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

A kind of well-log facies recognition method based on fuzzy theory and neutral net

Technical field

The present invention relates to petroleum well logging technology field, particularly relate to big data logging field.

Background technology

Well logging information and deposition are reflection and the governing factors of formation rock physical property, and therefore well-log information is all the time By the important information source as basis in oil and gas reservoir sedimentology research, well logging phase is then well logging information and Reservoir Sedimentological Bridge between feature.For most Oil/gas Well, well-log information is only synthesis letter covering full well section stratum Breath source, therefore well-log facies recognition analyzes method always as a most important research in oil exploration and exploitation geological research Means.

But, well logging information has the feature of ambiguity, has multi-solution and the ambiguity of geological Significance.Therefore, log well The identification of phase and analysis must be set up dividing in the comprehensive degree of depth of a large amount of existing deposition characteristicses and log parameter relation (log response) On analysis basis, the result of referring also to outcrop, core log and earthquake analysis simultaneously, choose building of applicable geology characteristic Mould method, could realize accurately identifying of well logging phase.

Further, since lack effective well logging phase automatic identifying method and technology, current well-log facies recognition is mainly logical Cross the manual identified realization of geological work personnel, and due to personnel's experience difference, subjective differences, the System level gray correlation of log data The factor such as different, the data volume faced by geological personnel is big, workload is heavy.Moreover, geological personnel experience difference, subjective because of The factors such as the systematical difference of the different instrument log data of element, different times so that traditional big discounting of well-log facies recognition accuracy Button.

It is that solution current oil industry is big that the advanced technologies such as big data analysis, degree of depth study are applied to oil-gas geology research The exploration of data analysis resources idle and trial.In recent years, petroleum industry establishes substantial amounts of cloud data center, but utilization rate is not Height, resource is by serious waste.One of them major reason is just a lack of big data processing platform (DPP) and corresponding big data technique Make full use of these to calculate, store resource.

Set up efficiently, accurately well-log facies recognition method be the active demand of present oil-gas geology research.

Content of the invention

For solving the deficiencies in the prior art, the present invention proposes a kind of well logging acquaintance based on fuzzy theory and neutral net Other method.

The technical scheme is that and be achieved in that：

A kind of well-log facies recognition method based on fuzzy theory and neutral net, first, builds fuzzy region convolutional Neural Network, will provide goal hypothesis region and target identification is put in same network, share convolutional calculation, and a training process is more The weight of new whole network；

It follows that log data carries out convolution and pondization operation, convolutional layer and pond through fuzzy region convolutional neural networks Change layer mutual, carry out fuzzy operation at convolutional layer and pond layer, from the beginning of the ground floor of fuzzy region convolutional neural networks, gradually Increase the number of plies of obfuscation, adjust the obfuscation number of plies for different data set, last of fuzzy region convolutional neural networks Layer obtains characteristic vector, this feature vector by a sliding window by the low dimensional vector of Feature Mapping to, then by spy Levying and being input to two full articulamentums, a full articulamentum is used for positioning, and another full articulamentum is used for classifying.

Alternatively, described convolutional layer formula is expressed as：

Pond layer formula is expressed as：

Wherein, biasAnd weightIt is fuzzy number, used here as symmetrical triangular fuzzy numbers,For mould Stick with paste the vector that array becomes, j-th fuzzy numberMembership function be：

Alternatively, during the training of fuzzy region convolutional neural networks, an associated losses function is defined：

Wherein, p_iIt is the prediction probability that this sample is form of logs,It is the label of sample, if surveying accordingly Well tracing pattern,It is 1, otherwiseIt is 0, N_clsIt is two sorted logic losses；t_iIt is four parameter compositions of prediction object boundary Vector,For the vector of tab area parameter composition, they 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 centre coordinate of object, width and length, x, x respectively_a, x^*Represent Target area respectively Territory, anchor region and tab area, return lossR is smooth loss function

Represent only when anchor region is positive sampleWhen, just calculate and return loss, otherwiseDo not calculate, Normalized parameter N_clsAnd N_regRepresent the length of low dimensional vector from maps feature vectors and the quantity of anchor region respectively.

