Stratum parameter inversion method based on physically-driven neural network

Abstract

The invention discloses a stratum parameter inversion method based on a physical driving neural network, and mainly solves the problems of poor inversion precision and low anti-noise capability of stratum parameters of a real well in an existing neural network inversion method. The method comprises the following specific steps: (1) constructing a stratum model and a logging response data set; (2) data preprocessing; (3) constructing a forward modeling neural network; (4) training and testing the forward modeling neural network; (5) constructing an inversion neural network based on physical driving; (6) training an inversion neural network; and (7) testing the trained inversion network. According to the invention, a deep neural network composed of a forward neural network module and a physical driving inversion module is constructed, and the problems of unsmooth boundary, low precision and weak anti-noise capability of the existing logging-while-drilling inversion are solved, so that stratum model parameters can be inversed in real time with relatively high precision.

Description

Technical Field

[0001] The invention belongs to well logging technology, and in particular to a while-drilling data processing technology. Background Art

[0002] In the process of logging while drilling (LWD), inversion of resistivity and formation boundary information is a key step in determining the properties of the formation and the distribution of oil and gas reservoirs. This technology can provide real-time resistivity data about the formation. By analyzing and inverting these data, the interface and boundary of the formation can be accurately determined, thereby inferring the properties of the formation and the distribution of oil and gas.

[0003] The core of resistivity inversion is to analyze the response characteristics of the formation interface using resistivity measurement tools. When the drill bit passes through different formations, the resistivity measurement tool will capture the resistivity differences between different formations. These differences can be converted into detailed formation boundary information through inversion algorithms to determine the changes in the formation and further infer the properties of the formation. For example, oil and gas reservoirs usually have low resistivity, so by identifying low-resistivity areas, potential oil and gas reservoirs can be effectively located. The formation boundary information obtained through inversion can better formulate drilling and completion plans to maximize oil and gas production.

[0004] Although there are some studies that have applied deep neural networks to the inversion of formation parameters in the logging process, the inversion method based on deep learning can quickly and accurately invert the logging response of while-drilling electromagnetic wave logging instruments in layered formations.

[0005] For example, reference 1: Shahriari M, Pardo D, Picón A, et al. A deep learning approach to the inversion of borehole resistivity measurements[J]. Computational Geosciences, 2020, 24: 971-994. This method applies deep neural networks to borehole resistivity inversion, focusing on the selection of loss function and error control, and proves the feasibility of using this method for resistivity inversion.

[0006] For example, reference 2: 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 deep learning inversion method based on BiLSTM was studied, which can evaluate the uncertainty of inversion parameters while inverting formation parameters.

[0007] For example, Reference 3: Zhu Gaoyang. Forward and inversion study of electromagnetic wave logging data while drilling in layered reservoirs based on deep learning [D]. Shandong University, 2020. A method based on deep learning DL is proposed to solve the time-consuming problem of inversion of induction logging data in layered anisotropic formations. And it has achieved good results in simulation data.

[0008] The deep learning-based azimuthal electromagnetic wave resistivity inversion method for drilling with application number 202410187481.5 sets the formation model as a three-layer formation model, sets the parameters required in the forward modeling process according to the inversion parameters, and adopts a one-dimensional forward modeling algorithm to obtain the electromagnetic wave resistivity data while drilling; the three-layer formation model includes an upper interface, an interlayer and a lower interface, and the azimuthal electromagnetic wave resistivity instrument while drilling is placed in the interlayer; the response data obtained by the one-dimensional forward modeling method based on the full tensor of the electromagnetic field of a multi-layer anisotropic medium is used as the electromagnetic wave resistivity data while drilling, and then the electromagnetic wave resistivity data while drilling is used together with the real data to train a deep learning model for azimuthal electromagnetic wave resistivity inversion while drilling, so that the best deep learning model after training and selection can use the received electromagnetic wave signal to output the resistivity data of the three-layer formation, and finally obtain the resistivity imaging of the three-layer formation model, which provides a reliable numerical calculation basis for drilling detection. Among them, the one-dimensional forward algorithm is the forward calculation of the response of the while-drilling azimuthal electromagnetic wave instrument in multi-layer media. Specifically, in solving the Maxwell equations in anisotropic media, based on the Hertz potential function, the interface amplitude is introduced to make the derivation process of electromagnetic wave propagation fully convergent, thereby enhancing the accuracy and interpretability of the inversion neural network. Summary of the invention

[0009] The technical problem to be solved by the present invention is to provide a formation parameter inversion method that allows the inversion neural network to comply with the physical law restrictions satisfied in the forward modeling process during the training process in order to further improve the inversion accuracy.

