Cave inversion and imaging method and system based on multi-scale unsupervised deep learning

Abstract

The present invention provides a karst cave inversion and imaging method and system based on multi-scale unsupervised deep learning, constructs multiple low-resistance geoelectric models, performs forward simulation to obtain observation data; builds a deep learning network, constructs a loss function containing multi-scale and smooth constraints, calculates the model gradient, uses a learning rate weighted constraint method to improve the model training ability in low-sensitivity areas, updates the deep learning network parameters, and trains the deep learning network to determine the mapping relationship between the observation data and the geoelectric model; uses the trained deep learning network to process the collected observation data, obtains the corresponding geoelectric model diagram, and realizes inversion.

Description

Technical Field

[0001] The present invention belongs to the technical field of geophysical exploration and relates to a cave inversion and imaging method and system based on multi-scale 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] In recent years, a large number of infrastructure projects such as transportation, water conservancy and hydropower have been put into construction. However, during the construction period, unfavorable geological bodies such as underground water-bearing caves are often encountered. If they are not explored, disasters such as sudden water inrush will occur, causing economic losses and casualties. Cross-hole resistivity CT detection is widely used in the detection of underground water-bearing caves due to its high sensitivity to low-resistance geological bodies such as water-bearing structures.

[0004] The effect of inversion and imaging of water-bearing caves mainly depends on the inversion method. The unsupervised deep learning inversion method does not rely on the actual resistivity model, but directly uses the simulated data to fit the observed data, which can be better applied in actual detection. However, the use of unsupervised deep learning inversion and imaging caves in cross-hole resistivity CT detection still has the following problems:

[0005] (1) Due to the obvious volume effect of the electric field, in the absence of prior information, the resistivity inversion imaging results are quite different from the actual model, especially for boundary characterization.

[0006] (2) The data sensitivity between boreholes is unevenly distributed, which is usually manifested as the sensitivity of the area close to the electrode is higher than that of the area far away from the electrode. The uneven sensitivity will lead to poor imaging effect of the abnormal body. Summary of the invention

[0007] In order to solve the above problems, the present invention proposes a cave inversion and imaging method and system based on multi-scale unsupervised deep learning. The present invention adds the model gradient calculated by the multi-scale method to the neural network, constructs an inversion architecture based on multi-scale unsupervised deep learning, improves the network's ability to depict the boundaries of the cave, and directly applies differentiated learning rates on the spatial grid based on spatial sensitivity analysis to enhance the learning ability far away from the hole area and improve the imaging effect of the cave in the low-sensitivity area.

[0008] According to some embodiments, the present invention adopts the following technical solutions:

[0009] A cave inversion and imaging method based on multi-scale unsupervised deep learning includes the following steps:

[0010] Construct multiple low-resistance geoelectric models and perform forward simulation to obtain observation data;

[0011] Build a deep learning network, construct a loss function with multi-scale and smooth constraints, calculate the model gradient, use the learning rate weighted constraint method to improve the model training ability in low-sensitivity areas, update the deep learning network parameters, and train the deep learning network to determine the mapping relationship between the observation data and the geoelectric model;

[0012] The trained deep learning network is used to process the collected observation data to obtain the corresponding geoelectric model map and realize inversion.

[0013] As an optional implementation method, a plurality of low-resistance geoelectric models are constructed, and the specific process of forward simulation to obtain observation data includes: according to the actual development form of the cave, a low-resistance geoelectric model with corresponding shape, size and spatial distribution is established;

[0014] For each low-resistance geoelectric model, cross-hole resistivity CT forward numerical simulation is performed to obtain observation data.

[0015] As an optional implementation method, a loss function containing multi-scale and smooth constraints is constructed. The specific steps for calculating the model gradient include: generating a prediction model through a neural network using the observed data, obtaining the predicted data after passing through the forward modeling network, calculating the data residual together with the observed data, calculating the model gradients based on multi-scale and smooth constraints respectively, and performing linear combinations.