Alternatively, first carry out the standardization of log data, initial data be converted into index without dimension test and appraisal value, Each index test and appraisal value is all in same number of levels, then carries out comprehensive test analysis.

Alternatively, the standardization carrying out log data uses following normalization method：

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

Wherein, x represents the data of every log, Sx represent standardization after borehole log data, M is corresponding well logging The average of curve data, S is the standard deviation of every borehole log data.

The invention has the beneficial effects as follows：

(1) according to the feature that data in the big data of well logging are fuzzy, incorporate fuzzy theory, propose fuzzy region convolutional Neural net Network FR-CNN, and progressive fuzzy method is proposed, it from the beginning of the ground floor of fuzzy region convolutional neural networks, is gradually increased fuzzy The number of plies changed, thus optimize network structure and parameter, it is achieved more preferable analytical performance and precision；

(2) adjust the number of plies of FR-CNN obfuscation for different log data collection, make the feature of extraction preferably reflect The characteristic of oil and gas reservoir itself, can solve log data fuzzy problem.

Brief description

In order to be illustrated more clearly that the embodiment of the present invention or technical scheme of the prior art, below will be to embodiment or existing In having technology to describe, the accompanying drawing of required use is briefly described, it should be apparent that, the accompanying drawing in describing below is only this Some embodiments of invention, for those of ordinary skill in the art, on the premise of not paying creative work, all right Obtain other accompanying drawing according to these accompanying drawings.

Fig. 1 is the structural representation of fuzzy region convolutional neural networks of the present invention；

Fig. 2 is symmetrical triangular fuzzy numbers coordinate schematic diagram.

Detailed description of the invention

Below in conjunction with the accompanying drawing in the embodiment of the present invention, the technical scheme in the embodiment of the present invention is carried out clear, complete Describe, it is clear that described embodiment is only a part of embodiment of the present invention, rather than whole embodiments wholely.Based on Embodiment in the present invention, it is every other that those of ordinary skill in the art are obtained under the premise of not making creative work Embodiment, broadly falls into the scope of protection of the invention.

Log data has the feature of ambiguity, the reason that cause this ambiguity is many, including noise, differ The data space of the log data that cause property, imperfection etc. cause pollutes, and also includes that different times, different instrument well logging bring Systematic data difference, the log data ambiguity that these problems are brought all constrains accurately identifying of well logging phase.

The present invention proposes a kind of well-log facies recognition method based on fuzzy theory and neutral net, builds log data Go out the data space of various dimensions, fuzzy theory is merged with degree of depth learning network R-CNN, survey in the case of proposing to solve fuzzy data The recognition methods of well phase, according to the feature that data in the big data of well logging are fuzzy, incorporates fuzzy theory, proposes fuzzy region convolution god Through network FR-CNN (Fuzzy R-CNN), it is further proposed that progressive blur method, from the beginning of convolutional neural networks ground floor, by The cumulative number of plies adding obfuscation, optimizes network structure and parameter, finally sets up theory and the method for FR-CNN, it is achieved preferably divide Analysis performance and precision.

Design the emphasis that suitable fuzzy region convolutional neural networks is the present invention, below to fuzzy region convolution of the present invention The structure of neutral net is described in detail.

Fuzzy region convolutional neural networks FR-CNN builds on the basis of degree of depth learning network R-CNN, such as Fig. 1 institute Showing, FR-CNN will provide goal hypothesis region and target identification is put in same network, share convolutional calculation, it is to avoid complicated Calculation procedure, it is only necessary to a training process just can update the weight of whole network, also accelerates detection speed simultaneously, reaches fast The purpose that speed is processed.

