CN118171582B - Electromagnetic wave logging inversion method and system for azimuth while drilling by combining residual neural network and L-M algorithm - Google Patents
Electromagnetic wave logging inversion method and system for azimuth while drilling by combining residual neural network and L-M algorithm Download PDFInfo
- Publication number
- CN118171582B CN118171582B CN202410581857.0A CN202410581857A CN118171582B CN 118171582 B CN118171582 B CN 118171582B CN 202410581857 A CN202410581857 A CN 202410581857A CN 118171582 B CN118171582 B CN 118171582B
- Authority
- CN
- China
- Prior art keywords
- model
- neural network
- drilling
- stratum
- residual neural
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
- 238000013528 artificial neural network Methods 0.000 title claims abstract description 115
- 238000005553 drilling Methods 0.000 title claims abstract description 92
- 238000004422 calculation algorithm Methods 0.000 title claims abstract description 58
- 238000000034 method Methods 0.000 title claims abstract description 50
- 238000012549 training Methods 0.000 claims abstract description 42
- 230000004044 response Effects 0.000 claims abstract description 38
- 230000015572 biosynthetic process Effects 0.000 claims abstract description 35
- 230000008859 change Effects 0.000 claims abstract description 24
- 238000012795 verification Methods 0.000 claims abstract description 23
- 238000012360 testing method Methods 0.000 claims abstract description 18
- 208000037170 Delayed Emergence from Anesthesia Diseases 0.000 claims description 16
- 230000000694 effects Effects 0.000 claims description 15
- 230000006870 function Effects 0.000 claims description 15
- 238000011156 evaluation Methods 0.000 claims description 10
- 238000010276 construction Methods 0.000 claims description 9
- 238000005070 sampling Methods 0.000 claims description 9
- 230000004913 activation Effects 0.000 claims description 7
- 238000007781 pre-processing Methods 0.000 claims description 7
- 230000002452 interceptive effect Effects 0.000 claims description 6
- 238000003708 edge detection Methods 0.000 claims description 5
- 210000002569 neuron Anatomy 0.000 claims description 5
- 238000004088 simulation Methods 0.000 claims description 4
- 230000008569 process Effects 0.000 description 11
- 238000010586 diagram Methods 0.000 description 9
- 238000004364 calculation method Methods 0.000 description 4
- 238000010606 normalization Methods 0.000 description 4
- 238000000605 extraction Methods 0.000 description 3
- 239000011159 matrix material Substances 0.000 description 3
- 238000005259 measurement Methods 0.000 description 3
- 238000002790 cross-validation Methods 0.000 description 2
- 238000013135 deep learning Methods 0.000 description 2
- 238000001514 detection method Methods 0.000 description 2
- 238000002474 experimental method Methods 0.000 description 2
- 230000007935 neutral effect Effects 0.000 description 2
- 230000009467 reduction Effects 0.000 description 2
- ORILYTVJVMAKLC-UHFFFAOYSA-N Adamantane Natural products C1C(C2)CC3CC1CC2C3 ORILYTVJVMAKLC-UHFFFAOYSA-N 0.000 description 1
- 230000002159 abnormal effect Effects 0.000 description 1
- 238000004458 analytical method Methods 0.000 description 1
- 230000015556 catabolic process Effects 0.000 description 1
- 238000012512 characterization method Methods 0.000 description 1
- 238000013527 convolutional neural network Methods 0.000 description 1
- 238000012937 correction Methods 0.000 description 1
- 238000006731 degradation reaction Methods 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000005457 optimization Methods 0.000 description 1
- 230000008092 positive effect Effects 0.000 description 1
- 238000012545 processing Methods 0.000 description 1
- 239000011435 rock Substances 0.000 description 1
- 230000000087 stabilizing effect Effects 0.000 description 1
- 230000000007 visual effect Effects 0.000 description 1
Classifications
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02A—TECHNOLOGIES FOR ADAPTATION TO CLIMATE CHANGE
- Y02A90/00—Technologies having an indirect contribution to adaptation to climate change
- Y02A90/30—Assessment of water resources
Landscapes
- Geophysics And Detection Of Objects (AREA)
Abstract
The invention belongs to the technical field of geophysical well logging, and relates to a method and a system for inversion of electromagnetic wave well logging while drilling of a combined residual neural network and an L-M algorithm, wherein the method comprises the following steps: s1, constructing a stratum resistivity forward model; s2, forward modeling is carried out according to the set stratum change parameters to obtain an azimuth electromagnetic wave logging response while drilling; s3, dividing a training set and a verification set, and constructing a test set; s4, constructing and training a residual neural network to obtain a residual neural network prediction model; s5, inputting the verification set into a residual neural network prediction model to obtain a prediction result, comparing the prediction result with stratum change parameters in a test set, and calculating a prediction error; s6, inputting a prediction result into an L-M algorithm, and carrying out inversion iteration to obtain optimal stratum resistivity model parameters; s7, constructing an actual formation resistivity model, and executing the steps S1-S6 to obtain actual formation resistivity model parameters. The invention can realize the rapid and accurate inversion of the multi-interface model.
Description
Technical Field
The invention belongs to the technical field of geophysical well logging, relates to a logging-while-drilling azimuth electromagnetic wave well logging inversion technology, and particularly relates to a logging-while-drilling azimuth electromagnetic wave well logging inversion method and system combining a residual neural network and an L-M algorithm.
