Construction method of neural network-based full waveform inversion travel time matching objective function
By constructing a travel time matching objective function based on a neural network for full waveform inversion, the travel time difference is automatically estimated and the velocity model is updated, which solves the problems of low accuracy and reliance on manual parameter adjustment in existing technologies and achieves high-precision underground medium inversion.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHINA PETROLEUM & CHEMICAL CORP
- Filing Date
- 2024-12-27
- Publication Date
- 2026-06-30
AI Technical Summary
Existing full-waveform inversion methods rely on linear assumptions when extracting accurate time difference information, resulting in low accuracy and the need for manual parameter adjustment, making it difficult to effectively invert the physical parameters of underground media.
A neural network-based full waveform inversion travel time matching objective function construction method is adopted. The travel time difference is automatically estimated by the neural network, the FWI travel time matching objective function is constructed, the velocity model is updated by the backpropagation wave field, and the travel time difference is iteratively optimized until the preset threshold is reached.
It achieves independence from linear assumptions, improves the accuracy and automation of full waveform inversion, and can more accurately invert the physical parameters of underground media.
Smart Images
Figure CN122307650A_ABST
Abstract
Description
Technical Field
[0001] This disclosure relates to the fields of velocity modeling and seismic imaging in oil and gas exploration, and in particular to a method for constructing a full-waveform inversion travel-time matching objective function based on neural networks. Background Technology
[0002] Under current technology, full waveform inversion (FWI) is an important technique in geophysical exploration. It inverts the physical parameters of the subsurface medium by comparing model prediction data with actual observed seismic data.
[0003] Currently, conventional full waveform inversion typically uses an objective function based on the L2 norm. This objective function requires the initial model of the full waveform inversion to reflect the kinematic characteristics of the real underground model, thereby avoiding the "period jump" problem in the lowest inversion frequency band.
[0004] To overcome the above problems, various waveform inversion objective functions have been proposed, such as:
[0005] (1) Objective function based on travel time matching of predicted data and observed data;
[0006] (2) Objective function derived from the matched filter based on predicted and observed data;
[0007] (3) Objective function based on matching the predicted data with the observed data envelopment;
[0008] (4) Objective functions defined based on other norms or distances between predicted and observed data.
[0009] Among these, the objective function based on the travel time matching of predicted and observed data can most directly reflect the kinematic characteristics of the model; however, the biggest challenge in applying this type of objective function lies in extracting accurate time difference information. Conventional methods for extracting time difference information include dynamic time programming, local cross-correlation, domain transformation, and linear filter coefficient analysis. These methods rely on linear assumptions, have drawbacks such as low accuracy, and require high-quality data and manual parameter adjustment. Summary of the Invention
[0010] The purpose of this invention is to provide at least one method for constructing a full-waveform inversion travel time matching objective function based on neural networks. By utilizing the nonlinear strength of neural networks, it is possible to automatically estimate accurate travel time information, which has the advantages of not relying on linear assumptions, high accuracy, and high degree of automation.
[0011] To address the aforementioned technical problems, at least one embodiment of this application provides a method for constructing a full-waveform inversion travel-time matching objective function based on a neural network, the method comprising:
[0012] Step S1: Excite the wavelet at the source location, perform forward wave field modeling in the current velocity model, record the forward wave field through a detector, and obtain a single-channel prediction vector; and obtain the travel time difference vector between the single-channel prediction vector and the single-channel observation vector through a preset neural network model;
[0013] Step S2: If the travel time difference vector is greater than a preset target threshold, construct an FWI travel time matching objective function based on the travel time difference vector, construct a full-waveform inversion associated source based on the FWI travel time matching objective function, obtain the backpropagation wavefield by backpropagating the associated source, update the current velocity model based on the backpropagation wavefield and the forward modeling wavefield, and re-acquire the travel time difference vector based on the updated current velocity model and the preset neural network model; repeat step S2 until the travel time difference vector is not greater than the preset target threshold.
[0014] Step S3: If the time difference vector is not greater than the preset target threshold, construct the final FWI time matching target function based on the time difference vector.