In Fig. 1, log data, through fuzzy region convolutional neural networks, carries out convolution and pondization operation.Fuzzy region is rolled up The core of long-pending neural metwork training is the mutual of convolutional layer and pond layer, therefore carries out obscuring at convolutional layer and pond layer and grasps Make.In order to avoid the fuzzy information loss excessively causing is too much, and extract feature in view of fuzzy region convolutional neural networks The degree that becomes more meticulous successively reduce, here change the obfuscation to each layer for the traditional fuzzy neutral net, the present invention proposes progressive Fuzzy method, i.e. from the beginning of the ground floor of fuzzy region convolutional neural networks, is gradually increased the number of plies of obfuscation, for difference Data set adjust the obfuscation number of plies, make the feature of extraction preferably reflect the characteristic of log, thus obtain best identified As a result, and improve recognition efficiency.

Last layer of fuzzy region convolutional neural networks obtains characteristic vector, and this feature vector passes through a little slip Window, by the low dimensional vector of Feature Mapping to, then inputs the feature into two full articulamentums, and a full articulamentum is used for Positioning, another full articulamentum is used for classifying.Provide several goal hypothesis region simultaneously at each sliding window, can be referred to as For anchor region, this region, centered on sliding window, has different transverse and longitudinal ratios and scaling.

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

Wherein,Represent be i-th layer of neuron j-th characteristic vector (x, y) value of position,Table ShowBe connected to the convolution kernel of m-th characteristic vector position (p, q) on weights.P_iAnd Q_iRepresent the height of convolution kernel respectively Degree and width, b_ijFor bias term, f (x) represents the activation primitive of neuron.

R-CNN pond layer formula is expressed as：

x_ij=f (β_ijdown(x_i-1j)+b_ij) (2)

Down (.) represents a down-sampling function, and typical operation is usually all of the different n*n blocks to input data Information is sued for peace, and so output data all reduce n times in two dimensions, and corresponding one of each output map belongs to certainly Oneself property taken advantage of a biasing β and additivity biasing b.

The input of convolutional neural networks and calculating process are all real numbers, the result obtaining all determining that property, and for number According to the situation that the data such as disappearance are fuzzy, the fuzzy region convolutional neural networks of the present invention introduces fuzzy theory, improved formula As follows：

Convolutional layer formula is expressed as：

Pond layer formula is expressed as：

Wherein biasAnd weightIt is fuzzy number, used here as symmetrical triangular fuzzy numbers,For mould Stick with paste the vector that array becomes, j-th fuzzy numberMembership function be

As in figure 2 it is shown, w_jIt is the symmetrical centre of fuzzy number,It is half length of fuzzy number,Represent the degree of membership at w.

During the training of fuzzy region convolutional neural networks, define an associated losses function：

Wherein p_iIt is the prediction probability that this sample is form of logs,It is the label of sample, if surveying accordingly Well tracing pattern,It is 1, otherwiseIt is 0, N_clsIt is the loss of two classification (0 or 1) logic.

t_iIt is the vector of four parameter compositions of prediction object boundary,For the vector of tab area parameter composition, they divide It is not：

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 centre coordinate of object, width and length, x, x respectively_a, x^*Represent estimation range respectively, Anchor region and tab area (y, w, h are in like manner).Return lossR is smooth loss function

Represent only when anchor region is positive sampleJust calculate and return loss, otherwiseDisregard Calculate.Normalized parameter N_clsAnd N_regRepresent the length of low dimensional vector from maps feature vectors and the quantity of anchor region respectively.

Use different well logging means can produce different data.As used natural gamma (GR), compensation sound wave (AC), mending Repay density (DEN), compensated neutron (CNL) and neutrovision porosity poor from density apparent porosity (SDN) etc. and there is different dimensions, Not having comparativity between data, therefore, the present invention needs first to carry out the standardization of log data, all changes initial data For index without dimension test and appraisal value, i.e. each index test and appraisal value is all in same number of levels, then carries out comprehensive test analysis.

Use following normalization method：

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

Wherein, x represents the data of every log, Sx represent standardization after borehole log data；M is corresponding well logging The average of curve data, S is the standard deviation of every borehole log data.