[0010] The technical solution adopted by the present invention to solve the above problems is a formation parameter inversion method based on a physical driven neural network, comprising the steps of:

[0011] The calculation steps of the one-dimensional forward algorithm are as follows: constructing a formation model, placing the logging while drilling instrument in the formation, inputting the real formation parameters into the one-dimensional forward algorithm, and the one-dimensional forward algorithm outputs the logging response value;

[0012] Forward neural network training steps: The forward neural network is based on a deep neural network. The real formation parameter values ​​are used as the training set. The logging response values ​​output by the one-dimensional forward algorithm are used as the first forward modeling labels, and the predicted formation parameters are output as the first forward modeling results. The error between the first forward modeling label and the first forward modeling result is used as a reference to constrain the training process.

[0013] Inversion neural network training steps: The inversion neural network is based on a deep neural network. The logging response values ​​calculated by the one-dimensional forward modeling algorithm are used as the training set. When training the forward neural network, the real formation parameters of the training set are used as the inversion labels, and the predicted formation parameters are output as the inversion results. The error between the inversion label and the inversion result is used as the inversion error. The trained forward neural network is then called, and the inversion result is used as the training set. The logging response values ​​output by the one-dimensional forward modeling algorithm are used as the second forward modeling label, and the predicted logging response values ​​are output as the second forward modeling result. The error between the second forward modeling label and the second forward modeling result is used as the second forward modeling error. The inversion error is combined with the second forward modeling error to calculate the total error, and the total error is used to constrain the training process of the inversion neural network.

[0014] Formation parameter inversion step: input the logging response value while drilling into the trained inversion neural network, and the inversion neural network outputs the predicted formation parameters to complete the inversion.

[0015] The present invention first uses the logging response data obtained by the one-dimensional forward algorithm for deep learning model training for electromagnetic wave formation parameter inversion while drilling, so that the best deep learning model after training and debugging can use the received electromagnetic wave logging data to invert the resistivity and formation boundary position information of the current formation, thereby obtaining the resistivity imaging of the formation and 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.

[0016] In particular, during the inversion neural network training process, in addition to using the logging response calculated by forward modeling as the training set, the trained forward neural network model is added to perform a second forward modeling on the inversion prediction value. The error of the second forward modeling output is used together with the inversion error to obtain the total error to strengthen the inversion training. The second forward modeling is added to the inversion training, and the forward modeling process satisfies certain physical laws. This makes the inversion network conform to the limitations of certain physical laws during the training process, rather than simply fitting the data, thereby enhancing the accuracy and interpretability of the inversion neural network.

[0017] Moreover, when the formation parameters predicted by the neural network are forward modeled for the second time, the forward modeling neural network is called instead of the one-dimensional forward modeling algorithm, that is, the forward modeling neural network is used instead of the physical drive. This process speeds up the training speed and introduces certain errors, which enhances the anti-noise performance of the inversion neural network.

[0018] The beneficial effect of the present invention is that, through the formation parameter inversion method based on the physical driven neural network, the rapid and high-precision inversion of the while-drilling azimuthal electromagnetic wave logging response is achieved. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 A forward neural network structure diagram provided for an embodiment;

[0020] Figure 2 A forward neural network training flow chart provided for an embodiment;

[0021] Figure 3 Comparison of the forward neural network prediction results and label values ​​provided in the embodiment;

[0022] Figure 4 A diagram of a physically driven inversion neural network structure provided for an embodiment;

[0023] Figure 5 A flow chart of physically driven inversion neural network training provided in an embodiment;

[0024] Figure 6 Case 1 for comparison between the inversion results and simulated formations provided in the embodiment; Figure 6 (a) is the inversion result of a simple three-layer formation model. Figure 6 (b) in the figure is the simulated formation inversion result;

[0025] Figure 7 Case 2 for comparison between the inversion results and simulated formations provided in the embodiment; Figure 7 (a) is the inversion result of the complex three-layer formation model. Figure 7 (b) in the figure is the simulated formation inversion result;

[0026] Figure 8Comparison of the inversion results of the real well logging response by the supervised descent method provided in the embodiment and the physical driven neural network provided in this embodiment; Figure 8 (a) is the formation inversion result obtained by supervised descent method; Figure 8 (b) in the figure is the inversion result of the neural network of the present invention. DETAILED DESCRIPTION

[0027] The inversion of formation parameters is finally completed using the trained inversion neural network. The inversion neural network is used to receive the real logging response value and output the predicted formation parameters, which are used for resistivity imaging of the formation and provide reliable geological guidance for logging.

[0028] The embodiment completes the training of the forward neural network before training the inversion neural network. The one-dimensional forward algorithm is calculated before the forward neural network training process. The input of the one-dimensional forward algorithm is the real formation parameters. Based on the Hertz potential function, the interface amplitude is introduced, and the Maxwell equations are solved, and the output is the calculated logging response value. In the forward neural network training process, the real formation parameter values ​​are used as the training set, the logging response values ​​output by the one-dimensional forward algorithm are used as labels, and the predicted formation parameters are output as the first forward result.

[0029] In the embodiment, when the inversion neural network is trained, the inversion neural network is used to receive the logging response value output by the one-dimensional forward algorithm and output the predicted formation parameters as the inversion result. The real formation parameters used as the training set when training the forward neural network are used as labels, and the inversion neural network outputs the predicted formation parameters as predicted values ​​to calculate the inversion error. In particular, during the inversion training, it is also necessary to call the forward neural network that has completed the training, and the forward neural network participating in the inversion training process is used to receive the predicted formation parameters output by the inversion neural network and output the predicted logging response value as the second forward result. According to the logging response value output by the one-dimensional forward algorithm as a label, the predicted logging response value output by the forward neural network is used as a predicted value to calculate the second forward error. During the inversion training, the inversion error is combined with the second forward error to generate a total error, and the total error is used to constrain the training process of the inversion neural network. The trained inversion neural network can realize the formation parameter inversion of the drilling azimuth electromagnetic wave based on the physical driven neural network.

[0030] In the process of inversion neural network training, physical restrictions of forward modeling rules are added to increase the anti-noise performance and smoothness of the inversion neural network. The training set of logging response is obtained by one-dimensional forward modeling method, and the forward model is obtained by forward neural network training for calling in the physical drive network. In the inversion process, in addition to the mean square error term between the conventional neural network prediction value and the label, the loss function of the neural network also adds the mean square error term between the response obtained by calling the forward network with the inversion prediction value and the actual logging response, thereby making certain physical restrictions on the neural network so that it can meet certain physical laws. Since there is a slight error between the predicted response obtained by calling the forward network and the calculated logging response, it is equivalent to adding certain noise in the training process, which further enhances the anti-noise ability of the trained inversion network. Using the trained inversion model, the formation parameters of the drilling azimuth electromagnetic waves measured by the logging instrument can be inverted in real time, and no prior formation information is required, so that the staff can easily obtain formation information accurately and timely based on the inversion imaging.

[0031] The specific steps include:

[0032] Step S1: Establish a forward model of electromagnetic wave resistivity while drilling. Set the formation model to a three-layer formation model, which includes an upper layer, a middle layer and a lower layer. The logging while drilling instrument is placed in the middle layer. A one-dimensional forward algorithm is used to receive the input of real formation parameters, calculate the electromagnetic wave logging response while drilling and output it. If the resistivity of two adjacent layers is the same, the formation model can be simplified to a two-layer model, including an upper layer and a lower layer. At this time, the logging while drilling instrument can be placed in the upper layer or the lower layer.

[0033] In the embodiment, the logging instrument is in the middle layer, the formation resistivity ranges from 1 to 100 Ωּ·m, and the formation thickness ranges from 0.5 to 12 m. The one-dimensional forward algorithm is used to calculate the logging response of the given formation parameters, and the logging response and the formation parameters together form a training 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, so as to eliminate the numerical overflow caused by the exponential increase term that may appear in the engineering calculation, thereby realizing the fast and convergent forward simulation of the while-drilling azimuthal electromagnetic wave logging response. The final training set consists of 400,000 sets of data, each set of which contains 3 formation resistivity values, 2 boundary values, and 8 logging response values ​​corresponding to the set of formation parameter values. After generating enough responses (400,000 groups), they are preprocessed as training sets and label data for the neural network. Preprocessing is used to speed up the training of deep learning models, accelerate convergence, and improve efficiency.

[0034] Step S2: Perform data preprocessing on the training set. Preprocessing can make the data distribution in the training 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 inversely and then the logarithm with base 10 is taken to reduce the difference in the order of magnitude of the data, and then standard normalization is performed; since the order of magnitude difference of the formation thickness data is small, the formation thickness is directly subjected to standard normalization; and the order of magnitude difference of the 8 well logging response data is large and unevenly distributed, so the 8 columns of well logging responses are subjected to column normalization, and normalized according to the maximum and minimum values ​​of the 8 columns of well logging responses to make their distribution more reasonable. By performing different data preprocessing on resistivity, formation thickness 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.

[0035] Step S3: Establish Figure 1 The forward neural network model shown in FIG. The forward neural network model has 8 layers, and its structure is as follows: input layer, first fully connected layer, first batch normalization, second fully connected layer, second batch normalization, third fully connected layer, third batch normalization, fourth fully connected layer, fourth batch normalization, fifth fully connected layer, fifth batch normalization, sixth fully connected layer, sixth batch normalization and output layer.

[0036] The parameters of each layer are set as follows: the input layer has 5 units; the first, second and fourth fully connected layers, the first, second and fourth batch normalization layers have 4096 units; the third and fifth fully connected layers, the third and fifth batch normalization layers have 2048 units; the sixth fully connected layer and the sixth batch normalization layer have 1024 units; the output layer has 8 units. The activation function used in the first and fourth fully connected layers and the output layer is the Sigmoid function, and the activation function used in other layers is the ReLU function.

[0037] Step S4: Use the preprocessed training set and the constructed forward neural network model for training. Figure 2 The figure shows a schematic diagram of the forward neural network training process, in which the preprocessed training set mentioned in step S1 and step S2 is trained by the neural network model described in step S3. The training compilation environment is Python 3.11.5, and Pytorch is used as the deep learning framework. During the training process, the initial learning rate is 0.004, and the learning rate reduction factor is 0.7. When the model loss function does not decrease after 50 epochs, the learning rate is reduced to 0.7 times the current value. The batch size is 8192, and the optimizer uses the Adam optimizer for training.

[0038] After the forward neural network training is completed, the test set data is used for testing and debugging. By adjusting the network structure, learning rate, and number of training cycles of the forward neural network, the error between the predicted value of the forward neural network and the response of the azimuth electromagnetic wave logging while drilling calculated by the one-dimensional forward algorithm is minimized. After the training is completed, the scatter points of the eight logging responses calculated by the one-dimensional forward method and the eight logging responses predicted by the corresponding forward neural network are plotted, and the average error is about 0.06%, as shown in Figure 2. Figure 3 Two examples are shown. The dotted line in the figure is the logging response calculated by the one-dimensional forward modeling method, and the dotted line is the logging response predicted by the forward neural network. It can be seen from the figure that the forward neural network has extremely high accuracy, with an average error of about 0.06%, which is highly consistent with the logging response calculated by the one-dimensional forward modeling method.

[0039] Step S5: Construct and train the inversion neural network. Figure 4 The inversion neural network model shown in FIG. The inversion neural network model has 8 layers, and its structure is as follows: input layer, first fully connected layer, first batch normalization, second fully connected layer, second batch normalization, third fully connected layer, third batch normalization, fourth fully connected layer, fourth batch normalization, fifth fully connected layer, fifth batch normalization, sixth fully connected layer, sixth batch normalization and output layer.

[0040] The parameters of each layer are set as follows: the input layer has 8 units; the first, second and fourth fully connected layers, the first, second and fourth batch normalization layers have 4096 units; the third and fifth fully connected layers, the third and fifth batch normalization layers have 2048 units; the sixth fully connected layer and the sixth batch normalization layer have 1024 units; the output layer has 5 units. The activation function used in the first and fourth fully connected layers and the output layer is the Sigmoid function, and the activation function used in other layers is the ReLU function.

[0041] Based on deep neural network, the logging response calculated by one-dimensional forward algorithm is used as training set, and formation parameters are used as labels to train the inversion neural network. During the training process, the loss function includes not only the mean square error between the inversion network prediction value and the label, but also the mean square error between the response of the inversion network prediction value performed by calling the trained forward neural network and the logging response calculated by the one-dimensional forward algorithm. The role of this error term is to add the physical law restrictions of the one-dimensional forward algorithm, and at the same time enhance the anti-noise ability and smoothness of the trained inversion neural network.

[0042] Step S6: Use the preprocessed training set, the trained forward network and the constructed inverse neural network model for training. Figure 5The figure shows the schematic diagram of the training process of the inversion neural network based on physical drive. First, the corresponding logging response of the formation is calculated by the one-dimensional forward modeling method and preprocessed as a training set. Then, the training set is trained by the inversion network constructed in step S5. The loss function of the training is divided into data item loss and physical losses ,in is the mean square error between the inversion network prediction value and the true formation parameters, The inversion neural network prediction value calls the mean square error between the forward neural network and the actual logging response. The inversion network passes through the total loss of the two Specifically, let the logging response be , the corresponding stratum label is , the prediction value of the neural network is , the neural network prediction data loss expression is: ,in, It is True value, It is predicted values, n is the number of samples; the physical item loss expression is: ,in The inversion network Predicted values The predicted response obtained by calling the forward neural network, ,in is the weight of the physical item loss, when If the physical term is too small, the effect will not be obvious. If it is too large, the loss function will not converge easily. Experiments have shown that A value of about 0.8 is more appropriate. Data item loss The meaning of is the fitting degree of the neural network to the inversion relationship, which is the mapping relationship from the logging response to the formation parameters through data fitting of the neural network, and The purpose is to test whether the formation parameters predicted by the neural network are consistent with the logging response obtained by the one-dimensional forward modeling algorithm after forward modeling (since the accuracy of the forward modeling neural network has been verified in step S4, the forward modeling neural network is called instead of the one-dimensional forward modeling algorithm when the formation parameters predicted by the neural network are forward modeled. This process speeds up the training speed and introduces certain errors, which enhances the anti-noise performance of the inversion neural network). The forward modeling process satisfies certain physical laws, so the physical law restrictions on the neural network are introduced during the training process, which enhances the accuracy and interpretability of the inversion neural network. During the training process of the inversion neural network, the initial learning rate is 0.0001, and the learning rate reduction factor is 0.7. When the model loss function does not decrease after 50 epochs, the learning rate is reduced to 0.7 times the current value, the batch size is 8192, and the optimizer uses the Adam optimizer for training, and the training is 1200 epochs.

[0043] After the inversion neural network training is completed, the test set data is used for testing and debugging. By adjusting the parameters of the inversion neural network such as the network structure, learning rate, and number of training cycles, the error between the predicted value of the inversion neural network and the actual formation parameter value is minimized.

[0044] Step S7: Inversion effect verification. In order to more intuitively verify the effect of the trained inversion neural network, we constructed the following Figure 6 The simple 3-layer stratigraphic model in (a) and Figure 7 The complex three-layer stratigraphic model of (a) is inverted using the inversion neural network trained in step S6. The inversion effect is shown in FIG. Figure 6 (b) and Figure 7 As shown in (b), it can be seen that for the simulated formation, the inverted formation parameters are basically consistent with the given formation parameters. Figure 8 The figure shows the inversion results of the actual well logging response. Since the formation parameters of the actual formation are unknown, they can only be compared through different inversion methods. Figure 8 (a) is the formation inversion result obtained by supervised descent method, and Figure 8 (b) is the inversion result of the neural network described in the present invention. The two have a high degree of consistency, which further verifies the functionality of the inversion neural network. In the actual application process, the received electromagnetic wave logging response is input into the trained physical drive neural network. 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 formation parameter inversion method based on physically driven neural network, characterized in that: The following steps are involved: The calculation steps of the one-dimensional forward algorithm are as follows: constructing a formation model, placing the logging while drilling instrument in the formation, inputting the real formation parameters into the one-dimensional forward algorithm, and the one-dimensional forward algorithm outputs the logging response value; Forward neural network training steps: The forward neural network is based on a deep neural network, with real formation parameter values ​​as the training set, the logging response values ​​output by the one-dimensional forward algorithm as the first forward modeling label, and the predicted formation parameters output as the first forward modeling result; The error between the first forward modeling label and the first forward modeling result is used as a reference to constrain the training process; Inversion neural network training steps: The inversion neural network is based on a deep neural network. The logging response values ​​calculated by the one-dimensional forward algorithm are used as the training set. When training the forward neural network, the real formation parameters of the training set are used as inversion labels, and the predicted formation parameters are output as the inversion results. The error between the inversion label and the inversion result is used as the inversion error. The trained forward neural network is then called, and the inversion result is used as the training set. The logging response values ​​output by the one-dimensional forward algorithm are used as the second forward label, and the predicted logging response values ​​are output as the second forward result. The error between the second forward modeling label and the second forward modeling result is taken as the second forward modeling error; the inversion error is combined with the second forward modeling error to calculate the total error, and the total error is used to constrain the training process of the inversion neural network; Formation parameter inversion step: input the logging response value while drilling into the trained inversion neural network, and the inversion neural network outputs the predicted formation parameters to complete the inversion.

2. The method according to claim 1, characterized in that: The formation parameters include formation resistivity value and formation thickness.

3. The method according to claim 2, characterized in that: The real formation parameters and logging response values ​​in the one-dimensional forward algorithm calculation step need to be preprocessed before being used in subsequent steps; the specific preprocessing is: The formation resistance is first taken inversely and then taken logarithm with base 10 for standard normalization. Directly perform standard normalization on the formation thickness; Perform column normalization on the logging response values.

4. The method according to claim 3, characterized in that: The constructed stratigraphic model is a three-layer stratigraphic model including an upper layer, a middle layer and a lower layer, and the logging while drilling instrument is placed in the middle layer.

5. The method according to claim 3, characterized in that: When two adjacent layers have the same resistivity, the constructed formation model is regarded as a two-layer formation model, including an upper layer and a lower layer, and the logging while drilling instrument can be placed in the upper layer or the lower layer.

6. The method according to claim 1, characterized in that: The mean square error between the inversion label and the inversion result is taken as the inversion error; the mean square error between the second forward modeling label and the second forward modeling result is taken as the second forward modeling error.

7. The method according to claim 1, characterized in that: The forward neural network model includes an input layer, a first fully connected layer and a first batch of normalized layers connected thereto, a second fully connected layer and a second batch of normalized layers connected thereto, a third fully connected layer and a third batch of normalized layers connected thereto, a fourth fully connected layer and a fourth batch of normalized layers connected thereto, a fifth fully connected layer and a fifth batch of normalized layers connected thereto, a sixth fully connected layer and a sixth batch of normalized layers connected thereto, and an output layer; The parameters of each layer are set as follows: the input layer has 5 units; the first, second and fourth fully connected layers, the first, second and fourth batch normalization layers have 4096 units; the third and fifth fully connected layers, the third and fifth batch normalization layers have 2048 units; the sixth fully connected layer and the sixth batch normalization layer have 1024 units; the output layer has 8 units; the activation function used in the first and fourth fully connected layers and the output layer is the Sigmoid function, and the activation function used in other layers is the ReLU function.

8. The method according to claim 1, characterized in that: The inversion neural network model includes an input layer, a first fully connected layer and a first batch of normalized layers connected thereto, a second fully connected layer and a second batch of normalized layers connected thereto, a third fully connected layer and a third batch of normalized layers connected thereto, a fourth fully connected layer and a fourth batch of normalized layers connected thereto, a fifth fully connected layer and a fifth batch of normalized layers connected thereto, a sixth fully connected layer and a sixth batch of normalized layers connected thereto, and an output layer; The parameters of each layer are set as follows: the input layer has 8 units; the first, second and fourth fully connected layers have 4096 units; the third and fifth fully connected layers have 2048 units each; the sixth fully connected layer has 1024 units; the output layer has 5 units; the activation function used in the first and fourth fully connected layers and the output layer is the Sigmoid function, and the activation function used in the remaining layers is the ReLU function.

Images