A method for inversion of formation parameters based on deep neural network for logging while drilling

Abstract

The present invention discloses a method for inverting formation parameters of LWD exploration based on deep neural network. The training phase includes the steps of constructing an azimuth electromagnetic wave exploration model of LWD, constructing a data set, preprocessing data, training a resistivity inversion neural network and training a formation boundary inversion neural network. In the testing phase, the received logging response is respectively input into the trained resistivity inversion neural network and the boundary inversion neural network. The resistivity inversion neural network outputs the predicted resistivity to the boundary inversion neural network. The boundary inversion neural network outputs the predicted formation boundary position. Finally, a formation image is drawn together according to the predicted resistivity and the formation boundary position. The present invention constructs a formation parameter inversion system composed of a resistivity inversion neural network and a formation boundary inversion neural network based on a priori resistance, so that formation parameters can be inverted in real time with high precision during the LWD exploration process.

Description

Technical Field

[0001] The present invention belongs to logging technology, specifically to the field of while-drilling data processing technology, and in particular to a while-drilling logging forward logging formation parameter inversion method based on deep neural network, which can be used for real-time inversion of forward logging formation parameters in the while-drilling logging process. Background Art

[0002] Look-Ahead While Drilling (LAWD) is an innovative geophysical logging technology designed to help drilling engineers predict the characteristics and structure of the formation ahead in real time during the drilling process. This technology combines LWD and look-ahead technology to provide information about the formation ahead during the drilling process, helping to optimize the drilling path and improve drilling safety and efficiency. As the focus of oil and gas exploration and development shifts to deep, ultra-deep and unconventional oil and gas reservoirs, traditional logging methods often lose accuracy when dealing with complex underground formations due to simplified calculations. Methods based on deep neural networks can better cope with complex formation conditions, capture the nonlinear characteristics of formations, improve the accuracy of inversion results, and provide more reliable formation parameter inversion results, thereby enhancing the success rate of oil and gas reservoir exploration. On the other hand, traditional LWD formation parameter inversion methods are usually time-consuming and difficult to meet the needs of real-time decision-making. Once the neural network is trained, its inversion speed far exceeds that of traditional inversion methods. It can update formation information in real time and quickly respond to changes in the drilling process, thereby providing information on the formation ahead during the drilling process, helping to optimize the drilling path and improve drilling safety and efficiency. Summary of the invention

[0003] The technical problem to be solved by the present invention is to invert the resistivity and thickness of the formation in front of the logging instrument through the response of the azimuthal electromagnetic wave logging while drilling during the logging while drilling process, and propose a method for inverting the forward exploration formation parameters of the logging while drilling based on a deep neural network.

[0004] The technical solution adopted by the present invention to solve the above technical problems is to establish a resistivity inversion neural network and a formation boundary inversion neural network, and use the inverted formation conductivity as prior information in the formation boundary inversion process to improve the inversion accuracy of the formation boundary, including the following steps:

[0005] Step S1: Constructing a formation model and logging response data set. A three-layer pre-exploration formation model is established, and the three-layer formation model includes an upper layer, a middle layer, and a lower layer. The logging instrument while drilling is always placed on the uppermost layer, and the parameters of the two lower layers are inverted. A one-dimensional forward algorithm is used to calculate the azimuthal electromagnetic wave logging response while drilling, and the data set is composed of formation parameters and logging response;

[0006] Step S2: After generating enough responses, preprocess them as training sets and label data for the neural network. The preprocessing is used to speed up the training of the deep learning model, accelerate convergence, and improve efficiency.

[0007] Step S3: Construct and train the resistivity inversion neural network. Based on the deep neural network, the well logging response calculated by the one-dimensional forward algorithm is used as the data set, and the formation resistivity is used as the label to train the resistivity inversion neural network. After the resistivity inversion neural network is trained, it is debugged. By adjusting the parameters such as the network structure, learning rate, and number of training cycles of the resistivity inversion neural network, the predicted value and label error of the resistivity inversion neural network are minimized;

