CN106446514A - Fuzzy theory and neural network-based well-log facies recognition method - Google Patents

Abstract

The invention provides a fuzzy theory and neural network-based well-log facies recognition method. The method comprises the steps of firstly, constructing a fuzzy area convolutional neural network, putting a given target hypothesis area and target recognition into the same network, sharing convolution calculation, and updating a weight of the whole network by a training process; and secondly, performing convolution and pooling operations on well-log data through the fuzzy area convolutional neural network, interacting a convolution layer with a pooling layer, performing fuzzy operation at the convolution layer and the pooling layer, gradually increasing a fuzzy layer number from a first layer of the fuzzy area convolutional neural network, adjusting the fuzzy layer number for different data sets, obtaining an eigenvector from a last layer of the fuzzy area convolutional neural network, mapping a feature of the eigenvector to a low-dimensional vector through a sliding window, and inputting the feature to two full connection layers, wherein one full connection layer is used for locating, and the other full connection layer is used for classification.

Description

Well logging phase identification method based on fuzzy theory and neural network

Technical Field

The invention relates to the technical field of petroleum logging, in particular to the field of big data logging.

Background

The logging information and the deposition are reflection and control factors of formation rock physical properties, so the logging information is always used as an important information source for the research of oil and gas reservoir sedimentology, and the logging phase is a bridge between the logging information and reservoir sedimentology characteristics. For most oil and gas wells, the logging information is the only comprehensive information source covering the whole well interval stratum, so the logging phase identification analysis method is always used as the most important research means in oil and gas exploration and geological research development.

However, the logging information has the characteristic of ambiguity, and has geologically significant multi-solution and ambiguity. Therefore, the identification and analysis of the logging facies must be established on the basis of a large amount of existing comprehensive depth analysis of sedimentary feature and logging parameter relations (logging response), and meanwhile, the accurate identification of the logging facies can be realized only by selecting a modeling method suitable for geological features by referring to results of field outcrop, core logging and seismic analysis.

In addition, due to the lack of effective automatic identification methods and technologies for logging facies, the current logging facies identification is mainly realized through manual identification of geological workers, and due to factors such as personnel experience difference, subjective difference and system difference of logging data, the data volume faced by the geological workers is large and the workload is heavy. Moreover, the traditional logging phase identification accuracy is greatly reduced due to factors such as experience difference of geologists, subjective factors, system difference of logging data of different instruments in different periods and the like.

The application of advanced technologies such as big data analysis and deep learning to oil and gas geological research is exploration and attempt for solving the problem that big data analysis resources in the current petroleum industry are idle. In recent years, a large number of cloud data centers are established in the petroleum industry, but the utilization rate is not high, and resources are seriously wasted. One important reason for this is the lack of large data processing platforms and corresponding large data technologies to fully utilize these computing and storage resources.

The establishment of an efficient and accurate well logging phase identification method is an urgent need of the current oil and gas geological research.

Disclosure of Invention

In order to overcome the defects of the prior art, the invention provides a log facies identification method based on a fuzzy theory and a neural network.

The technical scheme of the invention is realized as follows:

a well logging phase identification method based on fuzzy theory and neural network, firstly, constructing a fuzzy area convolution neural network, putting a given target hypothesis area and target identification into the same network, sharing convolution calculation, and updating the weight of the whole network in a training process;

then, the logging data is subjected to convolution and pooling operation through a fuzzy region convolution neural network, the convolution layer and the pooling layer are interacted, the fuzzy operation is carried out on the convolution layer and the pooling layer, the number of the fuzzification layers is gradually increased from the first layer of the fuzzy region convolution neural network, the number of the fuzzification layers is adjusted according to different data sets, the last layer of the fuzzy region convolution neural network obtains a feature vector, the feature vector maps the features to a low-dimensional vector through a sliding window, then the features are input into two full-connection layers, one full-connection layer is used for positioning, and the other full-connection layer is used for classification.