Background
The logging while drilling azimuth electromagnetic wave can obtain the stratum resistivity and stratum interface azimuth information at the same time, is an important means for offshore oil and gas field and complex oil and gas field exploration, and plays a vital role in while drilling geosteering. In the drilling process of a highly deviated well/horizontal well, rapid and accurate formation resistivity profile reconstruction is a basis for realizing accurate landing of a drill bit, however, as a measurement signal is not visual, the underground environment is complex and can be influenced by the anisotropy of the formation, the resistivity of upper and lower surrounding rocks and the layer thickness, and the problem of inversion multi-solution exists, so that the difficulty in processing and inversion of the azimuth electromagnetic wave information while drilling is increased. Therefore, the realization of the rapid and accurate inversion of the electromagnetic wave logging information of the azimuth while drilling is a key for realizing the detection of stratum interface, borehole correction and geosteering.
The conventional electromagnetic wave logging inversion method is based on mathematical inversion theory, needs a large amount of calculation resources, is easily influenced by data noise and model errors, and particularly the inversion process is sensitive to the selection of the initial values, and seriously influences the logging inversion result.
In recent years, by means of strong nonlinear characterization capability and feature extraction capability of deep learning neural network architecture, the problem of the calculation efficiency of the traditional inversion method (for example ：Li H, Liu G, Yang S, et al. Automated resistivity inversion and formation geometry determination in high-angle and horizontal wells using deep learning techniques. SPWLA 60th Annual Logging Symposium. The Woodlands, Texas, USA, June 15-19, 2019;Jin Y C, Wu X Q, Chen J F, et al. Using a physics-driven deep neural network to solve inverse problems for LWD azimuthal resistivity measurements. SPWLA 60th Annual Logging Symposium. The Woodlands, Texas, USA, June 15-19, 2019）, but the inversion accuracy is insufficient compared with that of the traditional gradient algorithm) is solved to a certain extent.
In summary, the existing electromagnetic wave inversion algorithm while drilling is difficult to meet the requirements of accuracy and real time, along with the continuous increase of domestic logging while drilling data, there is a need to develop a real-time and accurate electromagnetic wave logging while drilling inversion method and system to improve the inversion speed and accuracy of the resistivity profile of the electromagnetic wave logging while drilling stratum, and provide accurate and reliable stratum interface, adjacent stratum resistivity and geological structure information for real-time geosteering.
Disclosure of Invention
The invention aims at the technical problems of the prior art, and provides the method and the system for the electromagnetic wave logging inversion of the azimuth while drilling by combining the residual neural network and the L-M algorithm, which can improve the electromagnetic wave logging inversion speed of the azimuth while drilling on the basis of ensuring inversion precision, provide accurate and reliable stratum interface, adjacent layer resistivity and geological structure information for real-time geosteering, meet the need of real-time drilling geosteering, and achieve the goal of 'measurement while drilling'.
The invention provides a method for inverting electromagnetic wave logging while drilling of a position by combining a residual neutral network and an L-M algorithm, which comprises the following specific steps:
S1, constructing a stratum resistivity forward model, wherein the stratum resistivity forward model comprises a single-interface model, a double-interface model and a multi-interface model, and the change parameters of each model are the resistivity of each stratum and the edge detection distance;
S2, setting stratum change parameters, and obtaining electromagnetic wave logging response of the azimuth while drilling corresponding to the stratum resistivity forward model through forward modeling;
s3, dividing the logging response of the azimuth electromagnetic wave while drilling into a training set and a verification set according to a set proportion, and constructing a test set according to stratum change parameters set in the step S2;
S4, constructing a residual neural network by taking the logging response of the azimuth electromagnetic wave while drilling as input and the stratum parameter as output, and training the residual neural network by adopting a training set to obtain a residual neural network prediction model, wherein the stratum parameter report comprises stratum resistivity and stratum interface;
S5, inputting the result into a residual neural network prediction model through a verification set to obtain a prediction result, comparing the prediction result with stratum change parameters in a test set, and calculating a prediction error to evaluate the prediction effect and generalization capability of the residual neural network prediction model;
s6, inputting a prediction result as an initial value into an L-M algorithm, and inverting and iterating to obtain the optimal stratum resistivity model parameter;
S7, constructing an actual formation resistivity model according to the horizontal well interactive modeling result of the target zone target well, and executing the steps S1-S6 to obtain actual formation resistivity model parameters of the target zone target well target interval.
In some embodiments, after said step S2 and before step S3, a pre-processing step is further included: and normalizing the logging response of the electromagnetic wave of the azimuth while drilling to obtain the logging response of the electromagnetic wave of the azimuth while drilling.
In some embodiments, in the step S1, when the formation resistivity forward model is constructed, a plurality of points are taken at set sampling intervals within a formation resistivity setting range, a plurality of points are taken at set sampling intervals within a sounding distance setting range, different formation resistivities form different data pairs under the same model, and a plurality of different formation models are provided in the single-interface model, the double-interface model and the multi-interface model.
In some embodiments, in the step S1, the logging response of the electromagnetic wave while drilling of the forward model of the resistivity of all stratum is obtained by simulating the forward algorithm of the electromagnetic wave while drilling one-dimensional layered medium.
In some embodiments, in the step S4, the tuning parameters of the residual neural network prediction model are network parameters of the residual neural network, where the tuning parameters include a network layer number, a neuron number of each layer of the network, an activation function, a residual block number, an optimizer, a learning rate, and a regularization coefficient.
In some embodiments, in the step S4, the residual neural network adopts a ResNet network structure, and the residual network includes a plurality of residual blocks, where each residual block is composed of two basic blocks, each basic block includes two convolution layers and one jump connection, and a step size of a first convolution layer of each residual block is different from a step size of the remaining convolution layers.