[0015] At least one embodiment of this application also provides a system for constructing a full-waveform inversion travel-time matching objective function based on a neural network, comprising:
[0016] The travel time difference vector determination module is used to excite the wavelet at the source location, perform wavefield forward modeling in the current velocity model, record the forward modeled wavefield through a detector, and obtain a single-channel prediction vector; and to obtain the travel time difference vector between the single-channel prediction vector and the single-channel observation vector through a preset neural network model.
[0017] The travel time difference vector optimization module is used to construct an FWI travel time matching objective function based on the travel time difference vector when the travel time difference vector is greater than a preset target threshold; construct an accompanying source for full waveform inversion based on the FWI travel time matching objective function; obtain a backpropagation wavefield by backpropagating the accompanying source; update the current velocity model based on the backpropagation wavefield and the forward modeling wavefield; and re-acquire the travel time difference vector based on the updated current velocity model and the preset neural network model; and control the travel time difference vector optimization module to repeat the process multiple times until the travel time difference vector is not greater than the preset target threshold.
[0018] The construction module is used to construct the final FWI travel time matching objective function based on the travel time difference vector, provided that the travel time difference vector is not greater than the preset target threshold.
[0019] At least one embodiment of this application also provides an electronic device, including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the method described above.
[0020] At least one embodiment of this application also provides a computer-readable storage medium having a computer program stored thereon that, when executed by a processor, implements the steps of the method described above.
[0021] At least one embodiment of this application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the fiber optic detection method as described above.
[0022] The method for constructing the travel time matching objective function based on a neural network provided in this application, compared with the prior art, combines a neural network to improve the traditional method for estimating the travel time difference vector between explicit predicted data and observed data. It overcomes the shortcomings of traditional methods, such as linear assumptions, low accuracy, and reliance on manual parameter adjustment, and has the characteristics of nonlinearity, high accuracy, and high degree of automation.
[0023] In some optional embodiments, the method further includes:
[0024] Step S4: Perform inversion iteration based on the final FWI time-matching objective function.
[0025] In some optional embodiments, the FWI time-matching objective function includes:
[0026]
[0027] Where L is the FWI travel time matching objective function, and t is the travel time difference vector.
[0028] In some optional embodiments, updating the current velocity model based on the propagated wavefield and the forward wavefield includes:
[0029] Cross-correlation is performed on the returned wave field and the forward wave field to obtain the full waveform inversion gradient;
[0030] The current velocity model is updated based on the full waveform inversion gradient.
[0031] In some optional embodiments, the step of re-acquiring the travel time difference vector based on the updated current velocity model and the preset neural network model includes:
[0032] Based on the updated current velocity model, a new wavefield forward modeling is performed to obtain a single-channel prediction vector, and the travel time difference vector between the single-channel prediction vector and the single-channel observation vector is obtained again through the preset neural network model.
[0033] In some alternative embodiments, the system further includes:
[0034] The inversion iteration module is used to perform inversion iteration based on the final FWI travel time matching objective function. Attached Figure Description
[0035] One or more embodiments are illustrated by way of example with reference numerals in the accompanying drawings. These illustrations do not constitute a limitation on the embodiments. Elements with the same reference numerals in the drawings are denoted as similar elements. Unless otherwise stated, the figures in the drawings are not to be limited by scale.
[0036] Figure 1 A flowchart illustrating a method for constructing a full-waveform inversion travel-time matching objective function based on a neural network, as provided in this embodiment of the disclosure;
[0037] Figure 2 A schematic diagram of a neural network provided in an embodiment of this disclosure;
[0038] Figure 3 This is a schematic diagram illustrating a data alignment result for extracting time differences, provided in an embodiment of this disclosure. Detailed Implementation
[0039] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the various embodiments of the present invention will be described in detail below with reference to the accompanying drawings. However, those skilled in the art will understand that many technical details have been presented in the various embodiments of the present invention to enable the reader to better understand the present invention. However, the technical solutions claimed in the present invention can be implemented even without these technical details and various changes and modifications based on the following embodiments.
[0040] Under current technology, full waveform inversion is an important technique in geophysical exploration. It involves comparing model-predicted data with actual observed seismic data to invert the physical parameters of the subsurface medium.
