Gradient optimization method for joint inversion of tunnel resistivity and polarizability based on deep learning

Abstract

The present invention provides a tunnel resistivity and polarizability joint inversion gradient optimization method based on deep learning, which obtains tunnel three-dimensional geological data; according to the tunnel three-dimensional geological data and a preset JointGradOptNet network, an optimized prediction result of the joint inversion gradient of the tunnel resistivity and polarizability is obtained; wherein, in the training of the JointGradOptNet network, a tunnel three-dimensional geological model database is constructed based on the common water-containing body in front of the tunnel, and the resistivity input gradient and polarizability input gradient of the network are obtained by the difference between the inversion result of one iteration and the initial uniform geoelectric model, and the resistivity target gradient and polarizability target gradient of the network are obtained by the difference between the geoelectric model and the initial uniform geoelectric model; the present invention realizes relatively accurate imaging of the abnormal body in front of the tunnel face.

Description

Technical Field

[0001] The present invention relates to the field of geophysical exploration technology, and in particular to a tunnel resistivity and polarizability joint inversion gradient optimization method based on deep learning. Background Art

[0002] The statements in this section merely provide background art related to the present invention and do not necessarily constitute prior art.

[0003] As one of the most commonly used geophysical exploration methods, the resistivity detection method is widely used in engineering surveys, hydrogeology, environmental surveys, resource exploration and other fields. Since the resistivity method is sensitive to water body responses, it has been introduced into the field of tunnel advanced geological prediction to detect the source of sudden water disasters. In recent years, focused sounding observation devices with the functions of shielding rear interference and forward sounding of the face have been successfully applied in multiple tunnel projects. The focused sounding observation mode utilizes the principle of mutual repulsion of like currents, so that the face current produces a similar bunching effect, reducing the strong interference from near the face. By moving the like source power supply electrode backward and increasing the distance between the power supply and measurement electrodes, effective perception of water body information at different distances ahead is achieved. The observation data is processed by the inversion method to obtain the distribution information of the water body in front of the tunnel.

[0004] According to the inventors, the most commonly used tunnel resistivity data inversion method is linear inversion. By using observation data information at different distances, the position and morphological information of the water body in front of the tunnel can be captured. However, the traditional linear inversion method has problems such as heavy reliance on the initial model, easy to fall into local optimality and low computational efficiency, which need to be further improved. In recent years, with the emergence of deep learning methods, the geophysical field has gradually begun to use deep learning methods to solve inversion imaging problems, but the inversion accuracy of this method depends on the quality and type of data, there is an overfitting problem, poor generalization, and limited scope of application. The advantages and disadvantages of traditional methods and deep learning methods are complementary. How to effectively combine the advantages of the two and make up for their respective shortcomings is particularly critical. At present, the deep learning method is mainly based on the deep learning method in combination with the traditional method. Generally, only physical constraints are added to the loss function to assist training, which still cannot solve the inherent problems of the deep learning method. And there is no research to introduce deep learning methods into traditional methods.

[0005] There are two difficulties in introducing deep learning methods into traditional electrical inversion methods:

[0006] (1) Traditional electrical inversion methods update model parameters through repeated iterations to make the model approach the real model, while deep learning inversion methods train network parameters through large-scale data sets and directly learn the nonlinear mapping relationship between electrical data and models. The working logic of the two methods is completely different. Therefore, how to find a starting point to introduce deep learning into the traditional electrical inversion process and use the advantages of deep learning to assist in solving the shortcomings of traditional inversion methods is the primary issue for the efficient combination of deep learning and traditional methods.

[0007] (2) Due to the strong nonlinear mapping ability of deep learning methods, when the learning task is relatively simple, the learning results are prone to "generalizing from a single example", which is far from the objective law. In particular, most of the current deep learning resistivity inversion methods only use a single resistivity physical property parameter, which is prone to overfitting. Therefore, how to guide the deep neural network parameters to update in a meaningful direction by combining multiple physical property parameter inversion tasks and learn the real objective laws is a key issue to improve the robustness of deep learning methods. Summary of the invention

[0008] In order to address the shortcomings of the prior art, the present invention, on the one hand, provides a tunnel resistivity inversion gradient optimization method based on deep learning, proposes a deep learning resistivity inversion gradient optimization same-domain mapping idea for the characteristics of tunnel resistivity data, and constructs a tunnel resistivity inversion gradient optimization method based on deep learning. With traditional inversion as the skeleton, the correspondence between the initial gradient and the target gradient is learned through a deep neural network to optimize the traditional inversion gradient; on this basis, the present invention, on the other hand, provides a tunnel resistivity and polarizability joint inversion gradient optimization method based on deep learning, introduces polarizability data, constructs a dual-channel joint gradient optimization network with the initial resistivity gradient and the polarizability gradient as input and the dual target gradient as output, and draws on the traditional cross-gradient idea to introduce a collinear loss function to constrain the resistivity inversion and the polarizability inversion to ensure that the two physical parameters change in the same direction.

[0009] In order to achieve the above object, the present invention adopts the following technical solution:

[0010] The first aspect of the present invention provides a tunnel resistivity inversion gradient optimization method based on deep learning.

[0011] A tunnel resistivity inversion gradient optimization method based on deep learning includes the following processes:

[0012] Obtain 3D geological data of tunnels;

[0013] Based on the tunnel 3D geological data and the preset GradOptNet network, the optimized prediction results of the tunnel resistivity inversion gradient are obtained;

[0014] Among them, in the training of the GradOptNet network, a three-dimensional geological model database of the tunnel is constructed based on the common water-containing bodies in front of the tunnel. The input gradient of the GradOptNet network is obtained by the difference between the inversion result of one iteration and the initial uniform geoelectric model, and the target gradient of the network is obtained by the difference between the geoelectric model and the initial uniform geoelectric model.

[0015] A second aspect of the present invention provides a tunnel resistivity inversion gradient optimization system based on deep learning.

[0016] A tunnel resistivity inversion gradient optimization system based on deep learning, comprising:

[0017] The data acquisition module is configured to: acquire three-dimensional geological data of the tunnel;

[0018] The inversion gradient optimization module is configured to: obtain the optimized prediction result of the tunnel resistivity inversion gradient according to the tunnel three-dimensional geological data and the preset GradOptNet network;

[0019] Among them, in the training of the GradOptNet network, a three-dimensional geological model database of the tunnel is constructed based on the common water-containing bodies in front of the tunnel. The input gradient of the GradOptNet network is obtained by the difference between the inversion result of one iteration and the initial uniform geoelectric model, and the target gradient of the network is obtained by the difference between the geoelectric model and the initial uniform geoelectric model.

[0020] The third aspect of the present invention provides a tunnel resistivity and polarizability joint inversion gradient optimization method based on deep learning.

[0021] A tunnel resistivity and polarizability joint inversion gradient optimization method based on deep learning includes the following processes:

[0022] Obtain 3D geological data of tunnels;

[0023] Based on the tunnel 3D geological data and the preset JointGradOptNet network, the optimized prediction results of the joint inversion gradient of the tunnel resistivity and polarizability are obtained;

[0024] Among them, in the training of the JointGradOptNet network, a three-dimensional geological model database of the tunnel is constructed based on the common water-containing bodies in front of the tunnel. The resistivity input gradient and polarizability input gradient of the network are obtained by the difference between the inversion result of one iteration and the initial uniform geoelectric model. The resistivity target gradient and polarizability target gradient of the network are obtained by the difference between the geoelectric model and the initial uniform geoelectric model.

[0025] The fourth aspect of the present invention provides a tunnel resistivity and polarizability joint inversion gradient optimization system based on deep learning.

[0026] A tunnel resistivity and polarizability joint inversion gradient optimization system based on deep learning, comprising:

[0027] The data acquisition module is configured to: acquire three-dimensional geological data of the tunnel;

[0028] The joint inversion gradient optimization module is configured to: obtain the optimized prediction results of the joint inversion gradient of the tunnel resistivity and polarizability according to the three-dimensional geological data of the tunnel and the preset JointGradOptNet network;

[0029] Among them, in the training of the JointGradOptNet network, a three-dimensional geological model database of the tunnel is constructed based on the common water-containing bodies in front of the tunnel. The resistivity input gradient and polarizability input gradient of the network are obtained by the difference between the inversion result of one iteration and the initial uniform geoelectric model. The resistivity target gradient and polarizability target gradient of the network are obtained by the difference between the geoelectric model and the initial uniform geoelectric model.

[0030] Compared with the prior art, the present invention has the following beneficial effects:

[0031] 1. Aiming at how to introduce deep learning into the traditional inversion method, the present invention proposes a dual-driven inversion idea with traditional inversion method as the main method and deep learning as the auxiliary method, establishes a tunnel resistivity inversion gradient optimization method based on deep learning, takes the traditional inversion gradient as the entry point, uses deep neural network to learn the mapping relationship between gradients, and realizes more accurate imaging of the abnormal body in front of the tunnel face.

[0032] 2. To address the overfitting problem of deep learning methods, the present invention draws on the ideas of traditional joint inversion in geophysics and deep learning multi-task learning, proposes collinearity constraints, and establishes a tunnel resistivity / polarizability joint inversion gradient optimization algorithm based on deep learning to ensure the structural similarity of resistivity parameters and polarizability parameters, realizes the joint inversion of resistivity / polarizability, and improves the robustness of the algorithm to disturbances such as data noise.

[0033] Advantages of additional aspects of the present invention will be given in part in the following description, and in part will become obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] The accompanying drawings in the specification, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0035] Figure 1 A flow chart of the three-dimensional deep neural network GradOptNet method with gradient optimization as the task provided in Example 5 of the present invention.

[0036] Figure 2 The loss function calculation and network diagram of GradOptNet provided in Example 5 of the present invention.

[0037] Figure 3 This is a flow chart of the JointGradOptNet method for joint inversion gradient optimization of tunnel resistivity and polarizability provided in Example 5 of the present invention.

[0038] Figure 4 The loss function calculation and network diagram of JointGradOptNet provided in Example 5 of the present invention.

[0039] Figure 5 Each convolutional layer setting provided in Example 5 of the present invention.

[0040] Figure 6 Schematic diagram of the tunnel resistivity database provided in Example 5 of the present invention.

[0041] Figure 7 Schematic diagram of the polarizability database provided in Example 5 of the present invention.

[0042] Figure 8 This is the deep learning inversion result 1 provided in Example 5 of the present invention.

[0043] Fig. 9 This is the second deep learning inversion result provided in Example 5 of the present invention. DETAILED DESCRIPTION

[0044] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0045] It should be noted that the following detailed descriptions are exemplary and are intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present invention belongs.

[0046] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit exemplary embodiments according to the present invention. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprising" and / or "including" are used in this specification, it indicates the presence of features, steps, operations, devices, components and / or combinations thereof.

[0047] In the absence of conflict, the embodiments of the present invention and the features of the embodiments may be combined with each other.

[0048] Embodiment 1:

[0049] Embodiment 1 of the present invention provides a tunnel resistivity inversion gradient optimization method based on deep learning, comprising the following process:

[0050] Obtain 3D geological data of tunnels;

[0051] Based on the tunnel 3D geological data and the preset GradOptNet network, the optimized prediction results of the tunnel resistivity inversion gradient are obtained;

[0052] Among them, in the training of the GradOptNet network, a three-dimensional geological model database of the tunnel is constructed based on the common water-containing bodies in front of the tunnel. The input gradient of the GradOptNet network is obtained by the difference between the inversion result of one iteration and the initial uniform geoelectric model, and the target gradient of the network is obtained by the difference between the geoelectric model and the initial uniform geoelectric model.

[0053] Specifically, the process includes the following:

[0054] A 3D tunnel geological model database was built based on the common water bodies in front of the tunnel, including 10,134 numerical simulation data. For each sample data, the data set also has its corresponding geoelectric model and the corresponding traditional method's one-iteration inversion result;

[0055] Based on the 3D U-Net network architecture, a 3D deep neural network GradOptNet (Gradient Optimized Neural Network) with gradient optimization as the task was constructed for the 3D tunnel environment. A set of original gradient data was randomly extracted and input into GradOptNet for training, and the predicted gradient was output.

[0056] Calculate the data loss function, calculate the error between the output prediction gradient and the true target gradient, and perform gradient backpropagation to optimize the three-dimensional deep neural network GradOptNet;

[0057] The GradOptNet network is used to map the original gradient data of tunnel resistivity to the geological model. After the network training is completed, the test set data is used for testing to obtain the predicted target gradient, which is then added to the initial uniform model to obtain the optimized resistivity inversion result.

[0058] The construction of the tunnel 3D geological model database includes:

[0059] According to the common water-bearing geological structures in front of the tunnel, three types of typical tunnel anomalies were designed, mainly including fault model, single cave model and two cave models;

[0060] The database for tunnel resistivity inversion gradient optimization is used to obtain the input gradient of the network by the difference between the inversion result of one iteration and the initial uniform geoelectric model. The target gradient of the network is obtained by the difference between the geoelectric model and the initial uniform geoelectric model The training sample pairs were constructed;

[0061] For the actual application scenario of tunnel resistivity inversion, there is only one form of anomaly, namely low-resistance anomaly. Finally, the dataset is randomly divided into training set, validation set and test set in a ratio of 10:1:1;

[0062] A three-dimensional deep neural network GradOptNet with gradient optimization as the task is constructed. The network mainly consists of two parts: encoder and decoder. The decoder extracts high-level semantic information features of the data by compressing the input data size and expanding the number of data channels to obtain global information.

[0063] The reason why mean square error loss can be applied to regression problems is that under the assumption that the error between the model output and the true value follows a Gaussian distribution, minimizing the mean square error loss function is essentially consistent with the maximum likelihood estimation. Therefore, as long as this assumption holds true in scenarios such as regression problems, mean square error loss can achieve ideal prediction results. The deep learning inversion of resistivity can be regarded as a value regression problem in deep learning. Therefore, in this embodiment, the MSE metric is selected on the GradOptNet loss function. The MSE metric penalizes the error of the gradient prediction value relative to the target gradient value. Therefore, the mean square error loss function of GradOptNet is:

[0064]

[0065] in, is the gradient prediction value, m i,j,k is the target gradient value.

[0066] Embodiment 2:

[0067] Embodiment 2 of the present invention provides a tunnel resistivity inversion gradient optimization system based on deep learning, comprising:

[0068] The data acquisition module is configured to: acquire three-dimensional geological data of the tunnel;

[0069] The inversion gradient optimization module is configured to: obtain the optimized prediction result of the tunnel resistivity inversion gradient according to the tunnel three-dimensional geological data and the preset GradOptNet network;

[0070] Among them, in the training of the GradOptNet network, a three-dimensional geological model database of the tunnel is constructed based on the common water-containing bodies in front of the tunnel. The input gradient of the GradOptNet network is obtained by the difference between the inversion result of one iteration and the initial uniform geoelectric model, and the target gradient of the network is obtained by the difference between the geoelectric model and the initial uniform geoelectric model.

[0071] The working method of the system is the same as that described in Example 1, and will not be repeated here.

[0072] Embodiment 3:

[0073] Embodiment 3 of the present invention provides a tunnel resistivity and polarizability joint inversion gradient optimization method based on deep learning, comprising the following process:

[0074] Obtain 3D geological data of tunnels;

[0075] Based on the tunnel 3D geological data and the preset JointGradOptNet network, the optimized prediction results of the joint inversion gradient of the tunnel resistivity and polarizability are obtained;

[0076] Among them, in the training of the JointGradOptNet network, a three-dimensional geological model database of the tunnel is constructed based on the common water-containing bodies in front of the tunnel. The resistivity input gradient and polarizability input gradient of the network are obtained by the difference between the inversion result of one iteration and the initial uniform geoelectric model. The resistivity target gradient and polarizability target gradient of the network are obtained by the difference between the geoelectric model and the initial uniform geoelectric model.

[0077] Specifically, the process includes the following:

[0078] A 3D tunnel geological model database was built based on the common water bodies in front of the tunnel, including 10,134 numerical simulation data. For each sample data, the data set also has its corresponding geoelectric model and the corresponding traditional method's one-iteration inversion result.

[0079] Based on 10134 resistivity data of the tunnel resistivity inversion gradient optimization database, the corresponding polarizability numerical simulation data of the resistivity data are provided. For each polarizability data, the corresponding polarizability model and the polarizability inversion result of one iteration of the corresponding traditional inversion method are also provided.

[0080] Based on the resistivity three-dimensional deep neural network GradOptNet, a tunnel resistivity / polarizability joint inversion gradient optimization deep neural network JointGradOptNet is proposed;

[0081] Based on the joint inversion network JointGradOptNet, a set of joint original gradient data is randomly extracted, input into JointGradOptNet for training, and the predicted gradient is output;

[0082] Based on the data loss function, the gradient collinearity constraint is added, the error between the output prediction gradient and the true target gradient is calculated, and the gradient is returned to optimize the joint inversion network JointGradOptNet.

[0083] For the tunnel resistivity / polarizability joint inversion gradient optimization database, the resistivity input gradient of the network is obtained by the difference between the inversion result of one iteration and the initial uniform geoelectric model. With polarizability input gradient The resistivity target gradient of the network is obtained by the difference between the geoelectric model and the initial uniform geoelectric model. With polarizability target gradient The training sample pairs are constructed.

[0084] For the actual application scenario of tunnel resistivity inversion, there is only one form of anomaly, namely low-resistance anomaly. Finally, the dataset is randomly divided into training set, validation set and test set in a ratio of 10:1:1.

[0085] In this embodiment, based on the tunnel resistivity data, the corresponding polarizability numerical simulation data is continued. According to the equivalent resistivity formula, the equivalent resistivity of the model can be calculated through the resistivity ρ and polarizability η of the abnormal body model. The forward modeling result of the resistivity with induced polarization effect is obtained by using the equivalent resistivity. Continue to perform forward modeling through the geoelectric model and obtain the forward response result ρ without induced polarization effect s Finally, using the equivalent resistivity formula The forward response of the polarizability can be obtained.

[0086] A tunnel resistivity / polarizability joint inversion gradient optimization deep neural network JointGradOptNet is constructed. The network mainly consists of an encoder and two decoders, and the structures of the two decoders are exactly the same. The specific structure of the encoder and decoder of JointGradOptNet is the same as that of the encoder and decoder of GradOptNet. The input data of JointGradOptNe is the joint initial gradient of resistivity and polarizability, that is, a two-channel input data, and the two channels are the initial gradient of resistivity and the initial gradient of polarizability. JointGradOptNet mines relevant information and extracts common features from the joint initial gradient information of resistivity and polarizability through the encoder, and then inputs the common features into the two decoders respectively, and obtains the predicted target resistivity gradient and target polarizability gradient through decoding.

[0087] The output of the deep learning-based tunnel resistivity / polarizability inversion gradient optimization deep neural network (JointGradOptNet) is a gradient. Therefore, a gradient collinearity constraint can be added to the output gradient to conveniently implement the cross gradient of the traditional joint inversion, that is:

[0088]

[0089] in, is the resistivity target gradient, Polarizability target gradient.

[0090] Finally, the loss function of the deep learning-based tunnel resistivity / polarizability joint inversion gradient optimization method can be defined as:

[0091]

[0092] Embodiment 4:

[0093] Embodiment 4 of the present invention provides a tunnel resistivity and polarizability joint inversion gradient optimization system based on deep learning, including:

[0094] The data acquisition module is configured to: acquire three-dimensional geological data of the tunnel;

[0095] The joint inversion gradient optimization module is configured to: obtain the optimized prediction results of the joint inversion gradient of the tunnel resistivity and polarizability according to the three-dimensional geological data of the tunnel and the preset JointGradOptNet network;

[0096] Among them, in the training of the JointGradOptNet network, a three-dimensional geological model database of the tunnel is constructed based on the common water-containing bodies in front of the tunnel. The resistivity input gradient and polarizability input gradient of the network are obtained by the difference between the inversion result of one iteration and the initial uniform geoelectric model. The resistivity target gradient and polarizability target gradient of the network are obtained by the difference between the geoelectric model and the initial uniform geoelectric model.

[0097] The working method of the system is the same as that described in Example 1, and will not be repeated here.

[0098] Embodiment 5:

[0099] Embodiment 5 of the present invention provides a tunnel resistivity inversion gradient optimization method based on deep learning (such as Figure 1 As shown in the figure) and the tunnel resistivity / polarizability joint inversion gradient optimization method based on deep learning (as shown in the figure) Figure 2 ), including the following process:

[0100] Step S1, constructing a tunnel three-dimensional geological model database through computer numerical simulation, wherein the database includes a tunnel resistivity gradient optimization database (containing only resistivity data) and a tunnel resistivity / polarizability database (containing resistivity data and polarizability);

[0101] The method in this example is mainly aimed at typical abnormal bodies in tunnels, which mainly include three types: fault model, single cave model and two cave models. Here, models with different positions, sizes and values ​​are used to represent them, such as Figure 6 and Figure 7 shown.

[0102] The basic settings of the geoelectric model of this embodiment are as follows: the resistivity of the surrounding rock is set to 1000Ω·m, the resistivity value of the tunnel cavity is set to 10000000Ω·m, and the resistivity value of the low-resistance anomaly is set to 30Ω·m~200Ω·m. The size of this data model is 11×15×18, and it is powered by multi-isotropic source electrodes. A total of 3 survey lines are arranged, and each survey line has 9 measurement points; the surrounding rock polarizability of the polarizability model is set to 0.01, the polarizability value of the tunnel cavity is set to 0, and the polarizability value of the high-polarizability anomaly is set to 0.46-0.12;

[0103] Step S2, a three-dimensional deep neural network GradOptNet with gradient optimization as the task and a tunnel resistivity / polarizability joint inversion gradient optimization deep neural network JointGradOptNet based on the three-dimensional deep neural network GradOptNet;

[0104] In this embodiment, inversion gradient optimization is achieved through an end-to-end training method. The network mainly consists of two parts: an encoder and a decoder. The decoder extracts high-level semantic information features of the data by compressing the input data size and expanding the number of data channels to obtain global information.

[0105] Specifically, in this embodiment, four conv-down blocks are used to form an encoder, each of which contains two convolution operations and a maximum pooling operation, where each convolution operation contains a three-dimensional convolution operation, a batch normalization operation, and a ReLU activation function. The decoder upsamples the global information obtained by the encoder, and finally obtains a prediction result with the same size as the input original data. In the decoding process, the low-level local information obtained by the encoding layer is introduced into the decoding process through a shortcut connection, and the global information and local information are combined to ensure that the detailed information of the decoding result is not lost. The decoder is composed of four conv-up blocks. Each conv-up block first upsamples the features through a three-dimensional transposed convolution layer, and then performs two additional common three-dimensional convolution operations through the convolution layer. Each convolution layer is set as follows: Figure 5 shown.

[0106] Step S3, as in 3 and Figure 4 As shown, two loss functions are designed to complete the three-dimensional deep neural network GradOptNet with gradient optimization as the task and the tunnel resistivity / polarizability joint inversion gradient optimization deep neural network JointGradOptNet based on this.

[0107] In this example, the MSE metric is selected on the GradOptNet loss function. The MSE metric penalizes the error of the gradient prediction value relative to the target gradient value, so the mean square error loss function is defined as:

[0108]

[0109] in, is the gradient prediction value, m i,j,k is the target gradient value.

[0110] The output of the deep learning-based tunnel resistivity / polarizability inversion gradient optimization deep neural network (JointGradOptNet) is a gradient. Therefore, a gradient collinearity constraint can be added to the output gradient to conveniently implement the cross gradient of the traditional joint inversion, that is:

[0111]

[0112] in, is the resistivity target gradient, Polarizability target gradient.

[0113] Finally, the loss function of the deep learning-based tunnel resistivity / polarizability joint inversion gradient optimization method can be defined as:

[0114]

[0115] Step S4, training the GradOptNet network and the JointGradOptNet network.

[0116] The main network parameters and hardware conditions in this embodiment are: the calculation is implemented using a single-chip NVIDIA TITAN Xp; the network is built based on the PyTorch platform, the SGD optimizer batch size (batchsize) is 12, the learning rate (learning rate) is 0.125, and the number of epochs (epochs) of the learning algorithm in the entire training data set is 400.

[0117] Step S5, the GradOptNet network and the JointGradOptNet network construct the mapping relationship between the initial gradient data and the target gradient data, which can represent the inversion process. Substituting some results of the test set into Figure 8 and Fig. 9 shown. Figure 8 This is the inversion result of the GradOptNet network. Compared with the traditional inversion results, this method has high inversion accuracy and stable effect, especially in the description of the abnormal body boundary and resistivity value. Fig. 9 The inversion results of the GradOptNet network based on 3dBw noise intensity are compared with the inversion results of the JointGradOptNet network, which can effectively alleviate the overfitting problem of deep learning and improve the robustness of the method. The training time of the embodiment network is about 24 hours, and the test time of 860 sets of data is about 1 minute. The inversion efficiency can meet the requirements of engineering applications.

[0118] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of hardware embodiments, software embodiments, or embodiments combining software and hardware. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage and optical storage, etc.) containing computer-usable program code.

[0119] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0120] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0121] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0122] A person skilled in the art can understand that all or part of the processes in the above-mentioned embodiments can be implemented by instructing the relevant hardware through a computer program, and the program can be stored in a computer-readable storage medium, and when the program is executed, it can include the processes of the embodiments of the above-mentioned methods. The storage medium can be a disk, an optical disk, a read-only memory (ROM) or a random access memory (RAM), etc.

[0123] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A tunnel resistivity inversion gradient optimization method based on deep learning, Features: The process includes: Obtain 3D geological data of tunnels; Based on the tunnel 3D geological data and the preset GradOptNet network, the optimized prediction results of the tunnel resistivity inversion gradient are obtained; GradOptNet network, including: encoder and decoder; The encoder consists of four conv-down blocks, each of which consists of two convolution operations and a maximum pooling operation. Each convolution operation consists of a three-dimensional convolution operation, a batch normalization operation, and a ReLU activation function. The decoder consists of four conv-up blocks, each of which first upsamples the features through a 3D transposed convolutional layer and then performs two additional common 3D convolution operations through the convolutional layer; The decoder upsamples the global information obtained by the encoder to obtain a prediction result with the same size as the input original data. During the decoding process, the local information of the lower layer obtained by the encoding layer is introduced into the decoding process through a short-circuit connection to combine the global information with the local information. Among them, in the training of the GradOptNet network, a three-dimensional geological model database of the tunnel is constructed based on the common water-containing bodies in front of the tunnel. The input gradient of the GradOptNet network is obtained by the difference between the inversion result of one iteration and the initial uniform geoelectric model, and the target gradient of the network is obtained by the difference between the geoelectric model and the initial uniform geoelectric model.

2. The tunnel resistivity inversion gradient optimization method based on deep learning as claimed in claim 1, Features: The GradOptNet network uses the mean square error loss function, including: in, is the gradient prediction value, m i,j,k is the target gradient value.

3. The tunnel resistivity inversion gradient optimization method based on deep learning as claimed in claim 1, Features: Common water-bearing bodies include: fault model, single cave model and two cave models.

4. A tunnel resistivity inversion gradient optimization system based on deep learning, Features: include: The data acquisition module is configured to: acquire three-dimensional geological data of the tunnel; The inversion gradient optimization module is configured to: obtain the optimized prediction result of the tunnel resistivity inversion gradient according to the tunnel three-dimensional geological data and the preset GradOptNet network; GradOptNet network, including: encoder and decoder; The encoder consists of four conv-down blocks, each of which consists of two convolution operations and a maximum pooling operation. Each convolution operation consists of a three-dimensional convolution operation, a batch normalization operation, and a ReLU activation function. The decoder consists of four conv-up blocks, each of which first upsamples the features through a 3D transposed convolutional layer and then performs two additional common 3D convolution operations through the convolutional layer; The decoder upsamples the global information obtained by the encoder to obtain a prediction result with the same size as the input original data. During the decoding process, the local information of the lower layer obtained by the encoding layer is introduced into the decoding process through a short-circuit connection to combine the global information with the local information. Among them, in the training of the GradOptNet network, a three-dimensional geological model database of the tunnel is constructed based on the common water-containing bodies in front of the tunnel. The input gradient of the GradOptNet network is obtained by the difference between the inversion result of one iteration and the initial uniform geoelectric model, and the target gradient of the network is obtained by the difference between the geoelectric model and the initial uniform geoelectric model.

5. A joint inversion gradient optimization method of tunnel resistivity and polarizability based on deep learning, Features: The process includes: Obtain 3D geological data of tunnels; Based on the tunnel 3D geological data and the preset JointGradOptNet network, the optimized prediction results of the joint inversion gradient of the tunnel resistivity and polarizability are obtained; The JointGradOptNet network includes: an encoder and two decoders, and the two decoders have exactly the same structure; The encoder consists of four conv-down blocks, each of which consists of two convolution operations and a maximum pooling operation. Each convolution operation consists of a three-dimensional convolution operation, a batch normalization operation, and a ReLU activation function. The decoder consists of four conv-up blocks, each of which first upsamples the features through a 3D transposed convolutional layer and then performs two additional common 3D convolution operations through the convolutional layer; Extracting common features from the joint initial gradient of resistivity and polarizability through an encoder, inputting the common features into two decoders respectively, and obtaining the predicted target resistivity gradient and target polarizability gradient respectively through two decoders; Among them, in the training of the JointGradOptNet network, a three-dimensional geological model database of the tunnel is constructed based on the common water-containing bodies in front of the tunnel. The resistivity input gradient and polarizability input gradient of the network are obtained by the difference between the inversion result of one iteration and the initial uniform geoelectric model. The resistivity target gradient and polarizability target gradient of the network are obtained by the difference between the geoelectric model and the initial uniform geoelectric model.

6. The tunnel resistivity and polarizability joint inversion gradient optimization method based on deep learning as claimed in claim 5, Features: Based on the tunnel resistivity data and the corresponding polarizability numerical simulation data, according to the equivalent resistivity formula, the corresponding equivalent resistivity of the model is calculated through the resistivity and polarizability of the abnormal body model. The equivalent resistivity is used for forward modeling to obtain the resistivity forward modeling result with induced polarization effect. Continue to perform forward modeling through the geoelectric model and obtain the forward response result ρ without induced polarization effect s , using the equivalent resistivity formula The forward response of the polarizability is obtained.

7. The tunnel resistivity and polarizability joint inversion gradient optimization method based on deep learning as claimed in claim 5, Features: The loss function of JointGradOptNet includes: in, is the resistivity target gradient, is the true resistivity gradient, is the target gradient of polarizability, is the true gradient of polarizability.

8. A tunnel resistivity and polarizability joint inversion gradient optimization system based on deep learning, Features: include: The data acquisition module is configured to: acquire three-dimensional geological data of the tunnel; The joint inversion gradient optimization module is configured to: obtain the optimized prediction results of the joint inversion gradient of the tunnel resistivity and polarizability according to the three-dimensional geological data of the tunnel and the preset JointGradOptNet network; The JointGradOptNet network includes: an encoder and two decoders, and the two decoders have exactly the same structure; The encoder consists of four conv-down blocks, each of which consists of two convolution operations and a maximum pooling operation. Each convolution operation consists of a three-dimensional convolution operation, a batch normalization operation, and a ReLU activation function. The decoder consists of four conv-up blocks, each of which first upsamples the features through a 3D transposed convolutional layer and then performs two additional common 3D convolution operations through the convolutional layer; Extracting common features from the joint initial gradient of resistivity and polarizability through an encoder, inputting the common features into two decoders respectively, and obtaining the predicted target resistivity gradient and target polarizability gradient respectively through two decoders; Among them, in the training of the JointGradOptNet network, a three-dimensional geological model database of the tunnel is constructed based on the common water-containing bodies in front of the tunnel. The resistivity input gradient and polarizability input gradient of the network are obtained by the difference between the inversion result of one iteration and the initial uniform geoelectric model. The resistivity target gradient and polarizability target gradient of the network are obtained by the difference between the geoelectric model and the initial uniform geoelectric model.

Images