The invention provides a logging inversion system of azimuth electromagnetic wave while drilling, which is used for realizing the logging inversion method of the azimuth electromagnetic wave while drilling of the combined residual neural network and the L-M algorithm, and comprises the following steps:
The forward model construction module is used for constructing a stratum resistivity forward model;
The forward modeling module is used for setting stratum change parameters and acquiring an electromagnetic wave logging response of the azimuth while drilling corresponding to the stratum resistivity forward modeling module through forward modeling;
the data set construction module divides the logging response of the azimuth electromagnetic wave while drilling into a training set and a verification set according to a set proportion, and constructs a test set according to stratum change parameters set by the forward modeling module;
the model construction module is used for constructing a residual neural network by taking the logging response of the azimuth electromagnetic wave while drilling as input and the stratum parameter as output, and training the residual neural network by adopting a training set to obtain a residual neural network prediction model;
the evaluation module is used for obtaining a prediction result by inputting the verification set into the residual neural network prediction model, comparing the prediction result with stratum change parameters in the test set, and calculating a prediction error so as to evaluate the prediction effect and generalization capability of the residual neural network prediction model;
the inversion module inputs the prediction result as an initial value into an L-M algorithm, and inversion iteration is carried out to obtain optimal stratum resistivity model parameters;
An application module configured to: constructing an actual stratum resistivity model according to the horizontal well interactive modeling result of the target well of the target zone; and controlling and executing functions of the forward model building module, the forward modeling module, the data set building module, the model building module, the evaluation module and the inversion module, and outputting actual formation resistivity model parameters of the target interval of the target well.
In some embodiments, the system further comprises a preprocessing module, configured to normalize the electromagnetic wave logging response while drilling obtained by the forward modeling module to obtain a normalized electromagnetic wave logging response while drilling.
Compared with the prior art, the invention has the advantages and positive effects that:
(1) According to the method and the system for carrying out the electromagnetic wave logging inversion along with the drilling of the joint residual neural network and the L-M algorithm, the residual neural network is combined with the gradient algorithm, the inversion speed is rapidly improved by utilizing the strong nonlinear characteristic extraction capability of the residual neural network, meanwhile, the output of the residual neural network is used as an iteration initial value of the L-M algorithm, the stable reduction of a solving process and rapid convergence near an extreme point are realized, and the purpose of improving inversion precision is achieved. According to the invention, on the premise of ensuring inversion accuracy, the inversion speed of the electromagnetic wave logging while drilling azimuth is improved.
(2) The method and the system for inverting the electromagnetic wave logging while drilling in the azimuth of the combined residual neural network and the L-M algorithm have self-adaptability and self-learning capability, can automatically adjust the model according to actual data, better fit the actual data, realize the rapid and accurate inversion of a multi-interface model, have important significance for stratum evaluation and determination of stratum structures, and are suitable for the engineering requirements of while-drilling geosteering for precision and real time.
Drawings
FIG. 1 is a flow chart of an inversion method of electromagnetic wave logging while drilling of a joint residual neural network and an L-M algorithm according to an embodiment of the invention;
FIG. 2 is a schematic diagram of an inversion process of a residual neural network prediction model according to an embodiment of the present invention;
Fig. 3-4 are schematic diagrams of training time of a residual neural network according to an embodiment of the present invention;
FIG. 5 is a block diagram of an inversion system for electromagnetic wave logging while drilling according to an embodiment of the present invention;
FIG. 6 is a schematic diagram of a four-interface resistivity model constructed in accordance with an embodiment of the present invention;
FIG. 7 is a schematic diagram of the inversion result of the resistivity of the residual neural network of the four-interface stratum model according to the embodiment of the invention;
FIG. 8 is a schematic diagram of a four-interface distance inversion result of a residual neural network of a four-interface stratum model according to an embodiment of the present invention;
FIG. 9 is a schematic diagram of resistivity inversion results of a residual neural network and L-M joint inversion of a four-interface formation model according to an embodiment of the present invention;
FIG. 10 is a schematic diagram of a four-interface distance inversion result of a residual neural network and L-M joint inversion of a four-interface formation model according to an embodiment of the present invention;
FIGS. 11-16 are schematic diagrams of four-boundary residual neural networks and L-M joint inversion results under different forward model parameter conditions according to an embodiment of the invention.
In the figure, a forward model building module 1, a forward simulation module 2, a data set building module 3, a model building module 4, a model building module 5, an evaluation module 6, an inversion module 7, an application module 8 and a preprocessing module.
Detailed Description
The invention will now be described in more detail by way of exemplary embodiments with reference to the accompanying drawings. It is to be understood that elements, structures, and features of one embodiment may be beneficially incorporated in other embodiments without further recitation.
The existing azimuth electromagnetic wave inversion method while drilling cannot meet the requirements of accuracy and instantaneity at the same time. In order to improve inversion speed and accuracy of formation resistivity profile of azimuth electromagnetic wave logging while drilling, the invention provides an azimuth electromagnetic wave logging while drilling inversion method and system combining a residual neural network and an L-M algorithm, wherein the residual neural network is combined with a gradient algorithm, the inversion speed is rapidly improved by utilizing strong nonlinear characteristic extraction capacity of the residual neural network, meanwhile, the output of the residual neural network is used as an iteration initial value of the L-M algorithm, stable reduction of a solving process and rapid convergence near an extreme point are realized, and the purpose of improving inversion accuracy is achieved. According to the inversion method and system for the electromagnetic wave logging while drilling azimuth, provided by the invention, on the premise of ensuring inversion precision, the inversion speed of the electromagnetic wave logging while drilling azimuth is improved, and the engineering requirements of geosteering while drilling on precision and real time are met. The method and the system for inverting the electromagnetic wave logging while drilling azimuth of the combined residual neural network and the L-M algorithm are described in detail below with reference to the accompanying drawings.
