CN117744505A - Deep learning-based inversion method for electromagnetic wave resistivity of azimuth while drilling - Google Patents
Deep learning-based inversion method for electromagnetic wave resistivity of azimuth while drilling Download PDFInfo
- Publication number
- CN117744505A CN117744505A CN202410187481.5A CN202410187481A CN117744505A CN 117744505 A CN117744505 A CN 117744505A CN 202410187481 A CN202410187481 A CN 202410187481A CN 117744505 A CN117744505 A CN 117744505A
- Authority
- CN
- China
- Prior art keywords
- drilling
- resistivity
- deep learning
- electromagnetic wave
- azimuth
- 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.)
- Granted
Links
- 238000005553 drilling Methods 0.000 title claims abstract description 65
- 238000000034 method Methods 0.000 title claims abstract description 53
- 238000013135 deep learning Methods 0.000 title abstract description 16
- 238000012545 processing Methods 0.000 claims abstract description 11
- 230000004044 response Effects 0.000 claims abstract description 8
- 238000003384 imaging method Methods 0.000 claims abstract description 6
- 238000013136 deep learning model Methods 0.000 claims description 47
- 238000012549 training Methods 0.000 claims description 37
- 230000008569 process Effects 0.000 claims description 20
- 238000004422 calculation algorithm Methods 0.000 claims description 19
- 230000015572 biosynthetic process Effects 0.000 claims description 17
- 238000005755 formation reaction Methods 0.000 claims description 17
- 230000006870 function Effects 0.000 claims description 13
- 238000010606 normalization Methods 0.000 claims description 12
- 230000007246 mechanism Effects 0.000 claims description 11
- 238000012795 verification Methods 0.000 claims description 11
- 238000007781 pre-processing Methods 0.000 claims description 10
- 238000011176 pooling Methods 0.000 claims description 8
- 238000005259 measurement Methods 0.000 claims description 7
- 238000005457 optimization Methods 0.000 claims description 6
- 230000004913 activation Effects 0.000 claims description 5
- 230000009466 transformation Effects 0.000 claims description 4
- 238000010276 construction Methods 0.000 claims description 3
- 238000000605 extraction Methods 0.000 claims description 2
- 238000004364 calculation method Methods 0.000 abstract description 4
- 238000005516 engineering process Methods 0.000 abstract description 3
- 238000013528 artificial neural network Methods 0.000 description 9
- 238000002474 experimental method Methods 0.000 description 7
- 238000004088 simulation Methods 0.000 description 5
- 238000010200 validation analysis Methods 0.000 description 5
- 238000010586 diagram Methods 0.000 description 4
- 230000000694 effects Effects 0.000 description 4
- 238000011156 evaluation Methods 0.000 description 4
- 230000006399 behavior Effects 0.000 description 3
- 238000011161 development Methods 0.000 description 3
- 230000009286 beneficial effect Effects 0.000 description 2
- 239000011159 matrix material Substances 0.000 description 2
- 238000005070 sampling Methods 0.000 description 2
- 230000002411 adverse Effects 0.000 description 1
- 238000004458 analytical method Methods 0.000 description 1
- 238000013459 approach Methods 0.000 description 1
- 230000005540 biological transmission Effects 0.000 description 1
- 230000000052 comparative effect Effects 0.000 description 1
- 238000007405 data analysis Methods 0.000 description 1
- 238000001514 detection method Methods 0.000 description 1
- 230000005672 electromagnetic field Effects 0.000 description 1
- 230000008030 elimination Effects 0.000 description 1
- 238000003379 elimination reaction Methods 0.000 description 1
- 230000006698 induction Effects 0.000 description 1
- 238000000691 measurement method Methods 0.000 description 1
- 239000003208 petroleum Substances 0.000 description 1
- 238000003672 processing method Methods 0.000 description 1
- 238000012360 testing method Methods 0.000 description 1
- 230000000007 visual effect Effects 0.000 description 1
Landscapes
- Geophysics And Detection Of Objects (AREA)
Abstract
The invention relates to a logging-while-drilling data processing technology, and provides a drilling-while-drilling azimuth electromagnetic wave resistivity inversion method based on deep learning. The method and the device realize the rapid calculation of the electromagnetic wave resistivity logging response of the azimuth while drilling, and obtain the resistivity imaging of the stratum to provide reliable geosteering decision guidance for drilling.
Description
Technical Field
The invention relates to the technical field of data processing while drilling, in particular to a method for inverting the resistivity of electromagnetic waves in azimuth while drilling based on deep learning.
