Solitary rock imaging method and system based on deep learning of cross-hole resistivity CT

Abstract

The present invention provides a boulder imaging method and system based on deep learning of cross-hole resistivity CT. Aiming at the characteristics of non-image data under the cross-hole resistivity CT observation mode, a network architecture for feature extraction and imaging is established; drawing on the reference model commonly used in traditional inversion methods, a loss function guided by the reference model is constructed, and the background resistivity is used for guidance to ensure that the network is updated in the direction of generating a real model during training, and the mapping relationship between the observed data and the resistivity model is learned more accurately. The adaptive single time step multiple superposition calculation gradient is used to improve the accuracy of the gradient in the network and thus improve the imaging effect.

Description

Technical Field

[0001] The present invention belongs to the technical field of deep imaging, and relates to a boulder imaging method and system based on cross-hole resistivity CT 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] During subway construction, unfavorable geological structures such as boulders are often encountered during the construction period. If they are not discovered in advance, disasters such as machine jams may occur, delaying construction progress and even causing casualties. In order to ensure construction safety, it is necessary to accurately image and locate the boulders.

[0004] The cross-hole resistivity CT imaging method is to place detection electrodes in the borehole, establish an artificial electric field through the power supply electrode, perform encrypted sampling on the measuring electrode, and use the inversion technology to reconstruct the sampled data into a resistivity model. Traditional linear inversion technology is prone to fall into local optimality, resulting in erroneous data interpretation. Deep learning methods have super learning memory and nonlinear fitting capabilities, and have been widely studied in geophysical inversion imaging.

[0005] Existing cross-hole resistivity CT deep learning imaging methods all belong to supervised learning networks, that is, the training set includes observed data (potential / apparent resistivity) and labels (resistivity models matching the observed data), but it is difficult to obtain available labels (real resistivity models) in actual detection. This leads to a lack of available labels in the training set of supervised learning imaging methods, the network cannot be effectively trained, and it is difficult to obtain good imaging results.

[0006] According to the inventors' understanding, the application of unsupervised learning imaging methods to cross-hole resistivity CT detection faces the following problems:

[0007] (1) The potential and apparent resistivity data obtained by cross-hole resistivity CT detection are in non-image format, and it is impossible to establish a spatial correspondence with the resistivity model, making it difficult to extract features.

[0008] (2) The resistivity model generated by inversion of observation data has multiple solutions. If the network training is driven only by physical laws, it is easy to produce erroneous gradients, resulting in poor network imaging effect and difficulty in meeting engineering detection needs. Summary of the invention

[0009] In order to solve the above problems, the present invention proposes a boulder imaging method and system based on deep learning of cross-hole resistivity CT. The present invention establishes a network architecture for feature extraction and imaging for non-image data (potential / apparent resistivity) under the cross-hole resistivity CT observation mode. On this basis, the reference model is used to optimize the loss function, and the adaptive single time step multiple superposition calculation gradient is adopted to improve the accuracy of the gradient in the network and thus improve the imaging effect.

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

[0011] A method for solitary rock imaging based on deep learning of cross-hole resistivity CT includes the following steps:

[0012] A geoelectric model was established based on the resistivity model of historical boulders, and a database for unsupervised deep learning was established through cross-hole resistivity CT forward simulation;

[0013] Input the non-image data in the database into the deep learning network, obtain the feature map through the encoder network, use the convolution layer to output the predicted resistivity model, add the forward modeling module at the output end, calculate the forward modeling result of the predicted resistivity model, and train the network parameters by fitting with the input non-image data to form an unsupervised deep learning network;

[0014] A loss function guided by a reference model is constructed, and a gradient is calculated by multiple superpositions of an adaptive single time step. The unsupervised deep learning network is trained with data in a database so that the network can fit the mapping relationship between the observed data and the resistivity model.

[0015] The observation data of the target detection area is obtained, and the prediction model corresponding to the data is obtained using the trained unsupervised deep learning network to achieve imaging of the boulders in the detection area.