Referring to fig. 1, an embodiment of the first aspect of the present invention provides a method for inversion of electromagnetic logging while drilling azimuth by combining a residual neural network and an L-M algorithm, which specifically includes the steps of:
s1, constructing a stratum resistivity forward model, wherein the stratum resistivity forward model comprises a single-interface model, a double-interface model and a multi-interface model, and the change parameters of each model are stratum resistivity and edge detection distance.
Specifically, when the formation resistivity forward model is constructed, a plurality of points are taken at set sampling intervals in a formation resistivity setting range, a plurality of points are taken at set sampling intervals in a side detection distance setting range, different formation resistivities form different data pairs under the same model, and a plurality of different formation models are arranged in a single-interface model, a double-interface model and a multi-interface model. For example: the stratum resistivity is 1-1000 omega.m, the sampling interval is 10, 100 points are taken in total, the edge detection distance is 0-5m, the sampling interval is 0.1m, 50 points are taken in total, different stratum resistivities form different data pairs under the same model, 500000 different stratum models are contained in a single-interface model, and 500000 stratum models are simultaneously contained in a double-interface model and a multi-interface model. It should be noted that, the setting range and sampling interval of the formation resistivity and the edge detection distance can be defined according to actual requirements.
S2, setting stratum change parameters, and obtaining electromagnetic wave logging response of the azimuth while drilling corresponding to the stratum resistivity forward model through forward modeling.
Specifically, the logging response of the electromagnetic wave along with the drilling azimuth of all stratum resistivity forward models is obtained through simulation of the forward algorithm of the electromagnetic wave along with the drilling azimuth one-dimensional layered medium.
S3, dividing the logging response of the azimuth electromagnetic wave while drilling into a training set and a verification set according to a set proportion, and constructing a test set according to stratum change parameters set in the step S2.
It should be noted that, the proportion of the training set to the verification set may be divided according to actual requirements, and may be 80% for the training set and 20% for the verification set; the training set may be 90% and the verification set may be 10%.
S4, constructing a residual neural network by taking the logging response of the azimuth electromagnetic wave while drilling as input and the stratum parameter as output, and training the residual neural network by adopting a training set to obtain a residual neural network prediction model, wherein the stratum parameter report comprises stratum resistivity and stratum interface.
Specifically, the residual neural network prediction model directly transmits input to output as an initial result in a short-circuit connection mode, and the output result is expressed as:
H(X)=F(X)+X（1）
In the formula, H (X) is an output result, F (X) is a residual error, and X is an input.
It should be noted that, the residual neural network is a deep convolutional neural network, and the degradation problem when the depth of the neural network is increased is solved by introducing residual blocks. The residual block may pass the input data directly to the output layer to reduce the impact of depth increase on network performance. Variables that the residual neural network can customize include the number of network layers, the number of neurons per layer of the network, the activation function, the number of residual blocks, the optimizer, the learning rate, the regularization coefficient, etc. Reasonable selection and adjustment are needed in specific application scenes. Specifically, in some embodiments, the tuning of the residual neural network prediction model is a network parameter of the residual neural network, where the tuning includes a network layer number, a neuron number of each layer of the network, an activation function, a residual block number, an optimizer, a learning rate, and a regularization coefficient.
Specifically, in some embodiments, the residual neural network adopts ResNet network structure, the residual network includes a plurality of residual blocks, each residual block is composed of two basic blocks, and each basic block includes two convolution layers and one jump connection. The step size of the first convolutional layer of each residual block is different from the step sizes of the remaining convolutional layers, for example: the step size of the first convolution layer of each residual block is 2, and the step sizes of the rest convolution layers are 1. The convolution kernel size of the convolution layer is not limited to 3×3, and may be defined according to practical situations.
The activation function determines the output value of the neuron and can influence the nonlinear expression capacity and convergence speed of the residual neural network. Specifically, in some embodiments, the activation function employs a ReLU activation function.
It should be noted that a larger batch size may increase the training speed of the model, but may decrease the generalization ability of the model. Smaller batch sizes require more iterations, but can increase the generalization ability of the model. Specifically, in some embodiments, a batch size of 128 is selected.
Specifically, in some embodiments, the loss function selects a Mean Square Error (MSE) loss function.
The optimizer determines an updating strategy of the residual neural network prediction model parameters, and can influence the training speed and the convergence effect. Specifically, in some embodiments, the optimizer selects an Adam optimizer.
The learning rate determines the updating amplitude of the residual neural network prediction model parameters, and can influence the training speed and the convergence effect. Specifically, in some embodiments, the learning rate is set to 0.001.
The regularization coefficient determines the complexity of the residual neural network prediction model, and can influence the generalization capability and the robustness of the residual neural network prediction model. Specifically, in some embodiments dropout=0.5 is used to help control the complexity of the residual neural network prediction model, avoiding the over-fitting problem.