[0016] As an optional implementation, the specific process of using the learning rate weighted constraint method to improve the model training capability of the low-sensitivity zone includes: establishing a learning rate weighted function, and in the two-dimensional cross-hole resistivity CT, considering that the sensitivity decays toward the middle in the form of a quadratic function, dividing the learning rate weighted function into two sections according to the spatial position of the resistivity model grid.

[0017] As an optional implementation method, the specific steps for updating the parameters of the deep learning network include: using a learning rate weighted constraint method to calculate the model gradient weighted by the learning rate, and obtaining a gradient that takes into account both stable convergence and boundary characterization for network training to update the parameters of the neural network.

[0018] As a further limited implementation, the smoothness constraint loss function is:

[0019] L 1 =(G(m)-d obs ) T (G(m)-d obs )+λ(Cm) T (Cm)

[0020] m is the matrix form of the predicted geoelectric model, d obs The matrix form of the observed data, G(·) represents the forward operator, λ is the regularization factor used to balance the weights of the data term and the model term, and C is the smoothness matrix.

[0021] As a further limited implementation, the multi-scale constraint loss function is:

[0022]

[0023] in, is the characteristic coefficient generated after wavelet transform, sgn(·) represents the sign function, C is the wavelet transform operator, Represents the sensitivity matrix in the feature domain, expressed as: γ is a vector extended by elements with value γ.

[0024] A karst cave inversion and imaging system based on multi-scale unsupervised deep learning, comprising:

[0025] A geoelectric model construction module is configured to construct multiple low-resistance geoelectric models and perform forward simulation to obtain observation data;

[0026] A multi-scale gradient training module is configured to build a deep learning network, construct a loss function with multi-scale and smooth constraints, calculate the model gradient, use the learning rate weighted constraint method to improve the model training ability in low-sensitivity areas, update the deep learning network parameters, and train the deep learning network to determine the mapping relationship between the observation data and the geoelectric model;

[0027] The inversion module is configured to process the collected observation data using the trained deep learning network to obtain the corresponding geoelectric model map and realize inversion.

[0028] 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.

[0029] 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.

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

[0031] To address the problem of poor resistivity model boundary characterization, the present invention introduces multi-scale inversion into deep learning, constructs a loss function containing multi-scale constraints and smooth constraints, and ensures stable convergence of the neural network in the early stage of training by balancing the weight factors of the multi-scale and smooth constraint gradients. In the later stage of training, the boundaries of the anomaly are characterized at multi-scale to achieve morphological characterization of more complex geoelectric models.

[0032] In order to solve the problem that low-sensitivity areas are difficult to accurately image, the present invention assigns differential learning rate weights to grid cells at different distances from the borehole, thereby enhancing the data mining and inversion capabilities of the inversion network for low-sensitivity areas and improving the overall imaging effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] 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.

[0034] Figure 1 is a flow chart of a karst cave inversion and imaging method based on multi-scale unsupervised deep learning proposed in this embodiment;

[0035] Figure 2 is a flowchart of network training based on multi-scale gradient calculation proposed in this embodiment;

[0036] Figure 3 is a schematic diagram of a geoelectric model in a database established in an embodiment;

[0037] Figure 4 The figure is a multi-scale unsupervised deep learning inversion result in an embodiment. DETAILED DESCRIPTION

[0038] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0039] 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.

[0040] 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.

[0041] This embodiment discloses a cave inversion and imaging method based on multi-scale unsupervised deep learning, such as Figure 1 As shown, the following steps are included:

[0042] (1) Construct multiple low-resistance geoelectric models, perform forward simulation to obtain observation data, and build an unsupervised deep learning database;

[0043] The identification method of this embodiment is mainly aimed at the low-resistance body of the karst cave in the underground engineering, which is represented by geoelectric models with different resistivity, different positions and different shapes;