[0016] As an optional implementation method, a geoelectric model is established based on the resistivity model of the historical boulders, and the specific process of establishing an unsupervised deep learning database through cross-hole resistivity CT forward simulation includes:

[0017] The development data of common boulders are obtained, and high-resistance geoelectric models with various shapes, sizes and position distributions are constructed. For various geoelectric models, any two electrodes in the hole are powered, and the potential values ​​and apparent resistivity values ​​at other electrodes are calculated to obtain the potential and apparent resistivity data of the electrode points in the hole.

[0018] As an optional implementation, non-image data in the database is input into a deep learning network, a feature map is obtained through an encoder network, and a specific process of using a convolutional layer to output a predicted resistivity model includes:

[0019] The potential data and apparent resistivity data are input in a batch processing manner, and the neighborhood features are extracted by convolution transformation of the borehole apparent resistivity data. The data is then concatenated with the borehole potential data. The data is then mapped into a feature map that is spatially aligned with the prediction model through a fully connected layer, and the corresponding prediction resistivity model is obtained from the feature map through a convolution layer.

[0020] As an optional implementation, the steps of constructing a loss function guided by a reference model are:

[0021] The reference model term used to guide the inversion process is imposed in the loss function, and the calculation formula is:

[0022]

[0023] Where f(·) represents forward modeling, d obs is the observed data, m is the matrix form of the predicted geoelectric model, C is the smoothness matrix, m ref represents the matrix form of the reference geoelectric model, F is the weight matrix of the reference model, and η is the regularization factor used to balance the weights of the reference model terms with the data terms and model terms.

[0024] As a further example, the regularization factor calculation formula is as follows:

[0025]

[0026] Among them, η 0 is the initial value of the regularization parameter, epoch th The reference model abort threshold set for the network.

[0027] As an optional implementation, the specific process of calculating the gradient by using adaptive single time step multiple superpositions includes:

[0028] The observed data generates a prediction model m through a neural network, and the predicted potential data d is obtained through the forward modeling network. The model increment δm is obtained through calculation, and the data root mean square RMS is calculated; if the convergence condition is met, the loss function is calculated and the gradient is fed back to the network;

[0029] If RMS does not meet the convergence condition, update the current model m 2 =m+δm, and based on this, calculate the additional amount δm 2 ;

[0030] If the new RMS still does not converge, the update process is repeated, otherwise, all model increments are superimposed to update the network together.

[0031] As a further limited implementation, in the adaptive single time step multiple gradient superposition calculation process, a convergence criterion for multiple iteration termination is set:

[0032] When the data root mean square of the i-th iteration is greater than the set value, and the ratio of the difference between the data root mean square of the previous iteration and the data root mean square of this iteration is greater than the threshold, it does not converge;

[0033] When the RMS value of the data of the i-th iteration is greater than the set value, and the ratio of the difference between the RMS value of the data of the previous iteration and the RMS value of the data of this iteration is less than or equal to the threshold, it converges;

[0034] When the RMS value of the data at the i-th iteration is less than or equal to the set value, the algorithm converges.

[0035] A solitary rock imaging system based on deep learning of cross-hole resistivity CT, comprising:

[0036] The database construction module is configured to establish a geoelectric model based on the resistivity model of the historical boulders and to establish a database for unsupervised deep learning through cross-hole resistivity CT forward simulation;

[0037] An unsupervised deep learning network building module is configured to input non-image data in a database into a deep learning network, obtain a feature map through an encoder network, output a predicted resistivity model using a convolutional layer, add a forward modeling module at the output end, calculate a forward modeling result of the predicted resistivity model, train network parameters by fitting with the input non-image data, and form an unsupervised deep learning network;

[0038] A training module is configured to construct a loss function guided by a reference model, adopt an adaptive single time step multiple superposition calculation gradient, and train the unsupervised deep learning network with data in the database, so that the network can fit the mapping relationship between the observed data and the resistivity model;

[0039] The imaging module is configured to obtain observation data of the target detection area, and use the trained unsupervised deep learning network to obtain a prediction model corresponding to the data to achieve imaging of the boulders in the detection area.

[0040] 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.

[0041] 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.