Background
With the continuous development of highly deviated wells and horizontal wells, resistivity anisotropy becomes one of the main factors affecting formation evaluation measurement accuracy. With the development of oil and gas field exploration, electromagnetic wave resistivity logging is an important measurement method in geophysical exploration. The measured resistivity data provides a variety of formation information.
In the drilling process of the highly deviated well and the horizontal well, the parameter information of the stratum can be obtained by explaining the logging information of the electromagnetic wave logging instrument while drilling. How to quickly and accurately interpret logging data in a formation is important to geological exploration. In order to evaluate logging data in a reservoir, an efficient inversion model needs to be established to accurately obtain formation information from logging data measured in the formation by the logging instrument.
In the inversion of electromagnetic wave log data, an iterative inversion method is often used to process log data. The method needs to construct a cost function to measure the difference between the true value and the predicted value, and the error between the true data and the predicted data is continuously reduced in an iterative mode. However, since the iterative inversion method requires multiple calls to forward model, the time required is often long. And the inversion result is easy to have multiple solutions, and depends on the selection of an initial model. In practical logging, real-time inversion is often required to be performed on logging data, and the inversion method is difficult to meet the requirements. How to quickly and accurately interpret electromagnetic wave logging data to realize real-time inversion of the logging data is one of the problems which need to be solved in the current geological exploration.
Although some studies have been conducted to invert logging data in isotropic formations, most of the studies use shallow neural network structures and the accuracy of inversion results is greatly limited. In recent years, the rapid development of deep learning technology provides a new idea for the field of geological exploration. The inversion method based on deep learning can rapidly and accurately explain logging data of the electromagnetic wave logging instrument while drilling in the layered stratum.
For example, document 1: zhu Gaoyang Forward and reverse modeling of electromagnetic logging while drilling data in a layered reservoir based on deep learning [ D ]. Shandong university 2020. A novel method based on Deep Learning (DL) is presented to solve the time consuming problem of inversion of induction logging data in a layered anisotropic formation. And good results are obtained in the simulation data.
For example, document 2: shahriri M, pardo D, pic n A, et al A deep learning approach to the inversion of borehole resistivity measurements J Computational Geosciences, 2020, 24:971-994. This method is also applied to wellbore resistivity inversion via deep neural networks DNNs and focuses on error control and loss function selection, again demonstrating the feasibility of inversion using this method.
For example, document 3: fan J, zhang W, chen W, et al Inversion based on deep learning of logging-while-drilling directional resistivity measurements [ J ]. Journal of Petroleum Science and Engineering, 2022, 208:109677. A BiLSTM-based deep learning inversion method is studied, which not only can accurately invert formation parameters, but also can perform uncertainty estimation on inversion parameters.
Disclosure of Invention
The invention aims to solve the technical problem of real-time inversion of logging data of electromagnetic wave resistivity of azimuth while drilling, and provides an inversion method of electromagnetic wave resistivity of azimuth while drilling based on deep learning.
The invention adopts the technical scheme that the method for inverting the resistivity of the electromagnetic wave along with the drilling azimuth based on the deep learning comprises the following steps:
step S1: establishing a forward model of the electromagnetic wave resistivity of the azimuth while drilling, setting a stratum model to be a three-layer stratum model, setting stratum model parameters required in the forward process according to inversion parameters, and calculating the electromagnetic wave resistivity data of the azimuth while drilling by adopting a one-dimensional forward algorithm; the three-layer stratum model comprises an upper stratum, a middle layer and a lower stratum, and an azimuth electromagnetic wave resistivity logging instrument while drilling is arranged on the middle layer; the logging instrument comprises different types of transmitting coils and receiving coils, and different geological signals and resistivity signals are generated by setting different source distances and coil combinations;
step S2: preprocessing the electromagnetic wave resistivity data of the azimuth while drilling and the real measurement data of the instrument obtained by a one-dimensional forward algorithm to obtain data of a training set and a verification set, wherein the preprocessing is used for accelerating training of a deep learning model, accelerating convergence and improving efficiency;
step S3: based on a ResNet network, the convolution layer and the rest blocks except the activation function and the batch normalization layer in the pooling layer in the ResNet network are replaced by full-connection layers, and feature extraction is carried out on input data by combining a multi-head attention mechanism, so that the construction of a deep learning model for the inversion of the azimuth electromagnetic wave resistivity while drilling is completed;
step S4: aiming at the inversion problem of the resistivity of the electromagnetic waves in the azimuth while drilling, setting different depths and widths of a deep learning model; respectively inputting training sets into the deep learning models for training, and adjusting the super parameters of the deep learning models by using a Bayesian optimization parameter adjustment algorithm until the super parameters of the optimal deep learning models are determined, so as to complete the training of the deep learning models with different depths and widths; respectively inputting the verification sets into a deep learning model, and selecting the depth and width deep learning model with minimum training loss and minimum verification loss as the optimal deep learning model for inversion of the electromagnetic wave resistivity of the azimuth while drilling;
step S5: and inputting the received electromagnetic wave logging data into an optimal deep learning model, and outputting the resistivity and interface position information of the current stratum by the deep learning model to finally obtain the resistivity imaging of the stratum.