Optionally, the convolutional layer is formulated as:

the pooling layer formula is expressed as:

wherein, is offsetAnd weightAre fuzzy numbers, here symmetric triangular fuzzy numbers are used,is a vector of fuzzy numbers, the jth fuzzy numberThe membership function of (a) is:

optionally, in the training process of the fuzzy area convolutional neural network, a joint loss function is defined:

wherein p is_iIs the predicted probability that this sample is the log profile,is a label for the sample, and, if the corresponding log configuration,is 1, otherwiseIs 0, N_clsIs a two-classification logic loss; t is t_iIs a vector of four parameters that predicts the object boundaries,for marking the direction of the composition of the region parametersAmounts, which are respectively:

t_x＝(x-x_a)/w_at_h＝(y-y_a)/h_a

t_w＝log(w/w_a) t_h＝log(h/h_a)

wherein x, y, w and h represent the center coordinate, width and length of the object, respectively, and x, x_a，x^*Respectively representing prediction region, anchor region and label region, regression lossR is a smoothing loss function

Indicating that only if the anchor region is a positive sampleThen, the regression loss is calculated, otherwiseNot calculating, normalizing parameter N_clsAnd N_regRespectively representing low-dimensional directions mapped from feature vectorsThe length of the volume and the number of anchoring regions.

Optionally, the well logging data is normalized, the original data is converted into non-dimensionalized mapping evaluation values, and the mapping evaluation values are in the same quantity level, and then comprehensive evaluation analysis is performed.

Optionally, the normalization of the well log data is performed by the following normalization method:

Sx＝(x-M)/S,x∈{GR，AC，DEN，CNL，SDN，...}

wherein x represents the data of each logging curve, Sx represents the normalized logging curve data, M is the mean value of the corresponding logging curve data, and S is the standard deviation of each logging curve data.

The invention has the beneficial effects that:

(1) according to the characteristic of data fuzziness in big logging data, a fuzzy theory is integrated, a fuzzy area convolutional neural network FR-CNN is provided, a progressive fuzzy method is provided, the number of layers of fuzziness is gradually increased from the first layer of the fuzzy area convolutional neural network, so that the network structure and parameters are optimized, and better analysis performance and precision are realized;

(2) the number of layers of FR-CNN fuzzification is adjusted according to different logging data sets, so that the extracted features better reflect the characteristics of the oil and gas reservoir, and the problem of the fuzzification of the logging data can be solved.

Drawings

In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly described below, it is obvious that the drawings in the following description are only some embodiments of the present invention, and for those skilled in the art, other drawings can be obtained according to the drawings without creative efforts.

FIG. 1 is a schematic diagram of a fuzzy area convolutional neural network according to the present invention;

FIG. 2 is a diagram of symmetric triangular fuzzy number coordinates.

Detailed Description

The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.

The well logging data has the characteristic of ambiguity, the reasons for the ambiguity are manifold, including data space pollution of the well logging data caused by noise, inconsistency, imperfection and the like, and systematic data differences caused by well logging in different periods and different instruments, and the ambiguity of the well logging data caused by the problems restricts accurate identification of a well logging phase.

The invention provides a well logging phase identification method based on a Fuzzy theory and a neural network, which constructs a multidimensional data space for well logging data, fuses the Fuzzy theory and a deep learning network R-CNN, provides an identification method for well logging phases under the condition of solving the Fuzzy data, fuses the Fuzzy theory according to the characteristic of data fuzziness in well logging big data, provides a Fuzzy area convolution neural network FR-CNN (Fuzzy R-CNN), further provides a progressive Fuzzy method, gradually increases the number of layers of fuzziness from the first layer of the convolution neural network, optimizes the network structure and parameters, finally establishes the theory and the method of FR-CNN, and realizes better analysis performance and precision.

The design of a proper fuzzy area convolutional neural network is the key point of the invention, and the construction of the fuzzy area convolutional neural network of the invention is explained in detail below.

The fuzzy area convolution neural network FR-CNN is established on the basis of the deep learning network R-CNN, as shown in figure 1, the FR-CNN puts a given target hypothesis area and target identification into the same network, shares convolution calculation, avoids complex calculation steps, can update the weight of the whole network only by one training process, and simultaneously accelerates the detection speed and achieves the purpose of rapid processing.

In FIG. 1, the well log data is passed through a fuzzy area convolutional neural network for convolution and pooling operations. The core of fuzzy area convolutional neural network training is the interaction of the convolutional layer and the pooling layer, so that fuzzy operation is performed on the convolutional layer and the pooling layer. In order to avoid excessive information loss caused by excessive blurring and consider that the refinement degree of the feature extracted by the fuzzy area convolutional neural network is reduced layer by layer, the blurring of each layer by the traditional fuzzy neural network is changed, the invention provides a progressive blurring method, namely, the number of the blurring layers is gradually increased from the first layer of the fuzzy area convolutional neural network, and the number of the blurring layers is adjusted according to different data sets, so that the extracted feature better reflects the characteristic of a well logging curve, the optimal recognition result is obtained, and the recognition efficiency is improved.