[0042] Compared with the prior art, the present invention has the following beneficial effects:

[0043] The present invention aims at the characteristics of non-image data (potential / apparent resistivity) in the cross-hole resistivity CT observation mode, and establishes a network architecture for feature extraction and imaging;

[0044] The present invention draws on the reference model commonly used in traditional inversion methods, constructs a loss function guided by the reference model, and uses background resistivity as a guide to ensure that the network is updated in the direction of generating a real model during training, and more accurately learns the mapping relationship between the observed data and the resistivity model.

[0045] The present invention utilizes adaptive single time step multiple superposition gradient calculation, and the gradient of multiple iterative superposition is more accurate than that of a single time step, thereby ensuring the accuracy of neural network gradient update and alleviating the problem of many false anomalies. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] 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.

[0047] Figure 1 A flow chart of a boulder imaging method based on unsupervised deep learning of cross-hole resistivity CT proposed in this embodiment;

[0048] Figure 2 This is a schematic diagram of the encoder network proposed in this embodiment;

[0049] Figure 3 This is a flowchart of network training based on new gradient calculation proposed in this embodiment;

[0050] Figure 4 The unsupervised deep learning inversion result in one embodiment. DETAILED DESCRIPTION

[0051] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0052] 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.

[0053] 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.

[0054] This embodiment discloses a solitary rock imaging method based on unsupervised deep learning of cross-hole resistivity CT. Figure 1 As shown, the following steps are included:

[0055] Step S1, establishing a geoelectric model based on a resistivity model of common boulders;

[0056] The identification method of this embodiment is mainly aimed at isolated high-resistance rocks in underground engineering, which are represented here by geoelectric models with different resistivities, different positions and different shapes.

[0057] In this embodiment, according to the actual development of common boulders, high-resistance geoelectric models with various shapes, sizes and position distributions are constructed. For various geoelectric models, any two electrodes in the hole are powered, and the potential values ​​and apparent resistivity values ​​at other electrodes are calculated to obtain the potential and apparent resistivity data of the electrode points in the hole.

[0058] 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 for the background resistivity and 1000Ohm·m for the high-resistance anomaly.

[0059] The sample library of this embodiment has a total of 4880 samples, including 1280 samples of one rectangular block high-resistance body and 3600 samples of two rectangular block high-resistance bodies. The samples are randomly divided into training set, validation set and test set in a ratio of 10:1:1 (training set: 4066; validation set: 407; test set: 407).

[0060] Step S2, constructing a database by computer numerical simulation, wherein the database includes multiple sets of electrode point potential and apparent resistivity data in the hole;

[0061] This embodiment adopts a combination of three typical quadrupole observation devices, namely Bipole-Bipole, Dipole-Dipole and Pole-Tripole, with a data volume of 49,280, including 18,368 of Bipole-Bipole, 18,368 of Dipole-Dipole and 12,544 of Pole-Tripole. They are grouped by power supply point A, with a total of 64 groups, each with 770 data.

[0062] Step S3, inputting the non-image data into the deep learning network, obtaining a feature map through the encoder network, and then generating a predicted resistivity model through the convolution layer;

[0063] The non-image data of this embodiment includes potential data and apparent resistivity data, which are used as inputs of the unsupervised deep learning network. The encoder performs convolution transformation on the borehole apparent resistivity data to extract neighborhood features, and then splices it with the borehole potential data. Then, the data is mapped into a feature map that is aligned with the prediction model space through a fully connected layer, and the corresponding prediction resistivity model is obtained from the feature map through a convolution layer.

[0064] A forward modeling module is added at the output end to calculate the forward modeling results of the predicted resistivity model and train the network parameters by fitting with the input non-image data. The forward modeling module is established according to the physical laws of electric field propagation and converts the prediction model into prediction data.

[0065] Step S4, forward modeling the predicted resistivity model to obtain predicted potential and apparent resistivity data, and calculate the observed data loss function;

[0066] In this embodiment, a reference model term for guiding the inversion process is applied to the loss function, and the calculation formula is:

[0067]

[0068] Where f(·) represents forward modeling, d obs is the observed data, m is the matrix form of the predicted geoelectric model, C is the smoothness matrix, m ref represents the matrix form of the reference geoelectric model, F is the weight matrix of the reference model, and in this embodiment, F is set as the unit matrix. η is the regularization factor, which is used to balance the weights of the reference model terms, the data terms, and the model terms. The calculation formula is as follows:

[0069]

[0070] Among them, η 0 is the initial value of the regularization parameter, which can usually be 0.05 to 1. In this embodiment, it is 0.1. th The reference model termination threshold set for the network is selected as 30 times in this embodiment.