[0041] Currently, conventional full waveform inversion typically uses an objective function based on the L2 norm. This objective function requires the initial model of the full waveform inversion to reflect the kinematic characteristics of the real underground model, thereby avoiding the "period jump" problem in the lowest inversion frequency band.
[0042] To overcome the above problems, various waveform inversion objective functions have been proposed, such as:
[0043] (1) Objective function based on travel time matching of predicted data and observed data;
[0044] (2) Objective function derived from the matched filter based on predicted and observed data;
[0045] (3) Objective function based on matching the predicted data with the observed data envelopment;
[0046] (4) Objective functions defined based on other norms or distances between predicted and observed data.
[0047] Among these, the objective function based on the travel time matching of predicted and observed data can most directly reflect the kinematic characteristics of the model; however, the biggest challenge in applying this type of objective function lies in extracting accurate time difference information. Conventional methods for extracting time difference information include dynamic time programming, local cross-correlation, domain transformation, and linear filter coefficient analysis. These methods rely on linear assumptions, have drawbacks such as low accuracy, and require high-quality data and manual parameter adjustment.
[0048] To address the shortcomings of existing technologies as described above, the present invention aims to provide a method for constructing a full-waveform inversion travel-time matching objective function based on a neural network. By incorporating a neural network, this method improves upon traditional methods for estimating the travel-time difference vector between explicit predicted and observed data. It overcomes the shortcomings of traditional methods, such as linear assumptions, low accuracy, and reliance on manual parameter adjustments, and features nonlinearity, high accuracy, and a high degree of automation.
[0049] Example 1:
[0050] The embodiments of the present invention relate to a method for constructing a full waveform inversion travel-time matching objective function based on a neural network.
[0051] The following is a detailed description of the implementation details of the neural network-based full waveform inversion travel-time matching objective function construction method in this embodiment. The following content is only for the convenience of understanding and is not necessary for implementing this solution.
[0052] The neural network-based full-waveform inversion travel-time matching objective function construction method of this embodiment can be applied to electronic devices with communication, computing, and data storage capabilities. For example... Figure 1 As shown, the method for constructing the full waveform inversion travel-time matching objective function based on neural networks provided in this embodiment includes the following steps:
[0053] Step S1: Excite the wavelet at the source location, perform forward wave field modeling in the current velocity model, record the forward wave field through a detector, and obtain a single-channel prediction vector; and obtain the travel time difference vector between the single-channel prediction vector and the single-channel observation vector through a preset neural network model.
[0054] The data to be prepared includes single-channel prediction data (represented as vector p) and single-channel observation data (represented as vector d); the neural network model (represented as function F) provided by this invention extracts the travel time difference (represented as vector t) between the prediction data p and the observation data d.
[0055] It should be noted that constructing a pre-defined neural network (hereinafter referred to as the neural network) and training the neural network F... φ This neural network is used to automatically extract travel time information from seismic data. The input layer of the network is a random vector r, and the objective function is the precise travel time difference between each seismic event in the observed and predicted data. This process can be expressed as:
[0056] F φ (r) = t(p, d)
[0057] Where φ represents the parameters of the neural network, and its training loss function is:
[0058]
[0059] Among them, DA(F) φ (r), p, d) are the applied sequence alignment evaluation criteria, representing F φ (r) represents the optimal alignment scheme for p and d, and DA(·) can be any combination of sequence alignment criteria such as dynamic time programming, cross-correlation, and optimal transport distance.
[0060] The neural network architecture provided by this invention can be referenced. Figure 2 .
[0061] Step S2: If the travel time difference vector is greater than a preset target threshold, construct an FWI travel time matching objective function based on the travel time difference vector, construct a full waveform inversion companion source based on the FWI travel time matching objective function, obtain the backpropagation wavefield by backpropagating the companion source, update the current velocity model based on the backpropagation wavefield and the forward modeling wavefield, and re-acquire the travel time difference vector based on the updated current velocity model and the preset neural network model; repeat step S2 until the travel time difference vector is not greater than the preset target threshold.
[0062] In step S2, the travel time difference t extracted by the neural network is reduced to a preset target threshold, thereby obtaining a velocity model that reflects the true underground velocity. The preset target threshold can be set according to actual needs.