The present invention demarcates from existing log data, sets up the training dataset of FR-CNN, on this basis, by It is not quite similar in the information disclosed in different logging methods, so selecting defeated as FR-CNN of the combination of different log data Enter, so that it is determined that the log data combination that FR-CNN is optimum, and optimize network parameter when carrying out well-log facies recognition of FR-CNN And structure.

The feature that the present invention obscures according to data in the big data of well logging, incorporates fuzzy theory, proposes fuzzy region convolution god Through network FR-CNN, and propose progressive fuzzy method, from the beginning of the ground floor of fuzzy region convolutional neural networks, be gradually increased The number of plies of obfuscation, thus optimize network structure and parameter, it is achieved more preferable analytical performance and precision；And, the present invention is directed to Different log data collection adjusts the number of plies of FR-CNN obfuscation, makes the feature of extraction preferably reflect the spy of oil and gas reservoir itself Property, log data fuzzy problem can be solved.

The foregoing is only presently preferred embodiments of the present invention, not in order to limit the present invention, all essences in the present invention Within god and principle, any modification, equivalent substitution and improvement etc. made, should be included within the scope of the present invention.

Claims (5)

1. the well-log facies recognition method based on fuzzy theory and neutral net, it is characterised in that

First, build fuzzy region convolutional neural networks, goal hypothesis region will be given and same network is put in target identification In, share convolutional calculation, a training process updates the weight of whole network；

It follows that log data carries out convolution and pondization operation, convolutional layer and pond layer through fuzzy region convolutional neural networks Alternately, carry out fuzzy operation at convolutional layer and pond layer, from the beginning of the ground floor of fuzzy region convolutional neural networks, be gradually increased The number of plies of obfuscation, adjusts the obfuscation number of plies for different data sets, and last layer of fuzzy region convolutional neural networks obtains To characteristic vector, this feature vector by a sliding window by the low dimensional vector of Feature Mapping to, then that feature is defeated Entering to two full articulamentums, a full articulamentum is used for positioning, and another full articulamentum is used for classifying.

2. a kind of well-log facies recognition method based on fuzzy theory and neutral net as claimed in claim 1, it is characterised in that Described convolutional layer formula is expressed as：

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

Pond layer formula is expressed as：

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

Wherein, biasAnd weightIt is fuzzy number, used here as symmetrical triangular fuzzy numbers,It is fuzzy The vector that array becomes, j-th fuzzy numberMembership function be：

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

3. a kind of well-log facies recognition method based on fuzzy theory and neutral net as claimed in claim 2, it is characterised in that During the training of fuzzy region convolutional neural networks, define an associated losses function：

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

Wherein, p_iIt is the prediction probability that this sample is form of logs,It is the label of sample, if corresponding well logging song Line morphology,It is 1, otherwiseIt is 0, N_clsIt is two sorted logic losses；t_iBe prediction object boundary four parameters composition to Amount,For the vector of tab area parameter composition, they 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)

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

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

Wherein, x, y, w and h represent the centre coordinate of object, width and length, x, x respectively_a, x^*Represent estimation range, anchor respectively Determine region and tab area, return lossR is smooth loss function

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

Represent only when anchor region is positive sampleWhen, just calculate and return loss, otherwiseDo not calculate, normalizing Change parameter N_clsAnd N_regRepresent the length of low dimensional vector from maps feature vectors and the quantity of anchor region respectively.

4. a kind of well-log facies recognition method based on fuzzy theory and neutral net as described in any one of claims 1 to 3, its It is characterised by, first carries out the standardization of log data, initial data is converted into index without dimension test and appraisal value, each index Test and appraisal value is all in same number of levels, then carries out comprehensive test analysis.

5. a kind of well-log facies recognition method based on fuzzy theory and neutral net as claimed in claim 4, it is characterised in that The standardization carrying out log data uses following normalization method：

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

Wherein, x represents the data of every log, Sx represent standardization after borehole log data, M is corresponding log The average of data, S is the standard deviation of every borehole log data.