The last layer of the fuzzy area convolutional neural network gets the feature vector, which maps the feature to a low-dimensional vector through a small sliding window, and then inputs the feature to two fully-connected layers, one for localization and the other for classification. Several target hypothesis regions, which may be referred to as anchor regions, are simultaneously given at each sliding window, and the regions are centered on the sliding window and have different aspect ratios and scales.

The convolutional layer formula of the convolutional neural network R-CNN can be expressed as:

wherein,representing the value at the (x, y) position of the jth feature vector for the ith layer neuron,to representThe weight of the convolution kernel connected to the mth eigenvector at position (p, q). P_iAnd Q_iRespectively representing the height and width of the convolution kernel, b_ijFor the bias term, f (x) represents the activation function of the neuron.

The R-CNN pooling layer formula is expressed as:

x_ij＝f(β_ijdown(x_i-1j)+b_ij) (2)

down (.) represents a down-sampling function, and 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 corresponds to a multiplicative offset β and an additive offset b belonging to itself.

The input and calculation processes of the convolutional neural network are real numbers, the obtained results are deterministic, and for the condition of data blurring such as data missing, the fuzzy area convolutional neural network of the invention introduces a fuzzy theory, and the improved formula is as follows:

the convolutional layer formula is expressed as:

the pooling layer formula is expressed as:

wherein is offsetAnd weightAre fuzzy numbers, here symmetric triangular fuzzy numbers are used,is a vector of fuzzy numbers, the jth fuzzy numberIs a membership function of

As shown in FIG. 2, w_jIs the center of symmetry of the blur number,is half the length of the blur number,representing the degree of membership at w.

In the training process of the fuzzy area convolution neural network, a joint loss function is defined:

wherein p is_iIs the predicted probability that this sample is the log profile,is a label for the sample, and, if the corresponding log configuration,is 1, otherwiseIs 0, N_clsIs a two classification (0 or 1) logical loss.

t_iIs a vector of four parameters that predicts the object boundaries,the vectors formed by the parameters of the marked areas are respectively as follows:

t_x＝(x-x_a)/w_at_h＝(y-y_a)/h_a(7)

t_w＝log(w/w_a)t_h＝log(h/h_a)

wherein x, y, w and h represent the center coordinate, width and length of the object, respectively, and x, x_a，x^*Respectively representing a prediction region, an anchor region and a labeling region (y, w, h are the same). Loss of returnR is a smoothing loss function

Meaning that only when the anchor region is a positive sampleThe regression loss is calculated, otherwiseNo calculation is performed. Normalization parameter N_clsAnd N_regRepresenting the length of the low-dimensional vector mapped from the feature vector and the number of anchor regions, respectively.

Different logging tools are used to generate different data. If natural Gamma (GR), compensated acoustic wave (AC), compensated Density (DEN), Compensated Neutron (CNL), neutron apparent porosity and density apparent porosity difference (SDN) and the like are adopted and have different dimensions, data have no comparability, therefore, the invention needs to firstly carry out standardization of logging data, convert all original data into non-dimensionalized index mapping evaluation values, namely, all index mapping evaluation values are on the same quantity level, and then carry out comprehensive evaluation analysis.

The following normalization method was used:

Sx＝(x-M)/S,x∈{GR，AC，DEN，CNL，SDN，...}

wherein x represents the data of each logging curve, and Sx represents the normalized logging curve data; m is the mean of the corresponding log data, and S is the standard deviation of each log data.

The invention carries out calibration from the existing logging data, establishes a training data set of FR-CNN, and on the basis, because the information disclosed by different logging methods is different, selects the combination of different logging data as the input of FR-CNN, thereby determining the optimal logging data combination of FR-CNN and optimizing the network parameters and structure of FR-CNN when carrying out logging phase identification.

According to the characteristics of data fuzziness in big logging data, a fuzzy theory is integrated, a fuzzy area convolutional neural network FR-CNN is provided, a progressive fuzzy method is provided, the number of layers of fuzziness is gradually increased from the first layer of the fuzzy area convolutional neural network, so that the network structure and parameters are optimized, and better analysis performance and precision are realized; 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 above description is only for the purpose of illustrating the preferred embodiments of the present invention and is not to be construed as limiting the invention, and any modifications, equivalents, improvements and the like that fall within the spirit and principle of the present invention are intended to be included therein.