[0063] Step S3: If the time difference vector is not greater than the preset target threshold, construct the final FWI time matching target function based on the time difference vector.
[0064] If the travel time difference vector is not greater than the preset target threshold, the final FWI travel time matching objective function is constructed based on the travel time difference vector.
[0065] Example 2:
[0066] Based on the above embodiments, this embodiment further explains and illustrates the method for constructing the full waveform inversion travel-time matching objective function based on neural networks provided in the above embodiments.
[0067] In step S1: the wavelet is excited at the source location, the wave field is simulated in the current velocity model, and the wave field is recorded by the detector to obtain the single-channel prediction vector; and the travel time difference vector between the single-channel prediction vector and the single-channel observation vector is obtained by the preset neural network model.
[0068] The data to be prepared includes single-channel prediction data (represented as vector p) and single-channel observation data (represented as vector d); the neural network model (represented as function F) provided by this invention extracts the travel time difference (represented as vector t) between the prediction data p and the observation data d.
[0069] It should be noted that constructing a pre-defined neural network (hereinafter referred to as the neural network) and training the neural network F... φ This neural network is used to automatically extract travel time information from seismic data. The input layer of the network is a random vector r, and the objective function is the precise travel time difference between each seismic event in the observed and predicted data. This process can be expressed as:
[0070] F φ (r) = t(p, d)
[0071] Where φ represents the parameters of the neural network, and its training loss function is:
[0072]
[0073] Among them, DA(F) φ (r), p, d) are the applied sequence alignment evaluation criteria, representing F φ (r) represents the optimal alignment scheme for p and d, and DA(·) can be any combination of sequence alignment criteria such as dynamic time programming, cross-correlation, and optimal transport distance.
[0074] The neural network architecture provided by this invention can be referenced. Figure 2 .
[0075] In this step, the predicted data p of the neural network model comes from exciting the wavelet at the source location, performing wavefield forward modeling in the current velocity model, and recording the forward wavefield at the detector location. When the current velocity model accurately reflects the actual subsurface conditions, the predicted data p has a high degree of matching with the actual field observation data d; conversely, when the current velocity model fails to accurately reflect the actual subsurface conditions, the predicted data p has a low degree of matching with the observed data d. In the full waveform inversion, it is necessary to accurately extract the travel time difference between p and d, and iteratively update the current velocity model to gradually eliminate this time difference, thereby making the current velocity model gradually approach the true model.
[0076] The role of the neural network is to nonlinearly estimate the time difference relationship between the forward-modeled predicted data p and the observed data d. Conventional methods for estimating this time difference relationship are usually linear and rely on the structural similarity between p and d. This neural network method can overcome the above problems. First, a neural network is initialized, taking a random vector as input and outputting the time difference between p and d. The time difference output by the initially constructed neural network cannot accurately reflect the time difference relationship between p and d. By backfeeding the set neural network objective function, the neural network is trained to gradually converge, and the output time difference relationship meets the requirements of the objective function, thus correctly inverting the time difference relationship between p and d.
[0077] In step S2: if the travel time difference vector is greater than a preset target threshold, an FWI travel time matching objective function is constructed based on the travel time difference vector, an accompanying source for full waveform inversion is constructed based on the FWI travel time matching objective function, a backpropagation wavefield is obtained by backpropagating the accompanying source, and the current velocity model is updated based on the backpropagation wavefield and the forward modeling wavefield. Furthermore, the travel time difference vector is re-acquired based on the updated current velocity model and the preset neural network model. Step S2 is repeated until the travel time difference vector is not greater than the preset target threshold.
[0078] In some embodiments, updating the current velocity model based on the propagated wavefield and the forward wavefield includes:
[0079] Cross-correlation is performed on the returned wave field and the forward wave field to obtain the full waveform inversion gradient;
[0080] The current velocity model is updated based on the full waveform inversion gradient.
[0081] In some embodiments, the step of re-acquiring the travel time difference vector based on the updated current velocity model and the preset neural network model includes:
[0082] Based on the updated current velocity model, a new wavefield forward modeling is performed to obtain a single-channel prediction vector, and the travel time difference vector between the single-channel prediction vector and the single-channel observation vector is obtained again through the preset neural network model.