[0008] Step S4: Construct and train the formation boundary inversion neural network. Based on the deep neural network, the well logging response and formation resistivity calculated by the one-dimensional forward algorithm are used as data sets, and the formation boundary label is used to train the formation boundary inversion neural network; after the formation boundary inversion neural network is trained, it is debugged, and the prediction value and label error of the formation boundary inversion neural network are minimized by adjusting the parameters such as the network structure, learning rate and number of training cycles of the resistivity inversion neural network;

[0009] Step S5: Generate a simulated formation. First, input the logging response corresponding to the simulated formation into the resistivity inversion neural network to obtain the predicted resistivity; then input the logging response and the resistivity predicted by the resistivity inversion neural network into the formation boundary inversion neural network, and use the resistivity inversion neural network and the formation boundary inversion neural network for joint inversion; finally, draw a formation image based on the inverted resistivity and formation boundary.

[0010] The beneficial effect of the present invention is that, through the deep learning-based forward electromagnetic wave exploration inversion method, fast and high-precision inversion of forward electromagnetic wave logging while drilling is achieved. The logging response data and formation parameters obtained by the one-dimensional forward algorithm are used for the deep learning model training of the formation resistivity and boundary, and the resistivity is first inverted, and then the resistivity is brought into the formation boundary inversion neural network as a priori information to improve the inversion accuracy, and then the formation apparent resistivity image is drawn, providing reliable geological guidance for logging. The present invention does not rely on the prior information of the initial formation model, and the inversion speed can meet the real-time operation requirements. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] Figure 1 This is a schematic diagram of the azimuthal electromagnetic wave exploration model while drilling;

[0012] Figure 2 It is a dynamic while-drilling azimuthal electromagnetic wave exploration model;

[0013] Figure 3A resistivity inversion neural network structure diagram provided for an embodiment;

[0014] Figure 4 A resistivity inversion neural network training flow chart provided for an embodiment;

[0015] Figure 5 A structural diagram of a stratum boundary inversion neural network provided in an embodiment;

[0016] Figure 6 A flow chart of a formation boundary inversion neural network training provided in an embodiment;

[0017] Figure 7 A flow chart of the formation parameter joint inversion test provided in the embodiment;

[0018] Figure 8 Example 1 for comparing the simulated formation and the inversion result provided in the embodiment; (a) is the simulated formation, and (b) is the inversion result;

[0019] Fig. 9 Case 2 for comparing the simulated formation and the inversion result provided in the embodiment; (a) is the simulated formation, and (b) is the inversion result. DETAILED DESCRIPTION

[0020] The embodiment discloses a method for inverting formation parameters of LWD exploration based on deep neural network. During LWD exploration, as the logging instrument is drilled down, the resistivity of the current formation where the logging instrument is located is directly measured and known. What needs to be inverted is the upper boundary position and conductivity information of the two formations below the layer where the logging instrument is located.

[0021] In the process of formation parameter inversion of forward detection, unlike the synchronous inversion of resistivity and formation boundary, the resistivity is inverted first, and then the inverted resistivity is added as prior information to the inversion process of formation boundary, which improves the inversion accuracy. Using the trained inversion model, the electromagnetic wave measured by the logging instrument while drilling can be inverted in real time without the need for prior formation information, so that the staff can easily and accurately distinguish the formation information in front of the logging instrument in a timely manner based on the inversion imaging.

[0022] The forward logging formation parameter inversion method based on deep neural network includes the following steps:

[0023] Step S1: Establish Figure 1 The azimuth electromagnetic wave exploration model while drilling is shown. The formation model is set to a three-layer formation model, from top to bottom, it is formation 1, formation 2 and formation 3, and the corresponding conductivity is σ 1 , σ 2 and σ 3The logging instrument on the drilling instrument is always at the top layer of the model, that is, the current formation of the logging instrument is formation 1, also called the current formation. The location of the logging instrument is the original coordinate position Z of the model. 0 =0, the resistivity of formation 1 is known, the electromagnetic wave logging response is obtained through the transceiver antenna on the logging instrument, and the conductivity and boundary position of the next layer and the next two layers of the current formation are inverted based on the electromagnetic wave logging response. The current formation is formation 1, the next layer of the current formation is formation 2, and the next layer of the current formation is formation 3. The position of the formation boundary is represented by the distance from the upper boundary of the formation to the logging instrument. What needs to be inverted is the conductivity of formation 2 and formation 3 and the upper boundary distance Z between formation 2 and formation 3 0 The upper boundary position Z of layer 3 2 The upper boundary position Z of layer 2 1 The difference can be used to obtain the thickness of layer 2, that is, the thickness of the layer below the current layer.