Claims (5)

1. A well logging phase identification method based on fuzzy theory and neural network is characterized in that,

firstly, constructing a fuzzy area convolution neural network, putting a given target hypothesis area and target identification into the same network, sharing convolution calculation, and updating the weight of the whole network in a training process;

then, the logging data is subjected to convolution and pooling operation through a fuzzy region convolution neural network, the convolution layer and the pooling layer are interacted, the fuzzy operation is carried out on the convolution layer and the pooling layer, the number of the fuzzification layers is gradually increased from the first layer of the fuzzy region convolution neural network, the number of the fuzzification layers is adjusted according to different data sets, the last layer of the fuzzy region convolution neural network obtains a feature vector, the feature vector maps the features to a low-dimensional vector through a sliding window, then the features are input into two full-connection layers, one full-connection layer is used for positioning, and the other full-connection layer is used for classification.

2. The method for identifying well-log facies based on fuzzy theory and neural network as claimed in claim 1, wherein said convolutional layer formula is expressed as:

$v_{ij}^{xy}=f\left({\stackrel{\‾}{b}}_{ij}+{\mathrm{\Σ}}_m{\mathrm{\Σ}}_{p=0}^{P_i-1}{\mathrm{\Σ}}_{q=0}^{Q_i-1}{\stackrel{\‾}{W}}_{ijm}^{pq}{v}_{\left(i-1\right)m}^{\left(x+p\right)\left(y+q\right)}\right)$

the pooling layer formula is expressed as:

$x_j=f\left({\stackrel{\‾}{\mathrm{\β}}}_{ij}down\left(x_{i-1j}\right)+{\stackrel{\‾}{b}}_{ij}\right)$

wherein, is offsetAnd weightAre fuzzy numbers, here symmetric triangular fuzzy numbers are used,is a vector of fuzzy numbers, the jth fuzzy numberThe membership function of (a) is:

${\stackrel{\‾}{W}}_j\left(w\right)=\max \left\{1-\frac{|w-w_j|}{{\hat{w}}_j},0\right\}.$

3. the method as claimed in claim 2, wherein during the training of the fuzzy area convolutional neural network, a joint loss function is defined:

$L\left(\left\{p_i\right\},\left\{t_i\right\}\right)=\frac{1}{N_{cls}}{\mathrm{\Σ}}_i{L}_{cls}\left(p_i,p_i^{*}\right)+\mathrm{\λ}\frac{1}{N_{reg}}{\mathrm{\Σ}}_i{p}_i^{*}{L}_{reg}\left(t_i,t_i^{*}\right)$

wherein p is_iIs the prediction of the shape of the logging curveThe probability of the occurrence of the event,is a label for the sample, and, if the corresponding log configuration,is 1, otherwiseIs 0, N_clsIs a two-classification logic loss; t is t_iIs a vector of four parameters that predicts the object boundaries,the vectors formed by the parameters of the marked areas are respectively as follows:

t_x＝(X-X_a)/w_at_h＝(y-y_a)/h_a

t_w＝log(w/w_a)t_h＝log(h/h_a)

$\begin{array}{cc}{t}_x^{*}=\left(x^{*}-x_a\right)/w_a& t_y^{*}=\left(y^{*}-y_a\right)/h_a\end{array}$

$\begin{array}{cc}{t}_w^{*}=\log \left(w^{*}/w_a\right)& t_h^{*}=\log \left(h^{*}/h_a\right)\end{array}$

wherein x, y, w and h represent the middle of the object, respectivelyHeart coordinate, width and length, x_a，x^*Respectively representing prediction region, anchor region and label region, regression lossR is a smoothing loss function

${\mathrm{smooth}}_{L_1}\left(x\right)=\left\{\begin{array}{cc}0.5x^2& if\left|x\right|<1\\ \left|x\right|-0.5& otherwise\end{array}\right.$

Indicating that only the anchoring area is presentIs a positive sampleThen, the regression loss is calculated, otherwiseNot calculating, normalizing parameter N_clsAnd N_regRepresenting the length of the low-dimensional vector mapped from the feature vector and the number of anchor regions, respectively.

4. The method for identifying well-logging phase based on fuzzy theory and neural network as claimed in any one of claims 1 to 3, wherein the normalization of the well-logging data is performed first, the raw data are all converted into dimensionless index values, and the index values are all in the same quantitative level, and then the comprehensive evaluation analysis is performed.

5. The method for identifying the log facies based on the fuzzy theory and the neural network as claimed in claim 4, wherein the normalization of the log data is performed by the following normalization method:

Sx＝(x-M)/S,x∈{GR，AC，DEN，CNL，SDN，...}

wherein x represents the data of each logging curve, Sx represents the normalized logging curve data, M is the mean value of the corresponding logging curve data, and S is the standard deviation of each logging curve data.