Referring to fig. 2, the residual neural network training is performed according to the flow shown in fig. 2, and time and error variations of the training are recorded. As shown in fig. 3 and 4, the red line in the graph represents the trend of the Mean Square Error (MSE), the x-axis represents training rounds, and the y-axis represents the MSE value, using a logarithmic scale. As can be seen from a comparison of fig. 3 and 4, when the training round is greater than 1000 times, the tendency of MSE to decrease becomes gentle, stabilizing around 0.01, so in order to reduce the training time with guaranteed accuracy, in particular, in some embodiments, the training round may be set to 1000.
S5, inputting the verification set into a residual neural network prediction model to obtain a prediction result, and combining the prediction result with a test set
And comparing the real stratum parameters, and calculating a prediction error to evaluate the prediction effect and generalization capability of the residual neural network prediction model.
Specifically, in order to verify the performance of the residual neural network prediction model, the occurrence of overfitting linearity of the residual neural network prediction model is avoided. In some embodiments of the invention, a K-fold cross validation evaluation method is selected to validate the residual neural network prediction model, so that the data set is effectively utilized for training and testing, and the influence caused by unbalanced data set division is reduced. The method comprises the following specific steps:
(1) Dividing an original verification set into K=10 parts, wherein each part is called a 'fold', and keeping class distribution balance;
(2) Selecting one of the folds as a verification set, using the rest K-1 folds as a training set, training a residual neural network prediction model on the training set, testing the residual neural network prediction model on the verification set, and recording performance indexes of the residual neural network prediction model.
(3) And (3) repeating the step (1), selecting different folds as verification sets, and taking the average value of the performance indexes of 10 experiments as the performance evaluation index of the residual neural network prediction model until each fold is used as one verification set, namely, 10 experiments are carried out. Table 1 gives the K-fold cross validation results of the residual neural network prediction model. As can be seen from table 1, the training round was set to 1000 times, and the training error and the test error were tested, and after 10 times of verification, the mean square error of the training error and the test error were substantially identical, so that the stability of the network prediction model was verified.
TABLE 1
S6, inputting the prediction result into an L-M algorithm, and carrying out inversion iteration to obtain the optimal stratum resistivity model parameter.
The L-M (Levenberg-Marquardt) algorithm belongs to a nonlinear least square optimization algorithm, and a predicted value of a residual neural network predicted model is set to be f (x), wherein x is a model parameter, an observed value is y, and a residual function of the L-M (Levenberg-Marquardt) algorithm and the L-M (Levenberg-Marquardt) algorithm can be expressed asFurther, solving the residual neural network prediction model parameters so that the sum of squares of differences between the observed data and the residual neural network prediction model predicted values is minimum, and evaluating a loss function of generalization capability of the residual neural network prediction model as follows:
（2）
First, a jacobian matrix J (x) of the objective function is obtained, wherein Where e m is the mth component of the residual vector and x n is the nth component of the model parameters, the size of the jacobian matrix J (x) is (m, n), which can be written as:
（3）
in some embodiments of the invention, m=1, n=2.
Then, the step size is determined so that the objective functionTo a minimum, combined with Taylor expansionThe objective function may be written as:
（4）
Taking the derivative of deltax and letting the derivative be 0, we can obtain:
（5）
The non-negative parameter lambda of the L-M algorithm is introduced, and the above formula is rewritten, so that the following can be obtained:
（6）
wherein I is an identity matrix, lambda is a non-negative parameter of an L-M algorithm, and lambda is more than or equal to 0.
In some embodiments of the invention, the lambda value is dynamically adjusted based on iteration error variations.
In the L-M algorithm, the prediction result obtained in the step S5 is input into the L-M algorithm as an initial parameter value x 0, and iteration is performed. And (3) iteratively calculating the update quantity delta x of the current value x k according to the formula (6) each time, and finally obtaining updated model parameter values:
xk+1=xk+Δx（7）
In each iteration, the L-M algorithm adjusts the value of λ to achieve the effect of balancing convergence speed and accuracy. If the current iteration makes the error smaller, λ may be reduced; if the error becomes large, λ may increase. And when the iteration is completed until the error of a certain step is smaller than a preset threshold value, stopping the iteration by the L-M algorithm, and outputting the finally obtained stratum resistivity model parameters.
S7, constructing an actual formation resistivity model according to a horizontal well interactive modeling result of the target zone target well, and executing the steps S1-S6 to obtain actual formation resistivity model parameters of a target zone target well target interval, including formation resistivity and formation interfaces.
In some embodiments, after said step S2 and before step S3, a pre-processing step is further included: and normalizing the logging response of the electromagnetic wave of the azimuth while drilling to obtain the logging response of the electromagnetic wave of the azimuth while drilling. The mathematical expression of the normalization process is expressed as:
Wherein, X o represents the normalized data, X i represents the original sample data to be normalized, i represents the number of samples corresponding to the i-th sample, μ represents the feature mean vector of each dimension in the original sample data, and σ represents the feature standard deviation vector of each dimension in the original sample data.
Specifically, a standard fraction (z-score) process is performed to implement a normalization process for the logging response of the azimuth while drilling electromagnetic wave, so as to scale the data features to a certain data distribution centered on 0, which can maintain the original data feature distribution, does not change the distribution type, and can also make different features comparable. And the electromagnetic wave logging response while drilling is normalized, so that the deviation of the characteristic weight can be reduced when a model is built later, and the training efficiency and the prediction accuracy of the model are improved.
