Tunnel resistivity deep learning inversion method based on initial model constraint

Abstract

The application discloses a tunnel three-dimensional resistivity deep learning inversion method based on initial model constraint, and aims at the problems that the existing tunnel three-dimensional resistivity inversion result is easy to fall into local optimum, iteration is not easy to converge, and the generalization performance of the prediction result is insufficient by only adopting the forward data as the input value of the deep learning network, the resistivity prediction network is constructed, the initial model obtained by converting the forward data and the seismic exploration is simultaneously used as the input end of the resistivity prediction network, and the resistivity geological model is used as a label to perform training. Compared with the conventional deep learning resistivity inversion method by only using the forward data, the method can incorporate certain prior geological information through the constraint of the initial model, so that the stability and precision of the detection data processing are improved, and the generalization capability of the inversion is effectively improved.

Description

TECHNICAL FIELD

[0001] The present disclosure belongs to the field of geotechnical engineering exploration, and relates to a tunnel resistivity deep learning inversion method based on initial model constraint. BACKGROUND

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

[0003] The resistivity method is an electrical prospecting method based on the difference in electrical properties of rocks and minerals, which solves engineering, environmental, disaster and other geological problems by observing the distribution of artificially established underground stable current field. In recent years, it has been widely used in the fields of mineral exploration, oil and gas exploration and groundwater exploration. The resistivity inversion method can reconstruct the observed data into a resistivity model, which can realize the shape description and spatial positioning of the abnormal area in front of the tunnel face.

[0004] According to the inventors, the direct current resistivity method is an economical and efficient geophysical exploration method, which is simple to operate, sensitive to water-bearing structures and other characteristics, and is the mainstream method for detecting adverse geological structures. However, for tunnel resistivity detection data inversion, the inversion result is easy to fall into local optimum and prone to false geological interpretation. By incorporating deep learning algorithm into resistivity inversion, the accuracy of the inversion result can be improved. However, the current deep learning-based resistivity inversion only uses resistivity forward simulation data as input in the training stage, lacks the generality of scene external sample data, and often leads to insufficient generalization performance of the deep learning network prediction result. SUMMARY

[0005] To solve the above problems, the present disclosure provides a tunnel resistivity deep learning inversion method based on initial model constraint. The present disclosure is aimed at the problems that the existing three-dimensional resistivity inversion result of the tunnel is easy to fall into local optimum, the iteration is not easy to converge, and the use of forward simulation data as the input value of the deep learning network easily leads to insufficient generalization performance of the prediction result. By converting the stratum information obtained by seismic inversion into resistivity distribution and updating the resistivity value of the abnormal area by combining the simulated annealing method, the updated resistivity model is obtained as the initial model. By inputting the resistivity forward data and the initial model into the deep learning network at the same time and training with the resistivity geological model as the label, the stability and accuracy of the resistivity inversion imaging of the detection area in front of the tunnel face can be effectively improved.

[0006] To this end, the present disclosure adopts the following technical solutions:

[0007] A tunnel resistivity deep learning inversion method based on initial model constraint comprises the following steps:

[0008] Step S1: According to the predetermined region of the tunnel engineering exploration area, a predetermined number of three-dimensional resistivity models are constructed according to the geological conditions, and resistivity observation data corresponding to each model is generated through resistivity forward modeling to form a sample training database;

[0009] Step S2: The wave velocity information of the above-mentioned constructed geological model is obtained by using the seismic exploration method, and the stratum distribution in front of the tunnel face is obtained by inversion;

[0010] Step S3: The resistivity distribution of the inversion region is obtained by using the correlation expression between resistivity and seismic wave velocity, the resistivity value of the region with potential geological risk obtained by conversion is adjusted within a preset reference value range to obtain a resistivity perturbation interval, then the resistivity value in the perturbation interval is refined and adjusted by the simulated annealing method, and the updated resistivity model is used as the initial model in the subsequent training process to provide prior constraints for the minimization process of the objective function;

[0011] Step S4: A resistivity prediction network is constructed, the forward modeling data and the initial model are simultaneously used as the input end of the network, the resistivity model corresponding to the geological structure with potential geological risk is used as the label for training, and the resistivity distribution result of the geological model in front of the tunnel face is output.

