22} & i i f f | | x x | | < < 11 \\ | | x x | | - - 0.5 0.5 & o o t t h h e e r r w w i i s the s e e \end{matrix}

Indicates that only when the anchor region is a positive sample When , the regression loss is calculated, otherwise Not calculated, the normalization parameters N _cls and N _reg represent the length of the low-dimensional vector mapped from the feature vector and the number of anchor regions, respectively.

4. as described in any one of claim 1 to 3, the parallelization method of fuzzy area convolutional neural network under a kind of big data environment, it is characterized in that, at first carry out the standardization of logging data, original data is all converted into infinite Outlined index evaluation values, each index evaluation value is at the same quantitative level, and then comprehensive evaluation and analysis are carried out.

5. the parallelization method of fuzzy area convolutional neural network under a kind of big data environment as claimed in claim 4, it is characterized in that, carry out the normalization of logging data adopt following normalization method:

Sx=(x-M)/S, x∈{GR, AC, DEN, CNL, SDN,...}

Among them, x represents the data of each well log curve, Sx represents the normalized well log data, M is the mean value of the corresponding well log data, and S is the standard deviation of each well log data.