Referring to fig. 5, in a second aspect of the present invention, there is provided an inversion system for logging while drilling azimuth electromagnetic waves, for implementing the inversion method for logging while drilling azimuth electromagnetic waves of the joint residual neural network and the L-M algorithm according to the first aspect of the present invention, including:
The forward model construction module 1 is used for constructing a stratum resistivity forward model;
The forward modeling module 2 is used for setting stratum change parameters and acquiring the electromagnetic wave logging response of the azimuth while drilling corresponding to the stratum resistivity forward modeling through forward modeling;
The data set construction module 3 divides the logging response of the azimuth electromagnetic wave while drilling into a training set and a verification set according to a set proportion, and constructs a test set according to stratum change parameters set by the forward modeling module;
The model construction module 4 is used for constructing a residual neural network by taking the logging response of the azimuth electromagnetic wave while drilling as input and the stratum parameter as output, and training the residual neural network by adopting a training set to obtain a residual neural network prediction model;
the evaluation module 5 is used for inputting the result into the residual neural network prediction model through the verification set to obtain a prediction result, comparing the prediction result with the stratum parameter change number in the test set, and calculating a prediction error to evaluate the prediction effect and generalization capability of the residual neural network prediction model;
The inversion module 6 inputs the prediction result as an initial value into an L-M algorithm, and inversion iteration is carried out to obtain the optimal stratum resistivity model parameter;
An application module 7 arranged to: constructing an actual stratum resistivity model according to the horizontal well interactive modeling result of the target well of the target zone; the control execution forward model building module 1, the forward modeling module 2, the data set building module 3, the model building module 4, the evaluation module 5 and the inversion module 6 are used for outputting actual stratum parameters of a target interval of a target well.
In some embodiments, the system further includes a preprocessing module 8, configured to normalize the electromagnetic wave logging response while drilling obtained by the forward modeling module to obtain a normalized electromagnetic wave logging response while drilling. The mathematical expression of the normalization process is expressed as:
Wherein, X o represents the normalized data, X i represents the original sample data to be normalized, i represents the number of samples corresponding to the i-th sample, μ represents the feature mean vector of each dimension in the original sample data, and σ represents the feature standard deviation vector of each dimension in the original sample data.
Specifically, a standard fraction (z-score) process is performed to implement a normalization process for the logging response of the azimuth while drilling electromagnetic wave, so as to scale the data features to a certain data distribution centered on 0, which can maintain the original data feature distribution, does not change the distribution type, and can also make different features comparable. And the electromagnetic wave logging response while drilling is normalized, so that the deviation of the characteristic weight can be reduced when a model is built later, and the training efficiency and the prediction accuracy of the model are improved.
The following combines the concrete embodiment to verify the superiority and stability of the electromagnetic wave logging inversion method and system of the combined residual neural network and L-M algorithm while drilling.
1. And verifying inversion effect of the residual neural network prediction model.
And selecting the resistivity inversion and the stratum boundary inversion results of the multi-interface model to verify the inversion speed and inversion effect of the residual neural network prediction model. FIG. 6 shows four interface resistivity forward models established by the method and the system for the electromagnetic wave logging while drilling inversion of the joint residual neural network and the L-M algorithm, wherein inversion parameters comprise upper and lower interface distances d2 and d3 of layers where a measuring instrument is located, upper and lower interface distances d1 and d4 of adjacent layers, and formation resistivity R1, R2, R3, R4 and R5 of each layer. The simulated measuring instrument moves from the upper layer to the lower layer through the stratum interface, the stratum resistivity parameters from top to bottom are set to be 10 omega.m, 100 omega.m, 10 omega.m, 100 omega.m and 10 omega.m, and the stratum thickness is set to be 2m, 1m, 2m and 2m, so that the stratum resistivity and the interface inversion effect based on the residual neural network prediction model are shown in figures 7 and 8. FIG. 7 shows the inversion result of the resistivity of the layer where the logging instrument is located, and from the inversion result of the resistivity, it can be seen that the inversion result is basically consistent with the true value of the four-interface resistivity forward model, and individual abnormal points exist. Fig. 8 shows inversion results of the stratum interface, black diagonal lines represent the borehole track, black dots represent the inversion effect of the lower interface, and blue dots represent the inversion result of the upper interface, and it can be seen that the inversion result has perceptibility on the adjacent layer interface, but the error of the adjacent layer interface is larger than that of the current layer. The above analysis shows that the formation interface, whether resistivity or formation interface, has substantially approached the four-interface resistivity forward model. The method lays a foundation for taking the output of the residual neural network prediction model as an initial value of an L-M algorithm. Meanwhile, compared with the traditional inversion method, the inversion time of the residual neural network prediction model is improved by more than three orders of magnitude, the calculation speed is high, obvious advantages are achieved, and a foundation is provided for real-time geosteering engineering requirements.
2. And verifying the combined residual neural network prediction model and the inversion effect of the L-M algorithm.
Fig. 9 and 10 show the inversion results of the joint residual neural network prediction model and the L-M algorithm. It can be seen that the fitness of the borehole trajectory inverted by the joint residual neural network prediction model and the L-M algorithm is higher, compared with the borehole trajectory inverted by using the residual neural network prediction model alone in fig. 7 and 8, the error is smaller, and the resistivity inversion is also identical to the true value of the four-interface resistivity forward model. When the logging instrument enters the first layer, the inversion error is larger at the moment, and the interface fluctuation is larger. When the logging instrument drills to the second layer, the inversion error increases, but the position of the interface can still be determined. However, when the logging instrument is positioned in the middle layer, the inversion error is smaller, and the situation that the logging instrument just enters the stratum and has larger interference can be judged, so that the superiority and the stability of the logging-while-drilling azimuth electromagnetic wave logging inversion method and system of the combined residual neutral network and the L-M algorithm can be still proved.