[0024] Figure 2 The figure shows a dynamic azimuthal electromagnetic wave exploration model while drilling. The logging instrument is omitted in the figure and replaced by a dotted line with y=0. From left to right, it can be seen as a combination of inversion results at multiple times, that is, from left to right, it represents that as time goes by, the logging instrument is getting closer and closer to the underlying formation. The dynamic azimuthal electromagnetic wave exploration model while drilling can verify the inversion results when the logging instrument is at different distances from the front formation.

[0025] Step S2: Construct a data set for training. The well logging response of a given formation parameter is calculated using a one-dimensional forward algorithm, and the well logging response and formation parameters together form a data set, wherein the one-dimensional forward algorithm is a forward calculation of the response of the while-drilling azimuthal electromagnetic wave resistivity logging instrument in a multilayer medium, specifically, in solving the Maxwell equations in anisotropic media, based on the Hertz potential function, the interface amplitude is introduced, so that the derivation process of electromagnetic wave propagation is completely converged. The formation parameters include formation resistivity and formation boundary position. The formation resistivity value range is 1 to 100 Ωּ·m, and the formation boundary position value range is 0.5 to 12 m. The data set consists of 200,000 sets of data, each set of which contains 3 formation resistivity values, 2 formation boundary positions, and 24 well logging response values ​​corresponding to the set of formation parameter values. In the resistivity inversion neural network training process, the logging response is used as training data and the resistivity is used as the label; while in the formation boundary inversion neural network training process, the logging response and resistivity are used as training data and the formation boundary position is used as the label.

[0026] Step S3: preprocess the data set. Preprocessing can make the data distribution in the data set more reasonable, enable the neural network to better fit the target, accelerate convergence, and improve training efficiency and effect. In this embodiment, the formation resistance is first taken inverse to obtain conductivity, and then the logarithm with a base of 10 is taken to reduce the difference in the order of magnitude of the data, and the resistivity range of 1 to 100 Ωּ·m is transformed into conductivity -2 to 0, and then standard normalization is performed; since the order of magnitude difference of the formation boundary position is small, the formation boundary position is directly subjected to standard normalization; and the 24 well logging response data have large order of magnitude differences and uneven distribution, so the well logging response is column normalized, and the maximum and minimum values ​​of the 24 well logging responses are normalized to make their distribution more reasonable. By performing different data preprocessing on resistivity, formation boundary position and well logging response, the data set used for neural network training can be made more reasonable in numerical distribution, the convergence process of neural network training can be accelerated, and a more reliable and efficient data basis is provided for subsequent training and analysis.

[0027] Step S4: Establish Figure 3 The resistivity inversion neural network model shown in the figure. The model has 8 layers, and its structure is as follows: input layer, first fully connected layer, first batch normalization layer, second fully connected layer, second batch normalization layer, third fully connected layer, third batch normalization, fourth fully connected layer, fourth batch normalization layer, fifth fully connected layer, fifth batch normalization layer, sixth fully connected layer, sixth batch normalization layer and output layer. The parameters of each layer are set as follows: the input layer has 24 units, corresponding to 24 well logging responses; the first and fourth fully connected layers and their batch normalization layers have 4096 units; the second, third, fifth, and sixth fully connected layers and their batch normalization layers have 2048 units; the output layer has 2 units, corresponding to the 2 formation conductivities to be inverted. The activation function used in the first fully connected layer and the output layer is the Sigmoid function, and the activation function used in all other layers is the ReLU function.