The method has the beneficial effects that the problem of numerical overflow is effectively solved by a full tensor one-dimensional forward modeling method based on the multi-layer anisotropic medium electromagnetic field, and the rapid convergence of electromagnetic wave resistivity logging response calculation while drilling is realized. The response data obtained by forward modeling and instrument measured data are used for training a deep learning model for inversion of the electromagnetic wave resistivity of the azimuth while drilling, so that the trained and selected optimal deep learning model can output the resistivity of the current stratum and interface position information by utilizing the received electromagnetic wave logging data, and finally, the resistivity imaging of the stratum is obtained, and reliable geosteering decision guidance is provided for drilling. The invention does not depend on the selection of the initial stratum model, and the inversion speed meets the real-time operation requirement.
Drawings
FIG. 1 is a schematic flow chart of a method for inverting the resistivity of electromagnetic waves in azimuth while drilling based on deep learning;
FIG. 2 is a simplified model diagram of an electromagnetic wave resistivity instrument in azimuth while drilling provided by the invention;
FIG. 3 is a simplified model diagram of a three-layer formation structure provided by the invention;
FIG. 4 is a convolutional layer structure diagram of a deep learning model provided by an embodiment;
FIG. 5 is a network structure diagram of a deep learning model provided by an embodiment;
FIG. 6 is a logically exploded view of a multi-headed attention mechanism provided by an embodiment.
Detailed Description
The embodiment discloses a method for rapidly and accurately inverting the electromagnetic wave resistivity of a while-drilling azimuth by using a deep learning model, namely inversion of stratum parameters, wherein in the inversion process, the condition that the electromagnetic wave resistivity logging instrument of the while-drilling azimuth drills in a target layer is considered. Meanwhile, the influence of the upper and lower interface distances and the resistivity of the instrument distance when the logging instrument of the electromagnetic wave resistivity in the azimuth while drilling passes through the layers in the three-layer stratum is studied. According to the parameter inversion mode, a training set of the electromagnetic wave resistivity of the azimuth while drilling is obtained through a forward modeling method, and an inversion model is obtained through network training. And the inversion model is utilized to obtain the upper and lower interface distances of the instrument distance and the formation resistivity when the electromagnetic wave resistivity instrument in the azimuth while drilling passes through the layers. Thus, the user can conveniently and accurately acquire the underground condition in time according to inversion imaging.
As shown in fig. 1, the method for inverting the electromagnetic wave resistivity of the azimuth while drilling based on deep learning comprises the following steps:
step S1: and establishing a forward model of the electromagnetic wave resistivity of the azimuth while drilling, setting a stratum model to be a three-layer stratum model, setting stratum model parameters required in the forward process according to inversion parameters, and calculating the electromagnetic wave resistivity data of the azimuth while drilling by adopting a one-dimensional forward algorithm.
The purpose of establishing the forward modeling detection model of the azimuth electromagnetic wave resistivity while drilling is to calculate theoretical response data of the instrument in the underground stratum through a numerical simulation technology. In this model, we need to set the formation model parameters based on the inversion parameters, and implement numerical simulation by setting the desired instrument parameters. Instrument parameters as shown in fig. 2, in the z-axis direction of the model coordinates, T1 ', T2 ' are set as transmitting coils, R1, R2, R3 ' are receiving coils, R1, R2 are 4 inches 4 ʺ from the instrument measurement center, T1 ' are 22 inches 22 ʺ from the instrument measurement center, T2 ' are 36 inches 36 ʺ from the instrument measurement center, the distance between the transmitting coils and the receiving coils is called the source distance, we can use different source distances to generate electromagnetic wave signals, such as 24 inches source distance 24 ʺ, 82 inches source distance 82 ʺ, and 96 inches source distance 96 ʺ and by selecting the appropriate coil type to generate geologic signals and resistivity signals.