3. The inversion effect of the formation resistivity forward model with the same thickness, different resistivity distribution, the same resistivity distribution, different thickness, different resistivity, different thickness and the like of the combined residual neural network and the L-M algorithm is further tested.
The resistivity and the interface inversion result under different conditions are obtained according to the calculation result, and the performance is very stable as shown in figures 11-16.
According to the method and the system for inverting the electromagnetic wave logging while drilling azimuth by combining the residual neural network and the L-M algorithm, inversion accuracy is improved by 6% compared with the residual neural network, and the method and the system are high in accuracy and can be used for real-time inversion in a specific complex environment. The application example of the multi-interface shows that the technical scheme of the method and the system for the electromagnetic wave logging while drilling inversion of the joint residual neural network and the L-M algorithm is more stable, the accuracy in inversion of a thin layer is higher, and the calculated amount in inversion is only 1/4 of that in the traditional L-M algorithm, so that the method and the system have great advantages in multi-layer inversion, and mathematical simulation results prove that the method and the system integrate the advantages of the residual neural network and the L-M algorithm.
The above-described embodiments are intended to illustrate the present invention, not to limit it, and any modifications and variations made thereto are within the spirit of the invention and the scope of the appended claims.
Claims (8)
1. The method for inverting the electromagnetic wave logging while drilling of the azimuth by combining the residual neural network and the L-M algorithm is characterized by comprising the following specific steps of:
s1, constructing a stratum resistivity forward model, wherein the stratum resistivity forward model comprises a single-interface model, a double-interface model and a multi-interface model, and stratum change parameters of each model are stratum resistivity and edge detection distance;
S2, setting stratum change parameters, and obtaining electromagnetic wave logging response of the azimuth while drilling corresponding to the stratum resistivity forward model through forward modeling;
s3, dividing the logging response of the azimuth electromagnetic wave while drilling into a training set and a verification set according to a set proportion, and constructing a test set according to stratum change parameters set in the step S2;
S4, constructing a residual neural network by taking the logging response of the azimuth electromagnetic wave while drilling as input and the stratum parameter as output, and training the residual neural network by adopting a training set to obtain a residual neural network prediction model, wherein the stratum parameter report comprises stratum resistivity and stratum interface;
S5, inputting the result into a residual neural network prediction model through a verification set to obtain a prediction result, comparing the prediction result with stratum change parameters in a test set, and calculating a prediction error to evaluate the prediction effect and generalization capability of the residual neural network prediction model;
s6, inputting a prediction result as an initial value into an L-M algorithm, and inverting and iterating to obtain the optimal stratum resistivity model parameter;
S7, constructing an actual formation resistivity model according to the horizontal well interactive modeling result of the target zone target well, and executing the steps S1-S6 to obtain actual formation resistivity model parameters of the target zone target well target interval.
2. The method for inversion of electromagnetic logging while drilling of azimuth electromagnetic waves combining a residual neural network and an L-M algorithm according to claim 1, further comprising a preprocessing step after said step S2 and before said step S3: and normalizing the logging response of the electromagnetic wave of the azimuth while drilling to obtain the logging response of the electromagnetic wave of the azimuth while drilling.
3. The method for inversion of electromagnetic logging while drilling in combination with residual neural network and L-M algorithm according to claim 1 or 2, wherein in step S1, when the formation resistivity forward model is constructed, a plurality of points are taken at set sampling intervals within a formation resistivity setting range, a plurality of points are taken at set sampling intervals within a sounding distance setting range, different formation resistivities form different data pairs under the same model, and a plurality of different formation models are provided in a single-interface model, a double-interface model and a multi-interface model.
4. The method for inversion of electromagnetic logging while drilling of azimuth electromagnetic wave combining residual neural network and L-M algorithm according to claim 1 or 2, wherein in step S1, the response of electromagnetic logging while drilling of all stratum resistivity forward model is obtained by simulation of the forward algorithm of one-dimensional laminar medium of electromagnetic wave while drilling.
5. The method for inversion of electromagnetic logging while drilling of azimuth electromagnetic waves combining a residual neural network and an L-M algorithm according to claim 1 or 2, wherein in said step S4, the parameters of the residual neural network prediction model are parameters of the residual neural network, and the parameters include the number of network layers, the number of neurons of each layer of network, an activation function, the number of residual blocks, an optimizer, a learning rate, and a regularization coefficient.
6. The method of inversion of electromagnetic logging while drilling azimuth according to claim 1 or 2, wherein in step S4, the residual neural network adopts ResNet network structure, the residual network comprises a plurality of residual blocks, each residual block is composed of two basic blocks, each basic block comprises two convolution layers and one jump connection, and the step size of the first convolution layer of each residual block is different from the step sizes of the rest convolution layers.
7. An inversion system for logging while drilling azimuth electromagnetic waves, for implementing an inversion method for logging while drilling azimuth electromagnetic waves of a combined residual neural network and L-M algorithm as claimed in any one of claims 1 to 6, comprising:
The forward model construction module is used for constructing a stratum resistivity forward model;
The forward modeling module is used for setting stratum change parameters and acquiring an electromagnetic wave logging response of the azimuth while drilling corresponding to the stratum resistivity forward modeling module through forward modeling;
the data set construction module divides the logging response of the azimuth electromagnetic wave while drilling into a training set and a verification set according to a set proportion, and constructs a test set according to stratum change parameters set by the forward modeling module;
the model construction module is used for constructing a residual neural network by taking the logging response of the azimuth electromagnetic wave while drilling as input and the stratum parameter as output, and training the residual neural network by adopting a training set to obtain a residual neural network prediction model;
the evaluation module is used for obtaining a prediction result by inputting the verification set into the residual neural network prediction model, comparing the prediction result with stratum change parameters in the test set, and calculating a prediction error so as to evaluate the prediction effect and generalization capability of the residual neural network prediction model;
the inversion module inputs the prediction result as an initial value into an L-M algorithm, and inversion iteration is carried out to obtain optimal stratum resistivity model parameters;
An application module configured to: constructing an actual stratum resistivity model according to the horizontal well interactive modeling result of the target well of the target zone; and controlling and executing functions of the forward model building module, the forward modeling module, the data set building module, the model building module, the evaluation module and the inversion module, and outputting actual formation resistivity model parameters of the target interval of the target well.