[0044] The model size of this embodiment is 16m (X) × 32m (Z), the number of electrode points is 64, the electrode spacing is 1m, the survey line spacing is 14m, and the inversion grid size is 1m × 1m, that is, Figure 4 The size of each grid in the resistivity model is 200Ohm·m, the resistivity of the low-resistance anomaly is 20Ohm·m, Figure 3 The possible shapes of the anomalies are shown, including the arrangement and combination of "L", "T", step and rectangle. Among the single anomalies, there are 880 step-shaped, 640 "T" and "L" shapes; among the two anomalies, there are 2574 double step-shaped, 5354 double "T" shapes, and 2799 "T" + rectangle shapes. The database of this embodiment has a total of 4880 samples, 1280 samples of 1 rectangular block low-resistance body, and 3600 samples of two rectangular block low-resistance bodies. They are randomly divided into training set, verification set and test set in a ratio of 10:1:1.

[0045] (2) Inputting the observed data into the deep learning network, and generating a predictive resistivity model through the deep learning network;

[0046] (3) Forward simulation is performed on the predicted resistivity model to obtain the predicted data. Based on the predicted data, the observed data loss function is calculated to obtain the model gradient based on multi-scale and smooth constraints. After weighting using the learning rate, a new model gradient is obtained. The specific flow chart is as follows: Figure 2 shown.

[0047] The observation data in the sample library is used to generate a prediction model m through a neural network. After the forward modeling network, the prediction data d is obtained, and the data residual δd is calculated together with the observation data. The model gradients based on multi-scale and smooth constraints are calculated respectively, and linear combinations are performed. On this basis, the learning rate weighted constraint method is used to calculate the model gradient weighted by the learning rate, and the gradient that takes into account stable convergence and boundary characterization is obtained for network training to update the parameters of the neural network.

[0048] The smoothness constraint loss function is:

[0049] L 1 =(G(m)-d obs ) T (G(m)-d obs )+λ(Cm) T (Cm)

[0050] m is the matrix form of the predicted geoelectric model, d obsThe matrix form of the observed data, G(·) represents the forward operator, λ is the regularization factor, which is an empirical value when used to balance the weights of the data term and the model term, and C is the smoothness matrix.

[0051] The multi-scale constraint loss function is constructed as:

[0052]

[0053] Model terms in multi-scale constraint loss functions Using the one-norm and adding the damping factor μI for correction, the inversion equation in the wavelet domain is obtained:

[0054]

[0055] in, is the characteristic coefficient generated after wavelet transform, sgn(·) represents the sign function, C is the wavelet transform operator, Represents the sensitivity matrix in the feature domain, expressed as: γ is a vector extended by elements with value γ.

[0056] In this embodiment, the model gradient calculation formula based on multi-scale and smooth constraints is:

[0057]

[0058] in, Update the gradient of the characteristic parameters obtained by the inversion equation in the wavelet domain, δm 2 is the updated gradient of the spatial domain model parameters, where δd obs =f(m 0 )-d obs Represents the observed data residual. ω is the weight factor that balances the multi-scale and smooth constraint gradients.

[0059] In the early stage of training, ω needs to be taken as a small value to ensure stable convergence. On the basis of being able to accurately image the position and size of the anomaly, the ω value is increased to weaken the boundary blurring effect caused by the smooth constraint and to characterize the anomaly boundary at multiple scales.

[0060] W m δm is the model gradient weighted by the learning rate, where W m is the learning rate weighted matrix, expressed as follows:

[0061]

[0062] The size of the resistivity model in the inversion area is Nx(16m)×Ny(16m)×Nz(32m).

[0063] A learning rate weighting function suitable for unsupervised inversion grids is established. In 2D cross-hole resistivity CT, considering that the sensitivity usually decays toward the middle as a quadratic function, the weighting function is divided into two sections:

[0064]

[0065] Among them, x, z are the spatial positions of the resistivity model grid, and β1, β2 are empirical parameters. x Indicates the distance between the boreholes. In this embodiment, s x =14m. The resistivity model size of the inversion area is Nx(16m)×Nz(32m).

[0066] 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.

[0067] 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.

[0068] 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.

[0069] 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.

[0070] 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.

[0071] 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 cave inversion and imaging method based on multi-scale unsupervised deep learning, characterized in that: The following steps are involved: Construct multiple low-resistance geoelectric models and perform forward simulation to obtain observation data; Build a deep learning network, construct a loss function with multi-scale and smooth constraints, calculate the model gradient, use the learning rate weighted constraint method to improve the model training ability in low-sensitivity areas, update the deep learning network parameters, and train the deep learning network to determine the mapping relationship between the observation data and the geoelectric model; The trained deep learning network is used to process the collected observation data to obtain the corresponding geoelectric model map and realize inversion; The multi-scale constraint loss function is: Model terms in multi-scale constraint loss functions Use a norm and add a damping factor After correction, we get the inversion equation in the wavelet domain: in, The matrix form of the observation data is represents the forward operator, is the characteristic coefficient generated after wavelet transform, represents the symbolic function, is the wavelet transform operator, Represents the sensitivity matrix in the feature domain, expressed as: , is a value The vector of the elements of , To predict data, Update gradients for feature parameters.

2. The method for cave inversion and imaging based on multi-scale unsupervised deep learning according to claim 1, characterized in that: The specific process of constructing multiple low-resistance geoelectric models and performing forward simulation to obtain observation data includes: establishing low-resistance geoelectric models of corresponding shapes, sizes and spatial distributions according to the actual development morphology of the caves; For each low-resistance geoelectric model, cross-hole resistivity CT forward numerical simulation is performed to obtain observation data.

3. The method for cave inversion and imaging based on multi-scale unsupervised deep learning as claimed in claim 1, characterized in that: Construct a loss function with multi-scale and smooth constraints. The specific steps of calculating the model gradient include: generating a prediction model through a neural network using the observed data, obtaining the predicted data after the forward modeling network, calculating the data residual together with the observed data, calculating the model gradient based on multi-scale and smooth constraints respectively, and performing linear combination.

4. The method for cave inversion and imaging based on multi-scale unsupervised deep learning according to claim 1, characterized in that: The specific process of using the learning rate weighted constraint method to improve the model training capability of the low-sensitivity area includes: establishing a learning rate weighted function, and in the two-dimensional cross-hole resistivity CT, considering that the sensitivity decays toward the middle in the form of a quadratic function, dividing the learning rate weighted function into two sections according to the spatial position of the resistivity model grid.

5. The method for cave inversion and imaging based on multi-scale unsupervised deep learning according to claim 1, characterized in that: The specific steps for updating the parameters of the deep learning network include: using the learning rate weighted constraint method to calculate the model gradient weighted by the learning rate, and obtaining the gradient that takes into account both stable convergence and boundary characterization for network training to update the parameters of the neural network.

6. The method for cave inversion and imaging based on multi-scale unsupervised deep learning according to claim 1, characterized in that: The smoothness constraint loss function is: is the matrix form of the predicted geoelectric model, The matrix form of the observation data is represents the forward operator, is a regularization factor used to balance the weights of data terms and model terms, is the smoothness matrix.

7. A karst cave inversion and imaging system based on multi-scale unsupervised deep learning, characterized in that: include: A geoelectric model construction module is configured to construct multiple low-resistance geoelectric models and perform forward simulation to obtain observation data; A multi-scale gradient training module is configured to build a deep learning network, construct a loss function with multi-scale and smooth constraints, calculate the model gradient, use the learning rate weighted constraint method to improve the model training ability in low-sensitivity areas, update the deep learning network parameters, and train the deep learning network to determine the mapping relationship between the observation data and the geoelectric model; An inversion module is configured to process the collected observation data using the trained deep learning network to obtain a corresponding geoelectric model map and realize inversion; The multi-scale constraint loss function is: Model terms in multi-scale constraint loss functions Use a norm and add a damping factor After correction, we get the inversion equation in the wavelet domain: in, The matrix form of the observation data is represents the forward operator, is the characteristic coefficient generated after wavelet transform, represents the symbolic function, is the wavelet transform operator, Represents the sensitivity matrix in the feature domain, expressed as: , is a value The vector of the elements of , To predict data, Update gradients for feature parameters.

8. 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 6.

9. 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 6.

Images