[0012] Further, step S3 includes the following steps:

[0013] Step S31: Linear regression analysis is applied to the constructed geological model to derive the correlation expression of resistivity-seismic wave velocity:

[0014] ln(lnρ)=k(v-h)

[0015] Wherein, v is the seismic wave velocity, ρ is the resistivity, k is the resistivity-seismic wave velocity correlation coefficient, and h is a constant;

[0016] Step S32: Identify the resistivity abnormal value in the region obtained by conversion, and adjust it within a preset reference value range to obtain a resistivity perturbation interval [ρ a , ρ b ];

[0017] Step S33: Use the simulated annealing algorithm (VFSA) for global optimization to adjust the obtained resistivity data to find the optimal solution.

[0018] Further, step S33 includes the following steps:

[0019] Step S331: According to the resistivity perturbation interval [ρ a , ρ b ], randomly perturb the current resistivity data to obtain a new resistivity value:

[0020] P i+1 = p i + t(p b - p a )

[0021] In the formula, p i+1 is a new resistivity value, t is a random disturbance coefficient, and the range is set to 0.8 to 0.99, and is determined according to specific geological conditions and expected inversion accuracy requirements;

[0022] Step S332: According to the Metropolis criterion in the simulated annealing algorithm, it is judged whether to accept the new iteration resistivity value as the current solution, until there is no acceptable new resistivity value for continuous multiple iterations, or the value of the objective function has reached or is lower than the preset threshold value, then the simulated annealing process is completed.

[0023] Further, step S4 comprises the following steps:

[0024] Step S41: The updated resistivity model is taken as an initial model m0, and the initial model m0 and the forward observation data are taken as network inputs at the same time, and the constructed geological model is taken as a label to perform training;

[0025] Step S42: A full convolutional neural network (FCN) architecture is used to construct a resistivity prediction network, and the optimization process of the network is the minimization process of the objective function, and the expression of the objective function is:

[0026]

[0027] In the formula, Φ is the objective function of resistivity inversion, C is a smoothness matrix, which is used to make the change of the model parameters of adjacent grids smooth. A is a partial derivative matrix, and the elements of the matrix are the partial derivatives of the model parameters with respect to the theoretical observation data of the model. Δd is the difference between the theoretical observation data obtained by the forward and the measured observation data, Δm is the model parameter update amount, and λ is the Lagrange constant.

[0028] Further, the resistivity prediction is further configured to: the seismic data obtained by seismic exploration and the resistivity distribution converted based on the rock physical property correlation formula are combined with the initial model updated by the simulated annealing method as input data to directly predict the resistivity distribution result of the inversion area, and the rock physical property correlation formula is fitted by logging drilling data.

[0029] Compared with the prior art, the present application has the following beneficial effects:

[0030] The application aims at the problems that the conventional tunnel three-dimensional resistivity inversion result is easy to fall into local optimum, iteration is not easy to converge, and the generalization performance of the predicted result by using only the forward simulation data as the input value of the deep learning network, the data generated by the resistivity model forward and the three-dimensional initial model obtained by converting the seismic exploration are used as input information, the full convolution neural network structure is used to extract deep features of the data, and a three-dimensional tunnel resistivity prediction model is generated, so that the precision and generalization ability of the resistivity inversion imaging of the tunnel face front detection area are effectively improved.

[0031] In order to make the above objectives, characteristics and advantages of the present application more apparent, clear and easy to understand, the following preferred embodiments are specifically described below, and the accompanying drawings are used for reference. BRIEF DESCRIPTION OF DRAWINGS

[0032] The drawings constituting a part of the disclosure are used to provide further understanding of the disclosure, the illustrative embodiments of the disclosure and the description thereof are used to explain the disclosure, and do not constitute improper limitation on the disclosure.

[0033] Figure 1 It is a tunnel resistivity inversion flowchart based on an initial model constraint.

[0034] Figure 2 It is a resistivity prediction network structure diagram.

[0035] Figure 3 It is a three-dimensional resistivity inversion imaging structure schematic diagram of the application.

[0036] Figure 4 It is a three-dimensional resistivity inversion imaging result schematic diagram of the application. DETAILED DESCRIPTION

[0037] The disclosure will be further described below in combination with the drawings and embodiments.

[0038] It should be pointed out that the following detailed description is all exemplary, and aims to provide further description of the disclosure. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as generally understood by those skilled in the art to which the disclosure belongs.

[0039] The embodiment discloses a tunnel three-dimensional resistivity deep learning inversion method based on an initial model constraint, and the flow is as shown in Figure 1 The following steps are included:

[0040] In the initial model constraint based on the tunnel three-dimensional resistivity deep learning inversion method, the three-dimensional initial model is obtained by converting the seismic exploration data, and the three-dimensional initial model is used as the input of the deep learning network, so that the three-dimensional tunnel resistivity prediction model is generated by using the full convolution neural network structure to extract the deep features of the data, and the precision and generalization ability of the tunnel face front detection area are effectively improved. Figure 3The tunnel engineering exploration area with an abnormal area is shown. 100 inversion area sizes of 30 m x 30 m x 30 m (x x y x z, the tunnel section direction is x x z, the horizontal direction is x, the vertical direction is z, and the tunnel direction is y) are constructed in the tunnel engineering exploration area. The sizes, spatial positions, and shapes of the abnormal bodies in different models are different. The resistivity of the abnormal body is set to 10 ohm m, and the background resistivity of the geological model is set to 1000 ohm m. The velocity of the abnormal body is set to 3.9 km / s -1 , and the velocity of the geological model is set to 6.7 km / s -1 .

[0041] For the above 100 geological models, first, they are respectively divided into grid sizes of dx = dy = dz = 2 m, and the resistivity observation data generated by the forward modeling are taken as a training database; the distribution map of the seismic wave velocity can be obtained by performing seismic inversion, and by referring to the correlation formula between resistivity and velocity in the existing literature, and combining the linear regression method, the resistivity-velocity physical value correlation formula can be obtained as:

[0042] ln (ln p) = 0.396 x (v-1.792)

[0043] The resistivity disturbance interval of the abnormal body is set to [5, 20]. For the grid physical values in the abnormal body area, the simulated annealing method is used to update the grid parameters in the abnormal body area:

[0044] m i+1,j = m i,j + 0.9 x (20-15)

[0045] In the formula, j represents the jth model parameter, and i represents the ith disturbance. If there is no new resistivity value that can be updated or the set threshold is reached in several iterations, the simulated annealing process is completed; then, the updated resistivity model is taken as an initial model.

[0046] The forward observation data obtained by the forward modeling and the initial model are simultaneously taken as the input of the resistivity prediction network, and the geological model is taken as a label for training, and the network structure is as shown in Figure 2 . The network training process is a target function minimization process, and the grid parameter is updated by solving the inversion equation, and the training process is performed for a total of 200 rounds. The final inversion result is as shown in Figure 4 . It can be seen that for the distribution of the abnormal body in front of the tunnel face, the inversion method described in the application can clearly depict the position and shape of the low-resistivity abnormal body.

[0047] The above merely describes preferred embodiments of the present disclosure and is not intended to limit the present disclosure. The present disclosure can have various modifications and changes for those skilled in the art. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present disclosure shall be included in the protection scope of the present disclosure.

[0048] Although the specific embodiments of the present disclosure are described above with reference to the accompanying drawings, the present disclosure is not limited thereto, and those skilled in the art should understand that various modifications or changes made on the basis of the technical solutions of the present disclosure without creative labor are still within the protection scope of the present disclosure.

Claims

1. A deep learning inversion method for tunnel resistivity based on initial model constraints, characterized in that, Includes the following steps: Step S1: Construct a predetermined number of three-dimensional resistivity models based on the geological conditions of the predetermined area of ​​the tunnel engineering exploration area. Through resistivity forward modeling, generate resistivity observation data corresponding to each model to form a sample training database. Step S2: Use seismic detection methods to obtain the wave velocity information of the geological model constructed above, and obtain the stratigraphic distribution in front of the tunnel face through inversion; Step S3: Using the correlation expression between resistivity and seismic wave velocity, the resistivity distribution of the inversion area is obtained. The resistivity anomalies in the transformed area are identified and adjusted within a preset benchmark range to obtain a resistivity perturbation range. Subsequently, the resistivity values ​​in the perturbation range are refined and adjusted using the simulated annealing method. The updated resistivity model is used as the initial model in the subsequent training process to provide prior constraints for the objective function minimization process. Step S4: Construct a resistivity prediction network. By using forward modeling data and the initial model as inputs to the network, and using the resistivity model corresponding to the geological structure with potential geological risks as a label for training, the network outputs the resistivity distribution results of the geological model in front of the tunnel face.

2. The deep learning inversion method for tunnel resistivity based on initial model constraints according to claim 1, characterized in that, Step S3 includes the following steps: Step S31: Apply linear regression analysis to the constructed geological model to derive the correlation expression between resistivity and seismic wave velocity: ln(lnρ)=k(vh) In the formula, v is the seismic wave velocity, ρ is the resistivity, k is the resistivity-seismic wave velocity correlation coefficient, and h is a constant; Step S32: Identify the resistivity anomalies in the converted region and adjust them within a preset reference range to obtain a resistivity disturbance range [ρ]. a , ρ b ]; Step S33: Use the simulated annealing algorithm for global optimization, adjust the obtained resistivity data, and find the optimal solution.

3. The deep learning inversion method for tunnel resistivity based on initial model constraints according to claim 2, characterized in that, Step S33 includes the following steps: Step S331: Based on the resistivity perturbation range [ρ a , ρ b By randomly perturbing the current resistivity data, a new resistivity value is obtained: r i+1 =ρ i +t(p b -r a ) In the formula, ρ i+1 The new resistivity value is given, and t is the random perturbation coefficient, which is set to a range of 0.8 to 0.99, depending on the specific geological conditions and the expected inversion accuracy requirements. Step S332: Determine whether to accept the resistivity value of the new iteration as the current solution according to the Metropolis criterion in the simulated annealing algorithm. Continue until there are no acceptable new resistivity values ​​in multiple consecutive iterations, or the value of the objective function has reached or fallen below the preset threshold, then the simulated annealing process is complete.

4. The deep learning inversion method for tunnel resistivity based on initial model constraints according to claim 1, characterized in that, Step S4 includes the following steps: Step S41: Use the updated resistivity model as the initial model m0, and use the initial model m0 and forward modeling observation data as network inputs, with the constructed geological model as the label, for training. Step S42: Construct a resistivity prediction network using a fully convolutional neural network architecture. The optimization process of the network is the minimization process of the objective function, which is expressed as: In the formula, Φ is the objective function for resistivity inversion, C is the smoothness matrix, A is the Jacobian matrix, Δd is the difference between the theoretical observation data and the measured observation data obtained from forward modeling, Δm is the model parameter update amount, and λ is the Lagrange daily number.

5. The deep learning inversion method for tunnel resistivity based on initial model constraints according to claim 1, characterized in that, The resistivity prediction is further configured as follows: seismic data obtained from seismic exploration and resistivity distribution transformed based on rock property correlation formula are combined with the initial model updated by simulated annealing method as input data to directly predict the resistivity distribution of the inverted area, wherein the rock property correlation formula is obtained by fitting well logging data.

Images