[0071] Step S5, adaptively superimpose multiple gradients in a single time step to calculate the gradient and update the network;

[0072] First, the observed data generates a prediction model m through a neural network, and the forward modeling result d of the prediction model is obtained through the forward modeling network. The model increment δm is calculated, and the data root mean square RMS is calculated. If the convergence condition is met, the loss function is calculated and the gradient is returned to the network. If the RMS does not meet the convergence condition, the current model m is updated. 2 =m+δm, and based on this, calculate the additional amount δm 2 If the new RMS still does not converge, repeat the process. Otherwise, add all model increments to update the network together.

[0073] For the adaptive single time step multiple gradient superposition calculation method, the convergence condition for multiple iterations to terminate is as follows:

[0074]

[0075] RMS i represents the RMS value of the i-th iteration, R up It is another indicator of convergence. That is, in the late stage of network training, the RMS of some samples has converged to a lower level, and a single search can meet the network update requirements.

[0076] Step S6, training the network with the database so that the network can fit the mapping relationship between the observed data and the resistivity model;

[0077] Substituting some of the results of the test set into Figure 4 As shown in the figure, it can be seen that the resistivity value of the target area can usually reach more than 800Ohm·m. Against the background of 200Ohm·m, it can be clearly shown that there is a high-resistance anomaly in the area. This embodiment effectively improves the accuracy of cross-hole resistivity CT imaging of boulders. For a single block-shaped high-resistance anomaly and two block-shaped high-resistance anomalies, the volume, position and resistivity value can be imaged more accurately.

[0078] An embodiment of a boulder imaging system based on cross-hole resistivity CT deep learning is also provided.

[0079] In this embodiment, the system includes:

[0080] The database construction module is configured to establish a geoelectric model based on the resistivity model of the historical boulders and to establish a database for unsupervised deep learning through cross-hole resistivity CT forward simulation;

[0081] An unsupervised deep learning network building module is configured to input non-image data in a database into a deep learning network, obtain a feature map through an encoder network, output a predicted resistivity model using a convolutional layer, add a forward modeling module at the output end, calculate a forward modeling result of the predicted resistivity model, train network parameters by fitting with the input non-image data, and form an unsupervised deep learning network;

[0082] A training module is configured to construct a loss function guided by a reference model, adopt an adaptive single time step multiple superposition calculation gradient, and train the unsupervised deep learning network with data in the database, so that the network can fit the mapping relationship between the observed data and the resistivity model;

[0083] The imaging module is configured to obtain observation data of the target detection area, and use the trained unsupervised deep learning network to obtain a prediction model corresponding to the data to achieve imaging of the boulders in the detection area.

[0084] 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.

[0085] 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.

[0086] 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.

[0087] 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.

[0088] 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.

[0089] 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 solitary rock imaging method based on deep learning of cross-hole resistivity CT, Its characteristics are: The following steps are involved: A geoelectric model was established based on the resistivity model of historical boulders, and a database for unsupervised deep learning was established through cross-hole resistivity CT forward simulation; Input the non-image data in the database into the deep learning network, obtain the feature map through the encoder network, use the convolution layer to output the predicted resistivity model, add the forward modeling module at the output end, calculate the forward modeling result of the predicted resistivity model, and train the network parameters by fitting with the input non-image data to form an unsupervised deep learning network; A loss function guided by a reference model is constructed, and a gradient is calculated by multiple superpositions of an adaptive single time step. The unsupervised deep learning network is trained with data in a database so that the network can fit the mapping relationship between the observed data and the resistivity model. The observation data of the target detection area is obtained, and the prediction model corresponding to the data is obtained using the trained unsupervised deep learning network to achieve imaging of the boulders in the detection area.