As shown in fig. 3, in order to more accurately simulate the resistivity distribution characteristics of the underground stratum, the constructed three-layer stratum model comprises an upper stratum, a middle layer and a lower stratum, and the electromagnetic wave resistivity logging instrument in the azimuth while drilling is arranged in the middle layer; the logging instrument comprises different types of transmitting coils and receiving coils, and different geological signals and resistivity signals are generated by setting different source distances and coil combinations;
the model parameters of the three-layer stratum model comprise the distances from the electromagnetic wave resistivity logging instrument along with the drilling to the upper interface and the lower interface, and the resistivities of the upper stratum, the middle layer and the lower stratum, wherein the resistivities comprise horizontal resistivity and vertical resistivity, and the resistivity characteristics of each layer are simulated by adjusting the model parameters. Resistivity the electrical anisotropy of the formation as an electromagnetic wave transmission medium is reflected by consideration of the horizontal resistivity and the vertical resistivity.
When forward modeling is performed, a certain range value needs to be set to limit the value range of the parameters so as to ensure the feasibility and the rationality of the model. The formation resistivity takes the value of 0.2-2000 omega ּ m and the formation thickness takes the value of 0.5-15 m, and by gradually adjusting and changing the parameters, a richer forward modeling result can be obtained, and response signals of the simulation instrument in the underground stratum can be further analyzed and simulated.
The one-dimensional forward algorithm is forward calculation of the response of the electromagnetic wave resistivity logging instrument along with the drilling direction in a multi-layer medium, and particularly solves the problem of Maxwell's equation set in an anisotropic medium, and based on a Hertz potential function, interface amplitude is introduced, so that the deducing process of electromagnetic wave propagation is completely converged, numerical overflow caused by an index increase term possibly occurring in engineering calculation is eliminated, and fast and converged forward simulation of the electromagnetic wave resistivity logging response along with the drilling direction is realized.
Step S2: preprocessing the electromagnetic wave resistivity data of the azimuth while drilling and the real measurement data of the instrument obtained by a one-dimensional forward algorithm, and taking the preprocessed data as data of a training set and a verification set.
Further, the step S2 specifically includes: the data preprocessing section can be divided into two aspects. First, preprocessing is performed on electromagnetic wave resistivity data of azimuth while drilling obtained through forward modeling. The goal in this respect is to optimize the data to speed up convergence and to improve accuracy of the final training results.
In processing the while-drilling azimuth electromagnetic wave resistivity data, it is first necessary to consider that the geological signal and the resistivity signal are of different orders of magnitude. To solve this problem, we have adopted a logarithmic processing method. By taking the logarithm of the data, the range of the data can be narrowed to be closer to the same order of magnitude, thereby facilitating the convergence of the algorithm.
Second, there may be some outliers or outliers to the resulting while drilling azimuth electromagnetic wave resistivity data. These values may be due to measurement errors or other factors. In order to avoid adverse effects of these outliers on the final training results, the data needs to be processed to remove sample outliers. Outliers can severely impact data analysis and training results, and therefore require identification and elimination of these outliers to ensure reliability and accuracy of the final training results.
And finally, carrying out standard normalization processing on the electromagnetic wave resistivity data of the azimuth while drilling. The standard normalization process scales the data according to the mean value and standard deviation of the data, so that the distribution range of the data is in a unified interval. By normalizing the data to a standard, converting it to a value within a specific range, the convergence process of the algorithm can be accelerated. Standard normalization can ensure that the data has similar dimensions and distribution during the training process, preventing certain features from having excessive impact on the model training process.
On the other hand, a similar preprocessing step is required for the measured data. The measured data also has the problem that the geological signal and the resistivity signal are different orders of magnitude in value. Therefore, it is also necessary to perform the logarithmic processing and the standard normalization processing to accelerate the convergence speed of the algorithm and optimize the training result.
Through the data preprocessing in the two aspects, the electromagnetic wave resistivity data and the measured data of the azimuth while drilling are more consistent in value, the convergence process of an algorithm is quickened, and a more reliable and efficient data base is provided for subsequent training and analysis.
Step S3: based on the ResNet network, the convolution layer and the rest blocks except the activation function and the batch normalization layer in the pooling layer in the ResNet network are replaced by full-connection layers, so that the new network structure has stronger adaptability and expression capability when the nonlinear regression problem is processed, and the characteristic information is extracted by weighting input data by combining a multi-head attention mechanism, so that the construction of a deep learning model for the inversion of the resistivity of the electromagnetic waves in the azimuth while drilling is completed; in combination with a multi-head attention mechanism, the deep learning model can extract key features and improve the performance and accuracy of the model by more accurately understanding the relevance between input data.
