Geological radar full-waveform inversion optimization method and system based on convolutional neural network
Through the combination of convolutional neural network and differentiable convolution operator, the problem of inaccurate gradient calculation and slow calculation speed in full waveform inversion of geological radar is solved, and more accurate and efficient inversion results are achieved, significantly improving the calculation efficiency.
Patent Information
- Application Number
- CN202411057611.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-02
- Publication Date
- 2024-11-19
AI Technical Summary
There are problems such as inaccuracy and slow calculation speed in the full waveform inversion of geological radar, which leads to inaccurate inaccuracy of gradient calculation and optimization, which leads to inaccurate inversion results and huge consumption of computing resources.
Deep learning method based on convolutional neural network is adopted, and the gradient is calculated automatically by differentially differentially and backpropagating optimization is performed, combined with the time-domain finite difference forward evolution method of differentiable convolution operators, reducing inverse wavefield calculations and improving calculation efficiency.
Effectively suppress noise, improve the accuracy and calculation speed of inversion results, and significantly improve the efficiency and accuracy of full waveform inversion of geological radar.
Smart Images

Figure CN118981015A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of geophysical exploration, and in particular relates to a geological radar full waveform inversion optimization method and system based on convolutional neural network. 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] Ground Penetrating Radar (GPR) is a geophysical exploration method that uses high-frequency electromagnetic waves to detect underground targets. Its working principle is to transmit high-frequency electromagnetic waves into the ground through the transmitting antenna. When the electromagnetic waves encounter a target body with different dielectric parameters, they are reflected and received and recorded by the receiving antenna. The type and distribution of the target body are inferred through parameters such as radar waveform, electromagnetic field intensity, amplitude, spectrum characteristics and two-way travel time. GPR has the advantages of being non-destructive, efficient, high-resolution and sensitive to water bodies. The electromagnetic signal propagates quickly and does not require coupling agents. It is widely used in many fields such as geological exploration, archaeology, construction engineering, environmental science, forensic medicine and agriculture.
[0004] In geological exploration and geophysical research, ground penetrating radar can analyze the structure and composition of different underground strata, search for mineral resources, and monitor geological disasters such as landslides and subsidence; in archaeology and cultural heritage protection, it is used to detect underground archaeological sites, tombs, and ancient buildings, and help protect and restore cultural relics; in construction engineering and urban planning, ground penetrating radar can locate and detect underground pipelines, foundations, tunnels and other facilities, and evaluate the structural health of roads and bridges; in environmental science and ecological research, it is used to investigate the distribution of soil pollutants, evaluate groundwater resources and changes in water quality. Ground penetrating radar technology, due to its high efficiency, accuracy, and non-destructive characteristics, provides important data support and decision-making basis for many fields.
[0005] Full waveform inversion (FWI) is a high-resolution imaging method for inverting underground medium parameters based on full wavefield information. The principle of full waveform inversion of geological radar is to match the forward radar waveform and the actual observed radar waveform to minimize the difference between the two, thereby obtaining parameters such as the dielectric constant and conductivity of the medium. FWI can accurately invert the dielectric constant information of underground media, and has the characteristics of high precision and strong systematicity. It is widely used in theoretical research on numerical simulation of engineering inspections such as highways, bridges and tunnels. However, since this method uses a local optimization algorithm, if the initial model of the inversion is not set reasonably and the large-scale structural information in the initial model is inaccurate, it is very easy to cause the inversion to fall into a local minimum, resulting in a frequency jump phenomenon.
[0006] At present, there are the following difficulties in the full waveform inversion of geological radar:
[0007] First, the problem of gradient calculation and optimization of full waveform inversion of geological radar. Traditional full waveform inversion (FWI) usually requires a good initial model and uses an iterative optimization algorithm to update the model parameters. Specifically, FWI uses a cross-correlation algorithm to calculate the gradient, and this process requires additional calculation of the reverse time wave field. At the same time, the calculated gradient is often not optimized, resulting in inaccurate model updates. This inaccuracy is mainly reflected in two aspects: one is that it is easily interfered by noise, and the other is that it is affected by the gradient interference of the source point, which makes the inversion result inaccurate.