[0028] Step S5: Use the preprocessed data set to train the constructed resistivity inversion neural network. Figure 4The figure shows a schematic diagram of the resistivity inversion neural network training process. The well logging response obtained in step S2 is used as training data in the data set after preprocessing in step S3, and the resistivity is used as a label. The training data and the label are input into the resistivity inversion neural network constructed in step S4. The resistivity inversion neural network outputs the predicted resistivity as the inversion result. The resistivity inversion neural network training process is constrained by calculating the error between the predicted resistivity and the labeled resistivity. The training compilation environment is Python 3.11.5, and Pytorch is used as the deep learning framework. During the training process, the batch size is 8192. The optimizer is trained with the Adam optimizer. The loss function uses the mean square error function of the neural network predicted resistivity and the labeled resistivity value. The initial learning rate is 0.001, and the learning rate reduction factor is 0.85. When the model loss function does not decrease after 100 epochs, the learning rate is reduced to 0.85 times the current value. After training for 3000 epochs, the trained resistivity inversion neural network is obtained.

[0029] Step S6: Establish Figure 5 The boundary inversion neural network model shown in Figure 1 is a neural network model with 8 layers in total. Its structure is as follows: input layer, first fully connected layer, first batch normalization layer, second fully connected layer, second batch normalization layer, third fully connected layer, third batch normalization layer, fourth fully connected layer, fourth batch normalization layer, fifth fully connected layer, fifth batch normalization layer, sixth fully connected layer, sixth batch normalization layer and output layer. The parameters of each layer are set as follows: the input layer has 27 units, corresponding to 24 well logging responses and 3 formation resistivities; the first and fourth fully connected layers and their batch normalization layers have 4096 units; the second, third and fifth fully connected layers and their batch normalization layers have 2048 units; the sixth fully connected layer and its batch normalization layer have 1024 units; the output layer has 2 units, corresponding to the 2 formation boundary positions to be inverted. The activation function used in the first and fourth fully connected layers is the Sigmoid function, and the activation function used in all other layers is the ReLU function.

[0030] Step S7: Figure 6The figure shows a schematic diagram of the boundary inversion neural network training process. The well logging response and resistivity in the preprocessed data set are used as training data, and the formation boundary position is input as a label into the boundary inversion neural network built in step S6. The boundary inversion neural network outputs the predicted formation boundary position as the inversion result. The boundary inversion neural network training process is constrained by calculating the error between the predicted formation boundary position and the labeled formation boundary position. The training compilation environment is Python3.11.5, and Pytorch is used as the deep learning framework. The loss function uses the mean square error function of the formation boundary and the label boundary value predicted by the neural network. During the training process, the initial learning rate is 0.002, and the learning rate reduction factor is 0.9. When the model loss function does not decrease after 100 epochs, the learning rate is reduced to 0.9 times the current value. A total of 3000 epochs were trained, the batch size is 8192, and the optimizer uses the Adam optimizer for training.

[0031] Step S8: Testing step. After the resistivity inversion neural network and the boundary inversion neural network are trained, the joint inversion process in actual use is as follows: Figure 7 As shown in the figure, the logging instrument first inputs the currently received logging response into the trained resistivity inversion neural network, and the resistivity inversion neural network outputs the predicted resistivity; then the received logging response and the predicted resistivity are input into the boundary inversion neural network, and the boundary inversion neural network outputs the predicted formation boundary position; finally, the formation image is drawn together based on the predicted resistivity and formation boundary position. In order to intuitively verify the effect of the trained inversion neural network, we constructed the following Figure 8 (a) and Fig. 9 The simulation model of (a) is shown in Figure 1. Figure 7 The steps are combined for inversion. The inverted forward model is as follows Figure 8 (b) and Fig. 9 As shown in (b), it can be seen that for the simulated formation, the inverted exploration model is basically consistent with the given simulated formation. In the actual application process, the received electromagnetic wave logging response is input into the trained physical driven neural network, and the deep learning model will output the resistivity and interface position information of the current formation, and finally obtain the resistivity imaging of the formation.