The traditional neural network structure ResNet is a classical residual regression model, which performs well in processing image tasks, and its input is often an image with many channels, which performs well in visual images. However, the input to the nonlinear function is often a one-dimensional vector, and the convolution is not a global, but rather a local operation. This means that if the vector is reshaped into a matrix, the convolution kernel cannot affect the whole sequence, but only a part of the sequence. This is contrary to the goal of nonlinear regression. The invention focuses on solving the nonlinear regression problem, which requires more complex replacement operations for the traditional convolution layer and pooling layer, and adopts a full-connection layer to replace other blocks except for the activation function and the batch normalization layer in the convolution layer and pooling layer.
The convolution layer replaced by the full-connection layer is shown in fig. 4, and the pooling layer replaced by the full-connection layer is shown in fig. 5, and comprises the full-connection layer, the batch processing layer and the ReLU activation function. The convolution layer and the pooling layer are replaced by the full-connection layer, so that input data can be unfolded into a vector, and the input data can be subjected to linear transformation by using a weight matrix and an offset vector. The reserved batch normalization layer can also play a role as a regularization device, so that the requirement for regularization is reduced, and meanwhile, the initialization problem does not need to be concerned too much.
In a ResNet network, a convolution layer and a pooling layer play important roles, have the capability of extracting spatial features in input data, and effectively reduce the number of parameters in the network in a weight sharing mode, so that training efficiency and generalization performance are improved. The weight sharing mechanism enables the network to learn the local mode and global information of the input data and obtain better effects on specific tasks.
The invention processes the nonlinear regression problem by introducing the full connection layer and combines the reserved batch normalization method to improve the performance and the robustness of the model. The improved method for the ResNet network can better capture the spatial characteristics when processing the image data, and reduce the parameter quantity in a weight sharing mode, so that the efficiency is improved and better generalization performance is obtained in the training process.
As shown in fig. 6, the multi-headed attention mechanism allows the network to focus on different parts of the input data and process those parts by representing more complex functions than a simple weighted average. Given the same query, key, and value set, the deep learning model can learn different behaviors using the same attentive mechanisms and combine these different behaviors as knowledge. Such as capturing the dependency of various ranges within the sequence. Thus, it may be beneficial to allow the attention mechanism to use different subspace representations of queries, keys, and values in combination.
In practice, by applying multiple attention heads to the input data, each head will produce an attention pooled output. These attention-pooled outputs will be stitched together to form a larger output vector. To further process this stitched vector, another linear projection that can be learned is used for transformation. This linear projection will parameterize the stitched vector and produce the final output.
By introducing a multi-head attention mechanism and a linear projection transformation, the deep learning model is able to more fully capture the relevant information of the input data and represent different behaviors by learning different attentions. The final output will incorporate information from multiple heads, providing a richer, more accurate expression capability.
Step S4: aiming at the inversion problem of the resistivity of the electromagnetic waves in the azimuth while drilling, setting different depths and widths of a deep learning model; respectively inputting training sets into the deep learning models for training, and adjusting the super parameters of the deep learning models by using a Bayesian optimization parameter adjustment algorithm until the super parameters of the optimal deep learning models are determined, so as to complete the training of the deep learning models with different depths and widths; and respectively inputting the verification sets into the deep learning models, and selecting the depth and width deep learning models with minimum training loss and minimum verification loss as the optimal deep learning model for inversion of the electromagnetic wave resistivity of the azimuth while drilling.
In order to find the optimal super-parameter combination, a Bayesian optimization parameter adjustment algorithm is used, a Gaussian process method is adopted, and the prior parameter information is used for continuously updating the prior. The algorithm can quickly obtain results with a small number of iterations. Meanwhile, the algorithm is also suitable for the non-convex problem, and can avoid sinking into local optimum. Firstly defining an objective function, and then minimizing the objective function by using a Bayesian super-parameter optimization algorithm, so as to obtain the optimal super-parameter configuration of the electromagnetic wave resistivity inversion model of the azimuth while drilling.
In bayesian superparameter optimization, each superparameter is treated as a random variable whose a priori distribution is a gaussian distribution. The posterior distribution of these super parameters is updated using the electromagnetic wave resistivity logging data of the azimuth while drilling, resulting in a more accurate super parameter estimate. Finally, sampling some super-parameter configurations from the joint posterior distribution by using a sampling method, calculating values of the objective function on the configurations, and finding out a minimum point. In the examples we set the learning rate lr to be in the range of 0.0001 to 1, the batch size to be in the range of 128 to 512, and the epoch to be in the range of 50 to 100. After 100 iterations we get the parameter combination with the best evaluation index: the learning rate lr was 0.0003, batch size 468, epoch 92, and mean absolute error MAE 0.13043. These parameters are tuned to enable the model to exhibit optimal performance for a given task.
When focusing on the influence of depth and width of the deep learning model, consideration is performed by training residual regression models of different widths multiple times. It was found experimentally that both training loss and validation loss exhibited very small values when trained using a model of fixed width 36. Thus, based on previous experience, it was decided to initially set the width of each hidden layer to 36 and verify this selection by a series of comparative experiments. The method is characterized by comprising the following steps of (1) carrying out an experiment 1-a neural network depth parameter evaluation experiment: when the neural network depth is 10, 16, 28, 32, 46, 82, 100, 128, 256, 512, the corresponding early-stop epochs are 60, 44, 32, 37, 28, 25, 31, 26 respectivelyRespectively training time of 2h30min, 3h02min, 4h31min, 5h01min, 6h11min, 6h58min, 8h03min, 8h58min, 10h13min, 12h36min (10 -3 ) For 8.4561, 7.9982, 5.9785, 6.3965, 7.9312, 9.8091, 10.6924, 13.9311, 265.0108, 93.5367, respectively, corresponding validation losses (10 -3 ) 5.6972, 6.8958, 5.0348, 4.6978, 8.2676, 13.0932, 10.8426, 19.0, 1926.3720, 107.4625.
Based on the selection of the optimal depth, the influence of the width on the model is taken into account. The method is characterized by comprising the following steps of (1) carrying out an experiment 2-a neural network width parameter evaluation experiment: when the neural network width is 1, 4, 8, 16, 20, 24, 30, 50, 80, 100, the corresponding early-stop epochs are 100, 80, 67, 54, 38, 47, 76, 89, respectively, the corresponding training time is 24h42min, 11h35min, 10h11min, 6h02min, 6h55min, 4h56min, 3h59min, 5h11min, 7h19min, 8h31min, respectively, the corresponding training loss (10 -3 ) For 266.2759, 17.6886, 8.7851, 3.7678, 4.8671, 1.4943, 1.4724, 1.8368, 3.6952, 7.5337, respectively, corresponding validation losses (10 -3 ) 269.3576, 17.4628, 9.4620, 3.6868, 4.0725, 1.4701, 1.0257, 1.7680, 6.3965, 6.8163.
Experiment 1 is the effect of neural network depth on model performance. By experimentation, we found that when the depth of the neural network was less than 50, the test loss and validation loss varied less. And when the depth is equal to or exceeds 100, the loss of training and validation increases greatly. Based on these observations, we selected the optimal depth to be 32. On the basis of determining the optimal depth, we continue to consider the effect of width on model performance. We set the depth to 32 and do several exercises. Experiment 2 shows that when the width is less than 8, the loss of training and verification is larger, while the loss of training and verification at other values is similar. Therefore, we choose the optimal width to be 30.
Step S5: and inputting the received electromagnetic wave logging data into an optimal deep learning model, and outputting the resistivity and interface position information of the current stratum by the deep learning model to finally obtain the resistivity imaging of the stratum.
Claims (6)
1. The method for inverting the resistivity of the electromagnetic wave in the azimuth while drilling based on the deep learning model is characterized by comprising the following steps of:
step S1: establishing a forward model of the electromagnetic wave resistivity of the azimuth while drilling, setting a stratum model to be a three-layer stratum model, setting stratum model parameters required in the forward process according to inversion parameters, and acquiring the electromagnetic wave resistivity while drilling by adopting a one-dimensional forward algorithm; the three-layer stratum model comprises an upper stratum, a middle layer and a lower stratum, and an azimuth electromagnetic wave resistivity logging instrument while drilling is arranged on the middle layer; the logging instrument comprises different types of transmitting coils and receiving coils, and different geological signals and resistivity signals are generated by setting different source distances and coil combinations;
step S2: preprocessing electromagnetic wave resistivity data of the azimuth while drilling and instrument real measurement data obtained by a one-dimensional positive algorithm to obtain data of a training set and a verification set;
step S3: based on a ResNet network, the convolution layer and the rest blocks except the activation function and the batch normalization layer in the pooling layer in the ResNet network are replaced by full-connection layers, and feature extraction is carried out on input data by combining a multi-head attention mechanism, so that the construction of a deep learning model for the inversion of the azimuth electromagnetic wave resistivity while drilling is completed;
step S4: aiming at the inversion problem of the resistivity of the electromagnetic waves in the azimuth while drilling, setting different depths and widths of a deep learning model; respectively inputting training sets into the deep learning models for training, and adjusting the super parameters of the deep learning models by using a Bayesian optimization parameter adjustment algorithm until the super parameters of the optimal deep learning models are determined, so as to complete the training of the deep learning models with different depths and widths; respectively inputting the verification sets into a deep learning model, and selecting the depth and width deep learning model with minimum training loss and minimum verification loss as the optimal deep learning model for inversion of the electromagnetic wave resistivity of the azimuth while drilling;
step S5: and inputting the received electromagnetic wave logging data into an optimal deep learning model, and outputting the resistivity and interface position information of the current stratum by the deep learning model to finally obtain the resistivity imaging of the stratum.
2. The method of claim 1, wherein the model parameters of the three-layer formation model comprise distances of the electromagnetic resistivity logging instrument from the upper and lower interfaces while drilling, and the resistivities of the upper, middle, and lower formations comprise horizontal and vertical resistivities, and the model parameters are continuously adjusted to calculate instrument responses under different formation models.
3. The method of claim 1, wherein the resistivity of each stratum and the thickness of each layer are limited to be within a preset range when a forward model of the electromagnetic wave resistivity of the azimuth while drilling is established; the resistivity of each stratum is in the range of 0.2 to 2000 Ω -m, and the thickness of each stratum is in the range of 0.5 to 15 m.
4. The method of claim 1, wherein the preprocessing of the electromagnetic wave resistivity data of the azimuth while drilling obtained by the one-dimensional forward algorithm in step S2 includes a sample outlier removal process, a logarithmic process, and a normalization process;
the preprocessing of the measured data of the logging instrument of the electromagnetic wave resistivity of the azimuth while drilling comprises the logarithmic processing and the normalization processing.
5. The method of claim 1, wherein the output of each head of the multi-head attention mechanism is stitched in the deep learning model in step S3 and then subjected to linear projective transformation.
6. The method of claim 1, wherein in step S4, a deep learning model with a width of 36 and a depth of 32 is selected as the best deep learning model for electromagnetic wave resistivity inversion in azimuth while drilling.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202410187481.5A CN117744505B (en) | 2024-02-20 | 2024-02-20 | Depth learning-based inversion method for electromagnetic wave resistivity of azimuth while drilling |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202410187481.5A CN117744505B (en) | 2024-02-20 | 2024-02-20 | Depth learning-based inversion method for electromagnetic wave resistivity of azimuth while drilling |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN117744505A true CN117744505A (en) | 2024-03-22 |
| CN117744505B CN117744505B (en) | 2024-04-26 |
Family
ID=90261287
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202410187481.5A Active CN117744505B (en) | 2024-02-20 | 2024-02-20 | Depth learning-based inversion method for electromagnetic wave resistivity of azimuth while drilling |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN117744505B (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN118171582A (en) * | 2024-05-11 | 2024-06-11 | 中国石油大学(华东) | Electromagnetic wave logging inversion method and system for azimuth while drilling by combining residual neural network and L-M algorithm |
Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110968826A (en) * | 2019-10-11 | 2020-04-07 | 重庆大学 | Magnetotelluric deep neural network inversion method based on spatial mapping technology |
| WO2020257263A1 (en) * | 2019-06-18 | 2020-12-24 | University Of Houston System | Systems and methods for solving geosteering inverse problems in downhole environments using a deep neural network |
| US20210256671A1 (en) * | 2020-02-17 | 2021-08-19 | Halliburton Energy Services, Inc. | Borehole image blending through machine learning |
| CN114386412A (en) * | 2020-10-22 | 2022-04-22 | 四川大学 | Multi-modal named entity recognition method based on uncertainty perception |
| US20230041525A1 (en) * | 2020-01-25 | 2023-02-09 | Schlumberger Technology Corporation | Automatic model selection through machine learning |
| 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 |
-
2024
- 2024-02-20 CN CN202410187481.5A patent/CN117744505B/en active Active
Patent Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2020257263A1 (en) * | 2019-06-18 | 2020-12-24 | University Of Houston System | Systems and methods for solving geosteering inverse problems in downhole environments using a deep neural network |
| CN110968826A (en) * | 2019-10-11 | 2020-04-07 | 重庆大学 | Magnetotelluric deep neural network inversion method based on spatial mapping technology |
| US20230041525A1 (en) * | 2020-01-25 | 2023-02-09 | Schlumberger Technology Corporation | Automatic model selection through machine learning |
| US20210256671A1 (en) * | 2020-02-17 | 2021-08-19 | Halliburton Energy Services, Inc. | Borehole image blending through machine learning |
| CN114386412A (en) * | 2020-10-22 | 2022-04-22 | 四川大学 | Multi-modal named entity recognition method based on uncertainty perception |
| 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 |
Non-Patent Citations (3)
| Title |
|---|
| KYUBO NOH等: "Physics-Guided Deep-Learning Inversion Method for the Interpretation of Noisy Logging-While-Drilling Resistivity Measurements", GEOPHYSICAL JOURNAL INTERNATIONAL, vol. 235, no. 1, 1 June 2023 (2023-06-01), pages 150 - 165 * |
| 史晓锋, 李铮, 蔡志权: "随钻电磁波传播电阻率测井的快速反演新方法", 测井技术, vol. 25, no. 04, 22 August 2001 (2001-08-22), pages 266 - 269 * |
| 岳喜洲;马明学;李国玉;刘小刚;刘天淋;谢涛: "随钻方位电磁波电阻率测井技术与地质导向应用", 测井技术, vol. 65, no. 002, 31 December 2021 (2021-12-31), pages 1909 - 1920 * |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN118171582A (en) * | 2024-05-11 | 2024-06-11 | 中国石油大学(华东) | Electromagnetic wave logging inversion method and system for azimuth while drilling by combining residual neural network and L-M algorithm |
Also Published As
| Publication number | Publication date |
|---|---|
| CN117744505B (en) | 2024-04-26 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Colombo et al. | Physics-driven deep-learning inversion with application to transient electromagnetics | |
| CN109709603B (en) | Seismic horizon identification and tracking method and system | |
| Lopez-Alvis et al. | Deep generative models in inversion: The impact of the generator's nonlinearity and development of a new approach based on a variational autoencoder | |
| Aleardi et al. | 1D elastic full‐waveform inversion and uncertainty estimation by means of a hybrid genetic algorithm–Gibbs sampler approach | |
| CN108875122B (en) | Artificial intelligence method and system for calculating geological parameters by using logging-while-drilling data | |
| US20150153476A1 (en) | Method for constrained history matching coupled with optimization | |
| CN109799533A (en) | A kind of method for predicting reservoir based on bidirectional circulating neural network | |
| CN117744505B (en) | Depth learning-based inversion method for electromagnetic wave resistivity of azimuth while drilling | |
| Linde et al. | Distributed soil moisture from crosshole ground-penetrating radar travel times using stochastic inversion | |
| Shahriari et al. | Error control and loss functions for the deep learning inversion of borehole resistivity measurements | |
| Wang et al. | Bayesian inversion of logging-while-drilling extra-deep directional resistivity measurements using parallel tempering Markov chain Monte Carlo sampling | |
| Jin et al. | A physics-driven deep-learning network for solving nonlinear inverse problems | |
| Guo et al. | Application of supervised descent method to transient electromagnetic data inversion | |
| Shen et al. | Solving geosteering inverse problems by stochastic Hybrid Monte Carlo method | |
| Chen et al. | Integration of principal-component-analysis and streamline information for the history matching of channelized reservoirs | |
| CN115047531A (en) | Transient electromagnetic data inversion method based on ELM network | |
| Korjani et al. | Reservoir characterization using fuzzy kriging and deep learning neural networks | |
| CN116992754A (en) | Rapid inversion method for logging while drilling data based on transfer learning | |
| Grana et al. | Uncertainty quantification in seismic facies inversion | |
| Zhang et al. | Efficient surrogate modeling based on improved vision transformer neural network for history matching | |
| CN104316961A (en) | Method for obtaining geological parameters of weathered layer | |
| You et al. | Explainable convolutional neural networks driven knowledge mining for seismic facies classification | |
| Chen et al. | Integration of cumulative-distribution-function mapping with principal-component analysis for the history matching of channelized reservoirs | |
| Shen et al. | Data-driven interpretation of ultradeep azimuthal propagation resistivity measurements: Transdimensional stochastic inversion and uncertainty quantification | |
| Roberts et al. | Joint stochastic constraint of a large data set from a salt dome |
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 | ||
| GR01 | Patent grant |