[0008] Second, the problem of calculation speed of geological radar full waveform inversion. The traditional gradient descent method needs to specify the initial iteration step size when calculating the model update amount, which makes the model update process not flexible enough. If the initial step size is not selected properly, it may lead to a slow iteration process or even convergence to an erroneous solution. At the same time, since the number of model parameters involved in FWI is very large, and the time-domain finite difference forward algorithm needs to be called multiple times during inversion, the amount of calculation is huge, which makes the calculation speed an issue that cannot be ignored. As the scale of the model increases, the amount of calculation increases exponentially, which places extremely high demands on computing resources and time. Therefore, how to improve computing efficiency and reduce computing time while ensuring inversion accuracy has become an important challenge in FWI applications. Summary of the invention
[0009] In order to solve the above problems, the present invention proposes a geological radar full waveform inversion optimization method and system based on convolutional neural network. Aiming at the gradient calculation and optimization processing problems of geological radar full waveform inversion, the present invention uses automatic differentiation in deep learning to obtain gradients, and establishes a convolutional neural network for gradient optimization; aiming at the calculation speed problem of geological radar full waveform inversion, a time-domain finite difference forward modeling method based on differentiable convolution operator is proposed; finally, the geological radar full waveform inversion result is optimized, the calculation speed is improved, and the inversion result with good correspondence with the real model is obtained.
[0010] According to some embodiments, the present invention adopts the following technical solutions:
[0011] A geological radar full waveform inversion optimization method based on convolutional neural network includes the following steps:
[0012] (1) Constructing an initial model of medium parameters;
[0013] (2) setting an observation mode on the initial model;
[0014] (3) performing filtering and correction processing of a convolutional neural network on the initial model;
[0015] (4) Perform electromagnetic wave forward modeling on the model after filtering and correction, and calculate the simulated observation data of the receiving antenna position through the time-domain finite difference forward modeling method based on the differentiable convolution operator;
[0016] (5) Compare the simulated observation data with the real observation data of the real model, calculate the loss function, and add regularization constraints;
[0017] (6) Using the automatic differentiation method of deep learning to calculate the gradient, the process of back propagation calculation of the gradient passes through the convolutional neural network, and the convolutional neural network performs filtering and correction processing on the gradient; using the optimizer in deep learning to update the parameters of the model;
[0018] (7) Repeat steps (3)-(6) until the iterative convergence conditions are met and the full waveform inversion optimization result of the geological radar is obtained.
[0019] As an optional implementation, in the step (1), the initial model for constructing the medium parameters is a two-dimensional model or a three-dimensional model;
[0020] The medium parameters are selected from relevant electrical parameters and / or magnetic parameters of geological radar.
[0021] As a further feature, the medium parameter is specifically the relative dielectric constant of the medium;
[0022] Or, the medium parameter is specifically the conductivity of the medium;
[0023] Alternatively, the medium parameters are specifically the relative dielectric constant of the medium and the conductivity of the medium.
[0024] As an optional implementation, in step (1), the initial model is a model that does not contain abnormal objects and only contains background media; or a model that is a fuzzy model of the real model.
[0025] As an optional implementation, in step (2), the observation method is surface observation method, cross-hole observation method or drilling observation method.
[0026] As an optional implementation, in step (3), the convolutional neural network includes an upsampling part and a downsampling part, the upsampling part is composed of two convolutional layers, and the downsampling part is composed of two classic residual network blocks, and the number of channels of the downsampling part gradually increases, and the number of channels of the upsampling part gradually decreases, forming a symmetrical structure.
[0027] As an optional implementation, the specific process of step (4) includes: in the finite difference time domain forward modeling method, using the second-order central difference format represented by the following formula to replace the partial differential operation in the curl equation in the time domain Maxwell equations;
[0028]
[0029] f′(x) represents the derivative of function f(x) at point x, f(x) is the function to be approximated, x is the point at which the derivative is to be calculated, Δx is a small step size for finite difference approximation, O((Δx) 2 )) is the order of the error term, which indicates the error in the approximation and (Δx) 2 is proportional to;
[0030] Extract the convolution operator F from the formula i :(1,-1), such that:
[0031]
[0032] This converts the differential calculation into a convolution calculation.
[0033] As an optional implementation, in step (5), the real observation data is the real observation data of the receiving antenna position calculated on the real medium parameter model by a time-domain finite difference forward modeling method based on a differentiable convolution operator.
[0034] As an optional implementation, in step (5), the loss function is a mean square error loss function or a mean absolute error loss function.
[0035] As an optional implementation, in step (5), the regularization constraint is specifically total variation regularization.
[0036] As an optional implementation, in step (6), the optimizer is an Adam optimizer, an SGD optimizer, an LBFGS optimizer or an AdamW optimizer.
[0037] A geological radar full waveform inversion optimization system based on convolutional neural network, comprising:
[0038] a model building module configured to build an initial model of medium parameters;
[0039] An observation mode setting module, configured to set an observation mode on the initial model;
[0040] A filtering correction module is configured to perform a convolutional neural network filtering correction process on the initial model;
[0041] A forward modeling module is configured to perform electromagnetic wave forward modeling on the model after the filtering correction process, and calculate the simulated observation data of the receiving antenna position by using a time-domain finite difference forward modeling method based on a differentiable convolution operator;
[0042] A loss calculation and constraint module is configured to compare the simulated observation data with the real observation data of the real model, calculate the loss function, and add regularization constraints;
[0043] The optimization update module is configured to use the automatic differentiation method of deep learning to calculate the gradient. The process of back-propagation calculation of the gradient passes through the convolutional neural network, and the convolutional neural network performs filtering and correction processing on the gradient; the optimizer in deep learning is used to update the parameters of the model;
[0044] The iteration module is configured to repeatedly call the filtering correction module, the forward simulation module, the loss calculation and constraint module and the optimization update module until the iterative convergence conditions are met to obtain the full waveform inversion optimization result of the geological radar.
[0045] Compared with the prior art, the present invention has the following beneficial effects:
[0046] (1) Aiming at the problem of gradient calculation and optimization processing of full waveform inversion of geological radar, the present invention proposes to use automatic differentiation and back propagation in deep learning to realize gradient calculation, thereby eliminating the additional calculation of the reverse time wave field; it proposes to use convolutional neural network to optimize the gradient and model, obtain the corrected gradient and model, effectively suppress the noise and ghost in the traditional full waveform inversion effect, and better characterize the position and morphology of the anomaly; the proposed convolutional neural network is embedded in the iterative process of full waveform inversion of geological radar, and the inversion gradient is optimized during the iteration to obtain an accurate inversion effect.
[0047] (2) In order to solve the problem of computational speed of full waveform inversion of geological radar, the present invention proposes a time-domain finite-difference forward modeling method based on a differentiable convolution operator, which converts the differential calculation into a convolution calculation. Thus, some built-in acceleration measures of PyTorch on CUDA for convolution operations can be simply used to improve the computational speed of full waveform inversion of geological radar.
[0048] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, preferred embodiments are given below and described in detail with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS
[0049] 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.
[0050] Figure 1 It is a flow chart of a geological radar full waveform inversion optimization method based on a convolutional neural network in an embodiment of the present invention;
[0051] Figure 2 Schematic diagram of the convolutional neural network structure in an embodiment of the present invention;
[0052] Figure 3 1 is a comparison diagram of the effects before and after total variation regularization in an embodiment of the present invention;
[0053] Figure 4 A comparison diagram of the gradient map before the convolutional neural network and the gradient correction after the convolutional neural network in an embodiment of the present invention;
[0054] Figure 5 It is a schematic diagram of the real model, traditional inversion effect and full waveform inversion result of geological radar based on convolutional neural network in an embodiment of the present invention. DETAILED DESCRIPTION
[0055] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.
[0056] 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.
[0057] 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.
[0058] In the absence of conflict, the embodiments in this application and the features in the embodiments may be combined with each other.
[0059] Embodiment 1
[0060] A geological radar full waveform inversion optimization method based on convolutional neural network, comprising:
[0061] Construct an initial model of medium parameters;
[0062] After that, set the observation mode on the initial model;
[0063] The model is subjected to filtering and correction processing of the convolutional neural network; the processed model is subjected to electromagnetic wave forward simulation, and the simulated observation data of the receiving antenna position is calculated by the time-domain finite difference forward method based on the differentiable convolution operator; the loss function is calculated with the real observation data of the real model, and regularization constraints are added; the gradient is calculated using the automatic differentiation method of deep learning, and the convolutional neural network is used in the process of back propagation to calculate the gradient, which is filtered and corrected by the convolutional neural network; the optimizer in deep learning is used to update the model parameters; the above process is repeated until the iterative convergence conditions are met to obtain the full waveform inversion optimization result of the geological radar.
[0064] In some embodiments, the initial model is a two-dimensional model.
[0065] In some embodiments, the initial model is a three-dimensional model.
[0066] In some embodiments, the medium parameters in the initial model are selected from relevant electrical parameters and magnetic parameters of geological radar, and the medium parameters are specifically the relative dielectric constant of the medium.
[0067] Of course, in some other embodiments, the medium parameter is specifically the conductivity of the medium, or the medium parameter is specifically the relative dielectric constant of the medium and the conductivity of the medium.
[0068] In some embodiments, the initial model is a model that does not contain abnormal bodies and only contains background media.
[0069] In some embodiments, the initial model is a model obtained by fuzzy processing the real model.
[0070] In some embodiments, the observation method is a surface observation method, and the transmitting antenna and the receiving antenna are both set on the surface, maintaining the same offset distance or a changing offset distance.
[0071] In some embodiments, the observation method is a cross-hole observation method, in which two or more holes are drilled vertically on the ground or horizontally on the face, a transmitting antenna is arranged in one hole for excitation, and receiving antennas are arranged in the remaining holes for reception.
[0072] In some embodiments, the observation method is a drilling observation method, in which a hole is drilled vertically on the ground or horizontally on the face, and a transmitting antenna and a receiving antenna are arranged for excitation and reception.
[0073] The specific idea of the finite-difference time-domain forward modeling method based on the differentiable convolution operator is as follows:
[0074] The central difference format used by FDTD only involves the two sampling values before and after f(x), but not the value of f(x) itself. In the FDTD method, the second-order central difference format represented by the following formula is widely used to replace the partial differential operation in the curl equation in the time domain Maxwell equations.
[0075]
[0076] Extract the convolution operator F from the formula i :(1,-1), so that
[0077]
[0078] This converts the differential calculation into a convolution calculation, and we can simply use some built-in acceleration measures of PyTorch for convolution operations on CUDA.
[0079] The real observation data are the real observation data of the receiving antenna position calculated on the real medium parameter model by the time-domain finite difference forward method based on the differentiable convolution operator.
[0080] In some embodiments, the automatic differentiation method is specifically implemented based on PyTorch.
[0081] In some other embodiments, the automatic differentiation method is specifically implemented based on TensorFlow.
[0082] In some embodiments, when calculating the loss function with the real observation data obtained from the real model, the loss function is a mean square error (MSE) loss function.
[0083] In some other embodiments, the loss function is a mean absolute error (MAE) loss function.
[0084] Adding a regularization constraint to the loss function, wherein the regularization constraint is specifically Total Variation Regularization (TV Regularization);
[0085] In some embodiments, when an optimizer is used to update the initial model parameters, the optimizer is specifically an Adam optimizer.
[0086] Of course, in other embodiments, the optimizer is also an SGD optimizer, or an LBFGS optimizer, or an AdamW optimizer.
[0087] Embodiment 2
[0088] A full waveform inversion optimization method for geological radar based on convolutional neural network, combined with Figure 1 , specifically including the following process:
[0089] (1) Construct an initial two-dimensional model of relative dielectric constant and set up a cross-hole observation method;
[0090] (2) The model is filtered and corrected by a convolutional neural network, combined with Figure 2 , the convolutional neural network architecture of the present invention is given, which has only a simple four-block structure, performs two downsampling and two upsampling operations, and finally obtains the model parameter m k The same shape of m k ′.
[0091] Among them, the downsampling part consists of two classic residual network blocks.
[0092] First, perform the first convolution operation F on the input conv1 , use a 3×3 convolution kernel, a stride of 1, a padding of 1, and then perform batch normalization F bn1 , and then through the ReLU activation function F relu , and get output1.
[0093] Then, the second convolution operation F is performed on the output of the previous layer output1 conv2 , use a 3×3 convolution kernel, a stride of 1, a padding of 1, and then perform batch normalization F bn2 , and get output2.
[0094] If the dimension needs to be adjusted, the third convolution operation f is performed on the output of the previous layer output2 conv3 , a convolution operation is performed using a 1×1 convolution kernel to match the shape of the output, resulting in output3.
[0095] Add output2 or output3 to input input.
[0096] Finally, the ReLU activation function f relu , and get the output of the residual network block.
[0097] Using residual networks for downsampling can alleviate the gradient vanishing problem, improve training efficiency, improve the generalization ability of the network, and simplify the learning process.
[0098] Afterwards, the output results of the down-sampling by the two residual blocks are up-sampled twice by two up-sampling blocks to obtain the input shape required by the forward function.
[0099] The upsampling part consists of two convolutional layers.
[0100] First, perform the first layer of deconvolution operation F on the input trans , the convolution kernel size is 2×2, the stride is 2, and the padding is 1, and output 1 is obtained. The main function of this layer is to upsample and enlarge the width and height of the feature map.
[0101] Then, the second layer of conventional convolution operation F is performed on the output output1 of the previous layer conv2 , use a 3×3 convolution kernel, a step size of 1, and a padding of 1 to get output2. The function of this layer is to further process the upsampled feature map to give it better feature expression capabilities.
[0102] Will After ut2 passes through the sigmoid activation function, it is scaled to the appropriate threshold range to obtain the output of the final upsampling network block.
[0103] This upsampling module can effectively perform upsampling through deconvolution and enhance the expressiveness of features in combination with convolution operations. It has a simple and effective structure and is flexible and scalable.
[0104] Finally, the network of the present invention was built by combining the downsampling function of the residual network and the upsampling module.
[0105] Combination Figure 2 , the two residual blocks increase the number of input channels from 1 to 4, and then from 4 to 16, respectively, and use a 1×1 convolutional layer to adjust the number of channels and a stride of 2 for downsampling. Increasing the number of channels means that each layer can extract more features and improve the feature expression ability. Downsampling can expand the receptive field of each neuron, allowing the neural network to better capture global information. At the same time, with the increase in the number of network layers and downsampling operations, the network can gradually extract more abstract and higher-level features, thereby better recognizing complex patterns.
[0106] Combination Figure 2 , the two upsampling modules reduce the number of channels of the feature map from 16 to 4, and then from 4 to 1, respectively, and perform upsampling at the same time to restore the feature map, thereby obtaining the input shape required by the forward function. With the upsampling of the feature map, the feature representation required by the network gradually shifts from high-level abstract features to detailed features. Reducing the number of channels helps to streamline the feature representation, allowing the network to better focus on reconstructing and refining the detailed information of the image.
[0107] In the network architecture of the present invention, the number of channels gradually increases during downsampling, while the number of channels gradually decreases during upsampling, forming a balanced symmetrical structure. This design enables the network to have reasonable feature representation capabilities in both the encoding and decoding stages, which helps to improve the overall performance.
[0108] (3) Use the time-domain finite-difference method based on the differentiable convolution operator to perform forward simulation and obtain the observation data of the initial model;
[0109] In this embodiment, the finite difference time domain method based on the differentiable convolution operator can extract the convolution operators in three directions. Except for the different directions, the values are all (1, -1); after obtaining the convolution operator, the differential operation in the forward calculation can be replaced by the convolution operation commonly used in deep neural networks. In essence, the convolution operation is a template calculation based on the convolution kernel. This template calculation method is one of the most important computing cores and optimization directions in current high-performance computing. In this embodiment, multi-dimensional tensor programming in PyTorch, a common software platform for deep learning, is used to realize the parallel calculation of the forward process, thereby greatly accelerating the finite difference time domain forward calculation process.
[0110] (4) Obtain the loss function by comparing it with the observed data of the real model;
[0111] The loss function used in this embodiment is the mean square error loss function (MSE), which is specifically:
[0112]
[0113] Among them, f(x i ) represents the observed data obtained by the ith model through the finite-difference time-domain forward simulation based on the differentiable convolution operator, x i represents the model of the i-th iteration, y represents the real observation data obtained by the real model through the time-domain finite difference forward simulation based on the differentiable convolution operator, and N is the number of iterations.
[0114] The regularization method used in this embodiment is Total Variation Regularization (TV Regularization), combined with Figure 3 , total variation regularization can effectively maintain the edge and structural information of the image, while smoothing the image and removing noise. Specifically:
[0115]
[0116] Among them, the total variation term TV(u) is the sum of the gradients of the input data, u represents the original input data, represents the gradient of u in the horizontal direction at position (i, j), Represents the gradient of u in the vertical direction at position (i, j).
[0117] (5) Use PyTorch's automatic differentiation function to obtain the gradient of the loss function relative to the model. PyTorch is a powerful deep learning framework that implements gradient calculation through the backpropagation algorithm. PyTorch's automatic differentiation function is implemented by building a computational graph. The nodes in the computational graph represent tensors, and the edges represent operations. The gradient is calculated by backpropagation based on the computational graph. The gradient will pass through the convolutional neural network and combine with Figure 4 , the network will filter and correct the gradient;
[0118] (6) Use PyTorch's Adam optimizer for gradient descent. The core idea of the Adam optimizer is to introduce first-order moment estimation (momentum) and second-order moment estimation (square of gradient) on the basis of gradient descent, which correspond to the moving average of the gradient and the moving average of the square of the gradient, respectively. In each iteration, the Adam optimizer updates the model parameters according to these two estimated values;
[0119] (7) Model update;
[0120] In this embodiment, by using automatic differentiation and back propagation in deep learning, the additional calculation of the reverse time wave field is omitted; the convolutional neural network is used to optimize the gradient and model, such as Figure 5 As shown in the figure, the noise and ghosting in the traditional inversion are effectively suppressed, and the position and shape of the abnormal body are more accurately characterized. The time-domain finite-difference forward modeling method based on the differentiable convolution operator is used to convert the differential calculation into a convolution calculation. The built-in acceleration measures of PyTorch on CUDA are used to significantly improve the inversion calculation speed, and finally a good full-waveform inversion effect is obtained.
[0121] Embodiment 3
[0122] A geological radar full waveform inversion optimization system based on convolutional neural network, comprising:
[0123] a model building module configured to build an initial model of medium parameters;
[0124] An observation mode setting module, configured to set an observation mode on the initial model;
[0125] A filtering correction module is configured to perform a convolutional neural network filtering correction process on the initial model;
[0126] A forward modeling module is configured to perform electromagnetic wave forward modeling on the model after the filtering correction process, and calculate the simulated observation data of the receiving antenna position by using a time-domain finite difference forward modeling method based on a differentiable convolution operator;
[0127] A loss calculation and constraint module is configured to compare the simulated observation data with the real observation data of the real model, calculate the loss function, and add regularization constraints;
[0128] The optimization update module is configured to use the automatic differentiation method of deep learning to calculate the gradient. The process of back-propagation calculation of the gradient passes through the convolutional neural network, and the convolutional neural network performs filtering and correction processing on the gradient; the optimizer in deep learning is used to update the parameters of the model;
[0129] The iteration module is configured to repeatedly call the filtering correction module, the forward simulation module, the loss calculation and constraint module and the optimization update module until the iterative convergence conditions are met to obtain the full waveform inversion optimization result of the geological radar.
[0130] 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.
[0131] 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.
[0132] 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.
[0133] 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.
[0134] 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 by those skilled in the art within the spirit and principle of the present invention without creative labor shall be included in the protection scope of the present invention.
Claims
1. A full waveform inversion optimization method for geological radar based on convolutional neural network, characterized in that: The following steps are involved: (1) Constructing an initial model of medium parameters; (2) setting an observation mode on the initial model; (3) performing filtering and correction processing of a convolutional neural network on the initial model; (4) Perform electromagnetic wave forward modeling on the model after filtering and correction, and calculate the simulated observation data of the receiving antenna position through the time-domain finite difference forward modeling method based on the differentiable convolution operator; (5) Compare the simulated observation data with the real observation data of the real model, calculate the loss function, and add regularization constraints; (6) Using the automatic differentiation method of deep learning to calculate the gradient, the process of back propagation calculation of the gradient passes through the convolutional neural network, and the convolutional neural network performs filtering and correction processing on the gradient; Use the optimizer in deep learning to update the parameters of the model; (7) Repeat steps (3)-(6) until the iterative convergence conditions are met and the full waveform inversion optimization result of the geological radar is obtained.
2. The method for full waveform inversion optimization of geological radar based on convolutional neural network as claimed in claim 1, characterized in that: In the step (1), the initial model for constructing the medium parameters is a two-dimensional model or a three-dimensional model; The medium parameters are selected from relevant electrical parameters and / or magnetic parameters of geological radar; Or further, the medium parameter is specifically the relative dielectric constant of the medium; Or, the medium parameter is specifically the conductivity of the medium; Alternatively, the medium parameters are specifically the relative dielectric constant of the medium and the conductivity of the medium.
3. In step (1), the initial model is a model containing no abnormal bodies and only background medium; Alternatively, the initial model is a model obtained by fuzzy processing the real model.
4. The method for full waveform inversion optimization of geological radar based on convolutional neural network as claimed in claim 1, characterized in that: In the step (2), the observation method is a surface observation method, a cross-hole observation method or a drilling observation method.
5. The method for full waveform inversion optimization of geological radar based on convolutional neural network as claimed in claim 1, characterized in that: In the step (3), the convolutional neural network includes an upsampling part and a downsampling part, the upsampling part is composed of two convolutional layers, the downsampling part includes two classic residual network blocks, and the number of channels of the downsampling part gradually increases, while the number of channels of the upsampling part gradually decreases, forming a symmetrical structure.
6. The method for full waveform inversion optimization of geological radar based on convolutional neural network as claimed in claim 1, characterized in that: The specific process of step (4) includes: in the finite difference time domain forward modeling method, using the second-order central difference format represented by the following formula to replace the partial differential operation in the curl equation in the time domain Maxwell equations; f ′ (x) represents the derivative of function f(x) at point x, f(x) is the function to be approximated, x is the point at which the derivative is to be calculated, Δx is a small step size used for finite difference approximation, O((Δx) 2 )) is the order of the error term, which indicates the error in the approximation and (Δx) 2 is proportional to; Extract the convolution operator F from the formula i :(1,-1), such that: This converts the differential calculation into a convolution calculation.
7. The method for full waveform inversion optimization of geological radar based on convolutional neural network as claimed in claim 1, characterized in that: In the step (5), the real observation data is the real observation data of the receiving antenna position calculated on the real medium parameter model by the time-domain finite difference forward method based on the differentiable convolution operator.
8. The method for full waveform inversion optimization of geological radar based on convolutional neural network as claimed in claim 1, characterized in that: In the step (5), the loss function is a mean square error loss function or a mean absolute error loss function; In the step (5), the regularization constraint is specifically total variation regularization.
9. The method for full waveform inversion optimization of geological radar based on convolutional neural network as claimed in claim 1, characterized in that: In step (6), the optimizer is an Adam optimizer, an SGD optimizer, an LBFGS optimizer or an AdamW optimizer.
10. A geological radar full waveform inversion optimization system based on convolutional neural network, characterized in that: include: a model building module configured to build an initial model of medium parameters; An observation mode setting module, configured to set an observation mode on the initial model; A filtering correction module is configured to perform a convolutional neural network filtering correction process on the initial model; A forward modeling module is configured to perform electromagnetic wave forward modeling on the model after the filtering correction process, and calculate the simulated observation data of the receiving antenna position by using a time-domain finite difference forward modeling method based on a differentiable convolution operator; A loss calculation and constraint module is configured to compare the simulated observation data with the real observation data of the real model, calculate the loss function, and add regularization constraints; The optimization update module is configured to use the automatic differentiation method of deep learning to calculate the gradient. The process of back-propagation calculation of the gradient passes through the convolutional neural network, and the convolutional neural network performs filtering and correction processing on the gradient; Use the optimizer in deep learning to update the parameters of the model; The iteration module is configured to repeatedly call the filtering correction module, the forward simulation module, the comparison constraint module and the optimization update module until the iterative convergence conditions are met to obtain the full waveform inversion optimization result of the geological radar.
Citation Information
Cited By
Underground object three-dimensional reconstruction method and device
CN119625191A
CUDA (Compute Unified Device Architecture) and PyTorch mixed compiled ground penetrating radar rapid full waveform inversion method
CN120122225A