Claims

1. A method for inversion of formation parameters based on deep neural network for logging while drilling, characterized in that: include: Training phase: The steps of constructing the electromagnetic wave prospecting model while drilling are as follows: the current formation where the logging instrument is located is set to be the uppermost layer of the three formations in the prospecting model; the logging response is obtained through the electromagnetic wave transceiver antenna on the logging instrument, and the formation parameters of the next layer and the next two layers of the current formation are inverted according to the logging response; the formation parameters include resistivity and boundary position; Dataset construction steps: firstly, given the formation parameters, the logging response is obtained according to the formation parameters; The data set consists of several groups of data, each of which contains the resistivity of three layers, the position of the boundary between the next layer and the next two layers of the current layer, and several corresponding well logging responses; Data preprocessing steps: normalize the data set to make the data evenly distributed; Resistivity inversion neural network training steps: the well logging response in the preprocessed data set is used as training data, and the resistivity is used as a label, which is input into the resistivity inversion neural network; the resistivity inversion neural network outputs the predicted resistivity as the inversion result; the training process of the resistivity inversion neural network is constrained by calculating the error between the predicted resistivity and the resistivity used as the label; Formation boundary inversion neural network training steps: the well logging response and resistivity in the preprocessed data set are used as training data, and the formation boundary position is used as a label, which is input into the boundary inversion neural network; the boundary inversion neural network outputs the predicted formation boundary position as the inversion result; the boundary inversion neural network training process is constrained by calculating the error between the predicted formation boundary position and the formation boundary position as a label; Testing phase: The received logging response is input into the resistivity inversion neural network to obtain the predicted resistivity; the logging response and the resistivity predicted by the resistivity inversion neural network are then input into the formation boundary inversion neural network together, and the resistivity inversion neural network and the formation boundary inversion neural network are used for joint inversion; finally, a formation image is drawn based on the inverted resistivity and formation boundary.

2. The method according to claim 1, characterized in that: The formation boundary position is the distance from the upper boundary of the formation to the logging instrument.

3. The method according to claim 1, characterized in that: In the data set construction step, a one-dimensional forward algorithm is used to calculate the logging response of given formation parameters.

4. The method according to claim 1, characterized in that: The specific method for normalizing the data set in data preprocessing is: First take the reciprocal of the resistance, then take the logarithm of the reciprocal with base 10, and the obtained logarithmic result is the normalized result of the resistivity; The result of the standard normalization processing of the formation boundary position is used as the normalization processing result of the formation boundary position; The results of normalizing the maximum and minimum values ​​of the columns of the multiple logging responses are used as the normalization results of the multiple logging responses.

5. The method according to claim 1, characterized in that: The resistivity inversion neural network includes an input layer, a first fully connected layer, a first normalization layer, a second fully connected layer, a second normalization layer, a third fully connected layer, a third normalization layer, a fourth fully connected layer, a fourth normalization layer, a fifth fully connected layer, a fifth normalization layer, a sixth fully connected layer, a sixth normalization layer and an output layer in sequence; The boundary inversion neural network includes, in sequence, an input layer, a first fully connected layer, a first batch of normalized layers, a second fully connected layer, a second batch of normalized layers, a third fully connected layer, a third batch of normalized layers, a fourth fully connected layer, a fourth batch of normalized layers, a fifth fully connected layer, a fifth batch of normalized layers, a sixth fully connected layer, a sixth batch of normalized layers, and an output layer.

6. The method according to claim 5, characterized in that There are N logging responses corresponding to the formation parameter values ​​in each set of data in the data set, where N is an integer.

7. The method according to claim 6, characterized in that: The parameters of each layer in the resistivity inversion neural network are set as follows: the input layer has N units, corresponding to N logging responses; the first and fourth fully connected layers and their batch normalization layers have 4096 units; the second, third, fifth, and sixth fully connected layers and their batch normalization layers have 2048 units; the output layer has 2 units, corresponding to the two formation conductivities to be inverted; the activation function used in the first fully connected layer and the output layer is the Sigmoid function, and the activation function used in all other layers is the ReLU function.

8. The method according to claim 6, characterized in that: The parameters of each layer of the boundary inversion neural network are set as follows: the input layer is N+3 units, corresponding to N well logging responses and 3 formation resistivities; the first and fourth fully connected layers and their batch normalization layers are 4096 units; the second, third, and fifth fully connected layers and their batch normalization layers are 2048 units; the sixth fully connected layer and its batch normalization layer are 1024 units; the output layer is 2 units, corresponding to the 2 formation boundary positions to be inverted; The activation function used in the first and fourth fully connected layers is the Sigmoid function, and the activation function used in all other layers is the ReLU function.

Images