8. The while-drilling azimuth electromagnetic wave logging inversion system of claim 7, further comprising a preprocessing module for normalizing the while-drilling azimuth electromagnetic wave logging response obtained by the forward modeling module to obtain a normalized while-drilling azimuth electromagnetic wave logging response.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202410581857.0A CN118171582B (en) | 2024-05-11 | 2024-05-11 | Electromagnetic wave logging inversion method and system for azimuth while drilling by combining residual neural network and L-M algorithm |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202410581857.0A CN118171582B (en) | 2024-05-11 | 2024-05-11 | Electromagnetic wave logging inversion method and system for azimuth while drilling by combining residual neural network and L-M algorithm |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN118171582A CN118171582A (en) | 2024-06-11 |
| CN118171582B true CN118171582B (en) | 2024-07-12 |
Family
ID=91353050
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202410581857.0A Active CN118171582B (en) | 2024-05-11 | 2024-05-11 | Electromagnetic wave logging inversion method and system for azimuth while drilling by combining residual neural network and L-M algorithm |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN118171582B (en) |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113882853A (en) * | 2020-07-03 | 2022-01-04 | 中国石油化工股份有限公司 | Method for transmitting near-bit logging while drilling data |
| CN116992754A (en) * | 2023-06-25 | 2023-11-03 | 山东科技大学 | Rapid inversion method for logging while drilling data based on transfer learning |
Family Cites Families (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20120209528A1 (en) * | 2011-02-10 | 2012-08-16 | Baker Hughes Incorporated | Inversion-Based Method to Correct for the Pipe Residual Signal in Transient MWD Measurements |
| CN117744505B (en) * | 2024-02-20 | 2024-04-26 | 电子科技大学 | Depth learning-based inversion method for electromagnetic wave resistivity of azimuth while drilling |
-
2024
- 2024-05-11 CN CN202410581857.0A patent/CN118171582B/en active Active
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113882853A (en) * | 2020-07-03 | 2022-01-04 | 中国石油化工股份有限公司 | Method for transmitting near-bit logging while drilling data |
| CN116992754A (en) * | 2023-06-25 | 2023-11-03 | 山东科技大学 | Rapid inversion method for logging while drilling data based on transfer learning |
Also Published As
| Publication number | Publication date |
|---|---|
| CN118171582A (en) | 2024-06-11 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| US8855986B2 (en) | Iterative method and system to construct robust proxy models for reservoir simulation | |
| CN108875122B (en) | Artificial intelligence method and system for calculating geological parameters by using logging-while-drilling data | |
| KR102155101B1 (en) | System and method for predicting ground layer information, and a recording medium having computer readable program for executing the method | |
| CN111722283B (en) | Stratum velocity model building method | |
| CN111980662B (en) | Rapid processing method for lateral logging data of inclined shaft anisotropic stratum array | |
| CN112965103B (en) | Multi-pore reservoir prestack seismic probability multi-channel inversion method | |
| CN114723095A (en) | Missing well logging curve prediction method and device | |
| CN116992754A (en) | Rapid inversion method for logging while drilling data based on transfer learning | |
| CN106568647A (en) | Nerve network-based concrete strength predication method | |
| CN117744505B (en) | Depth learning-based inversion method for electromagnetic wave resistivity of azimuth while drilling | |
| Wong et al. | Use of neural network methods to predict porosity and permeability of a petroleum reservoir | |
| CN116911216B (en) | Reservoir oil well productivity factor assessment and prediction method | |
| CN118171582B (en) | Electromagnetic wave logging inversion method and system for azimuth while drilling by combining residual neural network and L-M algorithm | |
| CN113468466A (en) | Multi-working-condition one-dimensional wave equation solving method based on neural network | |
| Deutsch | Conditioning reservoir models to well test information | |
| CN107102379A (en) | A kind of method that seat earth watery prediction is carried out based on many attribution inversions | |
| CN115049173B (en) | Deep learning and Eaton method coupling driving stratum pore pressure prediction method | |
| Sarkheil et al. | The fracture network modeling in naturally fractured reservoirs using artificial neural network based on image loges and core measurements | |
| CN108765562B (en) | Oil-gas productivity evaluation method based on three-dimensional geological model | |
| Tang et al. | Combining machine learning and physics modelling to determine the natural cave property with fracturing curves | |
| CN115267927B (en) | Multi-boundary curtain type geosteering method based on ant colony-gradient series algorithm | |
| Talkhestani | Prediction of effective porosity from seismic attributes using locally linear model tree algorithm | |
| CN112147679A (en) | Lithology prediction method and device based on elastic parameters under fuzzy logic framework | |
| CN115421181B (en) | Three-dimensional geological model phase control attribute modeling method based on deep learning | |
| CN117688821B (en) | Prediction method for side slope stability of dumping site |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant |