DC resistivity inversion method and system based on unsupervised deep learning

Abstract

The present invention provides a DC resistivity inversion method and system based on unsupervised deep learning, constructs an unsupervised DC resistivity inversion network architecture, and adds a dynamic smooth constraint term to the loss function. A simulation sample library is constructed with reference to the geological conditions of the detection area, and an actual sample library is constructed using the domain migration method for the previous engineering data. The unsupervised neural network is pre-trained using the simulation sample library, and then the network is retrained using the actual sample library through a transfer learning method based on linear exploration and complete fine-tuning. Finally, the newly collected observation data is brought into the trained network, and a single actual sample is iteratively inverted multiple times based on the network constraints to obtain a predicted resistivity model. The present invention can obtain a resistivity model through unsupervised deep learning inversion for the actually collected DC resistivity observation data, and uses regularization constraints to improve the imaging effect.

Description

Technical Field

[0001] The present invention belongs to the technical field of geophysical exploration and relates to a DC resistivity inversion method and system based on unsupervised deep learning. Background Art

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

[0003] The DC resistivity detection method is sensitive to water bodies and is an effective method for detecting water-bearing structures. Due to its low economic cost, high exploration efficiency, and flexible construction, it can meet a variety of detection environments and engineering needs, such as hidden danger detection of water conservancy and hydropower projects, geological surveys of highway and railway transportation projects, etc., and has important engineering value. The observation data can be reconstructed into a resistivity model through the inversion method to solve geological problems. Since the inverse problem has multiple solutions, the selection of an effective inversion method is the key to ensuring the effectiveness of DC resistivity detection.

[0004] At present, the linear inversion method is the mainstream method in practical applications. However, its inversion results rely on the initial model and are prone to fall into local optimality, which in turn leads to incorrect geological interpretation. Using deep learning to solve inverse problems can overcome local optimality and has become a research hotspot in recent years. Existing DC resistivity inversion based on deep learning uses a supervised method to train the network, and the training set requires a large number of real resistivity models as labels. However, the real resistivity model is usually difficult to obtain in actual detection, and the network cannot be effectively trained due to lack of labels, resulting in poor results of existing methods in actual data. Unsupervised inversion methods can use physical laws and data mining to dual-drive the training network, get rid of dependence on real models, and have the feasibility of application in actual data. At present, no DC resistivity inversion method based on unsupervised deep learning has been invented.

[0005] There are three challenges in implementing the DC resistivity inversion method based on unsupervised deep learning:

[0006] First, there is no existing unsupervised DC resistivity inversion network. To this end, it is necessary to incorporate the physical laws of the electric field into the existing network architecture and train the network using only observed data, essentially getting rid of the reliance of network training on labels.

[0007] Second, due to the multi-solution nature of the inverse problem, errors are prone to occur in gradient calculation, resulting in poor network training results. On the basis of physical laws driving network training, it is necessary to impose constraints on the network training process to alleviate the multi-solution problem and improve the inversion imaging effect.

[0008] Third, the cost of obtaining actual data is high, and it is difficult to meet the volume required for network training, resulting in poor results of unsupervised DC resistivity inversion methods in actual detection. Transfer learning is a good way to solve this problem, but there is currently a lack of robust DC resistivity transfer learning methods. Summary of the invention

[0009] In order to solve the above problems, the present invention proposes a DC resistivity inversion method and system based on unsupervised deep learning. The present invention can obtain a resistivity model through unsupervised deep learning inversion for the actually collected DC resistivity observation data, and use regularization constraints to improve the imaging effect.

[0010] According to some embodiments, the present invention adopts the following technical solutions:

[0011] A DC resistivity inversion method based on unsupervised deep learning includes the following steps:

[0012] The forward modeling process is added into the neural network framework to construct an unsupervised DC resistivity inversion network;

[0013] Add a dynamic smoothness constraint term to the loss function used to drive the update of network parameters;

[0014] Design a geoelectric model based on the geological conditions of the detection area, and then construct a sample library containing simulated observation data greater than the set amount;

[0015] The unsupervised DC resistivity inversion network is pre-trained using the sample library to preliminarily determine the mapping relationship between the observed data and the resistivity model;

[0016] Using the actual collected data, the network is retrained through transfer learning based on linear exploration and complete fine-tuning to optimize the mapping relationship between the observed data and the resistivity model;

[0017] The newly collected observation data is brought into the trained network, and a single actual sample is iteratively inverted multiple times based on the network constraints, and the predicted resistivity model is finally output;

[0018] The final predicted resistivity model is used to perform direct current resistivity inversion.

[0019] As an optional implementation method, the forward modeling process is added to the neural network framework. The specific process of constructing an unsupervised DC resistivity inversion network includes: using the finite element / finite difference method to implement the point source forward modeling process as a forward modeling module, splicing it to the output end of the neural network, forward modeling the prediction model to obtain predicted data, updating the network parameters by fitting the predicted data and the input data, and realizing network training.

[0020] As an optional implementation, the specific process of adding a dynamic smooth constraint term to the loss function used to drive the update of network parameters includes:

[0021] Adding a dynamic smooth constraint term to the loss function, the calculation formula is:

[0022] Loss = (G(m)-d obs ) T (G(m)-d obs )+λ(Cm) T (Cm)

[0023] G(·) is the forward process, m represents the resistivity model, d obs is the observation data input to the network, C is the smooth constraint matrix used to approximate the first / second order derivative of m in space, and λ is the regularization parameter.

[0024] As an optional implementation, constructing a sample library includes constructing a simulation sample library, determining modeling parameters according to the electrode arrangement, electrode spacing and detection requirements of actual detection, using finite element / finite difference methods to perform forward numerical simulation, and obtaining observation data corresponding to a geoelectric model greater than a set amount as a simulation sample library.

[0025] As an optional implementation, constructing a sample library includes constructing an actual sample library, processing actual data collected from similar detection scenarios using a data domain migration method so that the data has features similar to those of the simulated data, and obtaining the actual sample library.

[0026] As an optional implementation, the specific process of retraining the network through transfer learning based on linear exploration and full fine-tuning includes:

[0027] After pre-training the network with the simulated sample library, the network is re-trained with the actual sample library;

[0028] The second training is to retrain the parameters of the last network layer using linear exploration and train all network layer parameters using full fine-tuning.

[0029] As an optional implementation method, the specific process of multiple iterative inversion of a single actual sample based on network constraints includes: after the observation data is input into the inversion network, the first predicted resistivity model is obtained by mapping, the corresponding data is obtained using the forward modeling module, and then the residual with the input observation data is calculated; if the residual is less than the set value, the resistivity model is output as the final model; if the residual is greater than the preset value, the loss function is calculated to update the network parameters, and then the model is regenerated; the above process is repeated until the residual converges.

[0030] A DC resistivity inversion system based on unsupervised deep learning, comprising:

[0031] A forward modeling network module is configured to add the forward modeling process into the neural network framework to construct an unsupervised DC resistivity inversion network;

[0032] A constraint module, configured to add a dynamic smoothness constraint term to a loss function used to drive an update of network parameters;

[0033] A sample library construction module is configured to design a geoelectric model with reference to the geological conditions of the detection area, and then construct a sample library containing simulated observation data greater than a set amount;

[0034] A preliminary training module is configured to pre-train the unsupervised DC resistivity inversion network using a sample library to preliminarily determine the mapping relationship between the observed data and the resistivity model;

[0035] The secondary training module is configured to use the actual collected data to perform secondary training on the network through transfer learning based on linear exploration and complete fine-tuning, thereby optimizing the mapping relationship between the observed data and the resistivity model;

[0036] The final optimization module is configured to bring the newly acquired observation data into the trained network, perform multiple iterative inversions on a single actual sample based on the network constraints, and finally output a predicted resistivity model;

[0037] The inversion module is configured to perform direct current resistivity inversion using the final predicted resistivity model.

[0038] A computer-readable storage medium stores a plurality of instructions, wherein the instructions are suitable for being loaded by a processor of a terminal device and executing the steps in the method.

[0039] A terminal device includes a processor and a computer-readable storage medium, wherein the processor is used to implement various instructions; and the computer-readable storage medium is used to store multiple instructions, wherein the instructions are suitable for being loaded by the processor and executing the steps in the described method.

[0040] Compared with the prior art, the present invention has the following beneficial effects:

[0041] In order to solve the problem that the existing resistivity deep learning inversion methods are difficult to apply to actual data because their training relies on labels and the physical nature of electric field propagation is not considered, the electric field forward modeling that represents the physical laws is added to the inversion network architecture to replace the label function, and an unsupervised DC resistivity inversion network architecture is constructed, which improves the generalization of the inversion network and lays a theoretical foundation for DC resistivity inversion and detection in engineering sites.

[0042] Aiming at the problem that the resistivity inverse problem has strong multi-solutions and the inversion network is difficult to accurately image, the present invention adds a dynamic smooth constraint for alleviating the multi-solutions in the loss function, improves the stability of network training, and ensures the inversion imaging effect of the network.

[0043] In view of the problem that actual data is difficult to meet the sample size required for network training, the present invention establishes a transfer learning method suitable for actual electrical data, reduces the network's demand for actual sample size, and realizes effective inversion of actual data through multiple iterations based on network constraints. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] 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.

[0045] Figure 1 A flow chart of a DC resistivity inversion method based on unsupervised deep learning proposed in this embodiment;

[0046] Figure 2 A schematic diagram of a geoelectric model in a simulation sample library established in an embodiment;

[0047] Figure 3 It is the unsupervised deep learning inversion result in the embodiment. DETAILED DESCRIPTION

[0048] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0049] It should be noted that the following detailed descriptions are all illustrative and 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.

[0050] 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.

[0051] This embodiment discloses a DC resistivity inversion method based on unsupervised deep learning, such as Figure 1 As shown, the following steps are included:

[0052] Step S1, constructing a large number of geoelectric models according to the detection scenario, and calculating the corresponding observation data through forward numerical simulation, and then building a simulation data sample library;

[0053] In some embodiments, a geoelectric model is designed with reference to the geological conditions of the detection area, and then a sample library containing a large amount of simulated observation data is constructed.

[0054] The modeling parameters are determined according to the actual electrode arrangement, electrode spacing, detection requirements, etc. The observation devices are usually Wenner-Schlumberger, monopole-dipole and dipole-dipole devices. The finite element / finite difference method is used for forward numerical simulation to obtain a large amount of observation data corresponding to the geoelectric model as a sample library.

[0055] The identification method of this embodiment is mainly aimed at faults, karst caves and other unfavorable geological bodies in underground engineering, which are represented by a geoelectric model containing abnormal bodies with different resistivity, position and shape;

[0056] like Figure 2 As shown, taking the detection environment of Dehou Reservoir in Yunnan as the background, the model size of this embodiment is set to 5m (X) × 186m (Y) × 32m (Z), with a total of three survey lines, a length of 186m, a survey line spacing of 1m, 63 electrode points, an electrode spacing of 3m, a background resistivity of 400-500Ohm·m, 1-3 blocky low-resistance anomalies, and a resistivity value of 10-20Ohm·m. The observation device uses a Wenner-Schlumberger and a dipole-dipole device;

[0057] Step S2, using a small amount of actual data in the early stage of the project as a sample library for secondary training and processing it through a domain transfer method;

[0058] Step S3, adding the forward modeling module into the neural network framework to construct an unsupervised DC resistivity inversion network architecture;

[0059] A dynamic smoothness constraint is added to the loss function to enhance the stability of network training.

[0060] In this embodiment, the point source forward modeling process implemented by the finite element / finite difference method is used as a forward modeling module, which is spliced ​​to the output end of the neural network. The prediction model is forward modeled to obtain prediction data, and the network parameters are updated by fitting the prediction data and input data to achieve network training.

[0061] The specific process of adding dynamic smoothness constraints to the loss function includes:

[0062] The loss function based on dynamic smoothness constraint is as follows:

[0063] Loss = (G(m)-d obs ) T (G(m)-d obs )+λ(Cm) T (Cm)

[0064] G(·) is the forward process, m represents the resistivity model, d obsis the observation data input to the network, C is the smooth constraint matrix, which is used to approximate the first / second order derivative of m in space, and λ is the regularization parameter. The Gauss-Newton method can be used to obtain the model gradient δm, and then update all network parameters.

[0065] δm=(J T J+λC T C) -1 J T δd obs

[0066] The dynamic regularization parameter λ is calculated as follows:

[0067] λ=λ0×(1.0-epoch / max_epoch) μ

[0068] λ0 is the initial value of the regularization parameter, max_epoch is the maximum number of training times set for the network, and μ is the rate of change factor.

[0069] Step S4, such as Figure 1 As shown, the unsupervised DC resistivity inversion network is trained using the simulated sample library and the actual data sample library in turn;

[0070] Using the actual collected data, the network is retrained through transfer learning based on linear exploration and complete fine-tuning to optimize the mapping relationship between the observed data and the resistivity model.

[0071] The specific process of secondary training of the network by the transfer learning method includes:

[0072] First, the data domain migration method is used to process the actual data so that it has similar characteristics to the simulated data. When the amount of data is small, linear fitting can be used. When the amount of data is large, existing technologies such as adversarial neural networks can be used to achieve it.

[0073] Secondly, perform linear exploration on the pre-trained network with the processed actual data and retrain the last linear layer.

[0074] Finally, the network is fully fine-tuned, updating the network parameters of all layers simultaneously.

[0075] The main network parameters and hardware conditions in this embodiment are: the network is built based on the PyTorch platform, and the calculation uses an NVIDIA TITAN RTX graphics card with 24G video memory, which contains 4608 stream processing units. The main network parameters are: SGD optimizer, batch size (batchsize) is set to 12, and batchsize = 8 in this article. The learning rate is 0.01-0.02, the momentum is 0.9, the weight decay is 1e-4, and the maximum number of epochs set for the network is 100.

[0076] Step S5, bringing the newly collected observation data into the trained network, performing multiple iterative inversions on a single actual sample based on network constraints, and finally outputting a predicted resistivity model.

[0077] After a single sample observation data is input into the inversion network, the first predicted resistivity model is obtained through mapping, and the corresponding data is obtained using the forward modeling module, and then the residual with the input observation data is calculated. If the residual is small, the resistivity model is output as the final model. If the residual is large, the loss function is calculated to update the network parameters and then the model is regenerated. Repeat this process until the residual converges.

[0078] The trained unsupervised deep learning network constructs the mapping relationship between the observed data and the resistivity model, which can replace the inversion process and substitute the actual data newly collected in Dehou Reservoir. The imaging results are as follows: Figure 3 As shown in the figure, the horizontal 50-85m area (L1) and 110-125m area (L2) are obvious low-resistance areas. Combined with geological data, it is speculated that they may be water-bearing areas in the karst zone or shallow Quaternary soil. The detection results have been verified by core drilling, and the detection effect can meet the requirements of engineering applications.

[0079] 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 a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, 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, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0080] 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.

[0081] 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.

[0082] 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 A step that specifies a function in one or more boxes.

[0083] 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.

[0084] Although the above describes the specific implementation mode of the present invention in conjunction with the accompanying drawings, it is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art on the basis of the technical solution of the present invention without creative work are still within the scope of protection of the present invention.

Claims

1. A DC resistivity inversion method based on unsupervised deep learning, characterized in that: The following steps are involved: The forward modeling process is added into the neural network framework to construct an unsupervised DC resistivity inversion network; Add a dynamic smoothness constraint term to the loss function used to drive the update of network parameters; Design a geoelectric model based on the geological conditions of the detection area, and then construct a sample library containing simulated observation data greater than the set amount; The unsupervised DC resistivity inversion network is pre-trained using the sample library to preliminarily determine the mapping relationship between the observed data and the resistivity model; Using the actual collected data, the network is retrained through transfer learning based on linear exploration and complete fine-tuning to optimize the mapping relationship between the observed data and the resistivity model; The newly collected observation data is brought into the trained network, and a single actual sample is iteratively inverted multiple times based on the network constraints, and the predicted resistivity model is finally output; The final predicted resistivity model is used to perform direct current resistivity inversion.

2. A DC resistivity inversion method based on unsupervised deep learning as claimed in claim 1, characterized in that: The forward modeling process is added to the neural network framework, and the specific process of constructing an unsupervised DC resistivity inversion network includes: using the finite element / finite difference method to implement the point source forward modeling process as a forward modeling module, splicing it to the output end of the neural network, forward modeling the prediction model to obtain predicted data, and updating the network parameters by fitting the predicted data with the input data to achieve network training.

3. A DC resistivity inversion method based on unsupervised deep learning as claimed in claim 1, characterized in that The specific process of adding dynamic smoothness constraints to the loss function used to drive the update of network parameters includes: Adding a dynamic smooth constraint term to the loss function, the calculation formula is: Loss=(G(m)-d obs ) T (G(m)-d obs )+λ(Cm) T (Cm) G(·) is the forward process, m represents the resistivity model, d obs is the observation data input to the network, C is the smooth constraint matrix used to approximate the first / second order derivative of m in space, and λ is the regularization parameter.

4. The DC resistivity inversion method based on unsupervised deep learning as claimed in claim 1, characterized in that: Constructing a sample library includes constructing a simulation sample library, determining modeling parameters according to the actual electrode arrangement, electrode spacing and detection requirements of detection, using the finite element / finite difference method to perform forward numerical simulation, and obtaining observation data corresponding to the geoelectric model that is greater than the set amount as a simulation sample library.

5. The DC resistivity inversion method based on unsupervised deep learning as claimed in claim 1, characterized in that: Constructing a sample library includes constructing an actual sample library, processing actual data collected from similar detection scenarios using a data domain migration method so that the data has features similar to those of the simulated data, and obtaining an actual sample library.

6. The DC resistivity inversion method based on unsupervised deep learning according to claim 1, characterized in that: The specific process of retraining the network through transfer learning based on linear exploration and full fine-tuning includes: After pre-training the network with the simulated sample library, the network is re-trained with the actual sample library; The second training is to retrain the parameters of the last network layer using linear exploration and train all network layer parameters using full fine-tuning.

7. The DC resistivity inversion method based on unsupervised deep learning according to claim 1, characterized in that: The specific process of multiple iterative inversion of a single actual sample based on network constraints includes: after the observation data is input into the inversion network, the first predicted resistivity model is obtained through mapping, the corresponding data is obtained using the forward modeling module, and then the residual with the input observation data is calculated; if the residual is less than the set value, the resistivity model is output as the final model; if the residual is greater than the preset value, the loss function is calculated to update the network parameters, and then the model is regenerated; the above process is repeated until the residual converges.

8. A DC resistivity inversion system based on unsupervised deep learning, characterized in that: include: A forward modeling network module is configured to add the forward modeling process into the neural network framework to construct an unsupervised DC resistivity inversion network; A constraint module, configured to add a dynamic smoothness constraint term to a loss function used to drive an update of network parameters; A sample library construction module is configured to design a geoelectric model with reference to the geological conditions of the detection area, and then construct a sample library containing simulated observation data greater than a set amount; A preliminary training module is configured to pre-train the unsupervised DC resistivity inversion network using a sample library to preliminarily determine the mapping relationship between the observed data and the resistivity model; The secondary training module is configured to use the actual collected data to perform secondary training on the network through transfer learning based on linear exploration and full fine-tuning, thereby optimizing the mapping relationship between the observed data and the resistivity model; The final optimization module is configured to bring the newly acquired observation data into the trained network, perform multiple iterative inversions on a single actual sample based on the network constraints, and finally output a predicted resistivity model; The inversion module is configured to perform direct current resistivity inversion using the final predicted resistivity model.

9. A computer-readable storage medium, characterized in that: A plurality of instructions are stored therein, and the instructions are suitable for being loaded by a processor of a terminal device and executing the steps of the method described in any one of claims 1 to 7.

10. A terminal device, characterized in that: The method comprises a processor and a computer-readable storage medium, wherein the processor is used to implement various instructions; and the computer-readable storage medium is used to store a plurality of instructions, wherein the instructions are suitable for being loaded by the processor and executing the steps in the method according to any one of claims 1 to 7.

Images