2. A boulder imaging method based on deep learning of cross-hole resistivity CT according to claim 1, Its characteristics are: The specific process of establishing a geoelectric model based on the resistivity model of the historical boulders and building an unsupervised deep learning database through cross-hole resistivity CT forward simulation includes: The development data of common boulders are obtained, and high-resistance geoelectric models with various shapes, sizes and position distributions are constructed. For various geoelectric models, any two electrodes in the hole are powered, and the potential values ​​and apparent resistivity values ​​at other electrodes are calculated to obtain the potential and apparent resistivity data of the electrode points in the hole.

3. A boulder imaging method based on deep learning of cross-hole resistivity CT according to claim 1, Its characteristics are: The specific process of inputting non-image data in the database into the deep learning network, obtaining the feature map through the encoder network, and using the convolution layer to output the predicted resistivity model includes: The potential data and apparent resistivity data are input in a batch processing manner, and the neighborhood features are extracted by convolution transformation of the borehole apparent resistivity data. The data is then concatenated with the borehole potential data. The data is then mapped into a feature map that is spatially aligned with the prediction model through a fully connected layer, and the corresponding prediction resistivity model is obtained from the feature map through a convolution layer.

4. A boulder imaging method based on deep learning of cross-hole resistivity CT according to claim 1, Its characteristics are: The steps to construct a loss function guided by the reference model are: The reference model term used to guide the inversion process is imposed in the loss function, and the calculation formula is: Where f(·) represents forward modeling, d obs is the observed data, m is the matrix form of the predicted geoelectric model, C is the smoothness matrix, m ref represents the matrix form of the reference geoelectric model, F is the weight matrix of the reference model, and η is the regularization factor used to balance the weights of the reference model terms with the data terms and model terms.

5. A boulder imaging method based on deep learning of cross-hole resistivity CT according to claim 4, Its characteristics are: The regularization factor calculation formula is as follows: Among them, η 0 is the initial value of the regularization parameter, epoch th The reference model abort threshold set for the network.

6. A boulder imaging method based on deep learning of cross-hole resistivity CT according to claim 1, Its characteristics are: The specific process of using adaptive single time step multiple superposition to calculate the gradient includes: The observed data generates a prediction model m through a neural network, and the predicted potential data d is obtained through the forward modeling network. The model increment δm is obtained through calculation, and the data root mean square RMS is calculated; if the convergence condition is met, the loss function is calculated and the gradient is fed back to the network; If the RMS does not meet the convergence condition, update the current model m 2 = m + δm, and calculate the new increment δm based on this 2 ; If the new RMS still does not converge, the update process is repeated, otherwise, all model increments are superimposed to update the network together.

7. A boulder imaging method based on deep learning of cross-hole resistivity CT according to claim 6, Its characteristics are: In the adaptive single time step multiple gradient superposition calculation process, multiple iteration termination convergence criteria are set: When the data root mean square of the i-th iteration is greater than the set value, and the ratio of the difference between the data root mean square of the previous iteration and the data root mean square of this iteration is greater than the threshold, it does not converge; When the RMS value of the data of the i-th iteration is greater than the set value, and the ratio of the difference between the RMS value of the data of the previous iteration and the RMS value of the data of this iteration is less than or equal to the threshold, it converges; When the RMS value of the data at the i-th iteration is less than or equal to the set value, the algorithm converges.

8. A solitary rock imaging system based on deep learning of cross-hole resistivity CT, Its characteristics are: include: The database construction module is configured to establish a geoelectric model based on the resistivity model of the historical boulders and to establish a database for unsupervised deep learning through cross-hole resistivity CT forward simulation; An unsupervised deep learning network building module is configured to input non-image data in a database into a deep learning network, obtain a feature map through an encoder network, output a predicted resistivity model using a convolutional layer, add a forward modeling module at the output end, calculate a forward modeling result of the predicted resistivity model, train network parameters by fitting with the input non-image data, and form an unsupervised deep learning network; A training module is configured to construct a loss function guided by a reference model, adopt an adaptive single time step multiple superposition calculation gradient, and train the unsupervised deep learning network with data in the database, so that the network can fit the mapping relationship between the observed data and the resistivity model; The imaging module is configured to obtain observation data of the target detection area, and use the trained unsupervised deep learning network to obtain a prediction model corresponding to the data to achieve imaging of the boulders in the detection area.

9. A computer-readable storage medium, Its characteristics are: 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, Its characteristics are: 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