[0083] In step S2, the travel time difference t extracted by the neural network is reduced to a preset target threshold to obtain a velocity model that can reflect the true underground velocity.
[0084] Using the travel time difference *t* from the converged neural network output, a target function for full waveform inversion is constructed. Furthermore, an accompanying seismic source for full waveform inversion is constructed, and the propagated wavefield is obtained by backpropagating this source. This propagated wavefield is then cross-correlated with the forward propagation wavefield recorded in the first step to form the full waveform inversion gradient, which is used to update the current velocity model. This update can reduce the travel time difference between the predicted data *p* extracted by the neural network and the observed data *d*.
[0085] In step S3: if the time difference vector is not greater than the preset target threshold, the final FWI time matching target function is constructed based on the time difference vector.
[0086] In some embodiments, the FWI time-matching objective function includes:
[0087]
[0088] Where L is the FWI travel time matching objective function, and t is the travel time difference vector.
[0089] If the travel time difference vector is not greater than the preset target threshold, the final FWI travel time matching objective function is constructed based on the travel time difference vector.
[0090] In some embodiments, the method further includes:
[0091] Step S4: Perform inversion iteration based on the final FWI time-matching objective function.
[0092] It should be noted that the time difference data alignment results extracted by the method disclosed in this embodiment can be used as a reference. Figure 3 Wherein: the red and blue curves are the predicted earthquake records and the observed earthquake records, respectively, and the black dashed line represents the travel time correspondence extracted from the two data.
[0093] The method for constructing the travel time matching objective function based on a neural network provided in this application, compared with the prior art, combines a neural network to improve the traditional method for estimating the travel time difference vector between explicit predicted data and observed data. It overcomes the shortcomings of traditional methods, such as linear assumptions, low accuracy, and reliance on manual parameter adjustment, and has the characteristics of nonlinearity, high accuracy, and high degree of automation.
[0094] Example 3:
[0095] Based on the above embodiments, another embodiment of this application relates to a system for constructing a full waveform inversion travel-time matching objective function based on a neural network.
[0096] The following provides a detailed description of the implementation details of the neural network-based full waveform inversion travel-time matching objective function construction system in this embodiment. The following implementation details are provided for ease of understanding and are not essential for implementing this solution. The neural network-based full waveform inversion travel-time matching objective function construction system provided in this embodiment includes:
[0097] The travel time difference vector determination module is used to excite the wavelet at the source location, perform wavefield forward modeling in the current velocity model, record the forward modeled wavefield through a detector, and obtain a single-channel prediction vector; and to obtain the travel time difference vector between the single-channel prediction vector and the single-channel observation vector through a preset neural network model.
[0098] The travel time difference vector optimization module is used to construct an FWI travel time matching objective function based on the travel time difference vector when the travel time difference vector is greater than a preset target threshold; construct an accompanying source for full waveform inversion based on the FWI travel time matching objective function; obtain a backpropagation wavefield by backpropagating the accompanying source; update the current velocity model based on the backpropagation wavefield and the forward modeling wavefield; and re-acquire the travel time difference vector based on the updated current velocity model and the preset neural network model; and control the travel time difference vector optimization module to repeat the process multiple times until the travel time difference vector is not greater than the preset target threshold.
[0099] A construction module is used to construct a final FWI travel time matching objective function based on the travel time difference vector, provided that the travel time difference vector is not greater than the preset target threshold. Wherein:
[0100] In the travel time difference vector determination module: the wavelet is excited at the source location, the wave field is simulated in the current velocity model, and the forward wave field is recorded by the detector to obtain the single-channel prediction vector; and the travel time difference vector between the single-channel prediction vector and the single-channel observation vector is obtained by the preset neural network model.
[0101] The data to be prepared includes single-channel prediction data (represented as vector p) and single-channel observation data (represented as vector d); the neural network model (represented as function F) provided by this invention extracts the travel time difference (represented as vector t) between the prediction data p and the observation data d.
[0102] It should be noted that constructing a pre-defined neural network (hereinafter referred to as the neural network) and training the neural network F...φ This neural network is used to automatically extract travel time information from seismic data. The input layer of the network is a random vector r, and the objective function is the precise travel time difference between each seismic event in the observed and predicted data. This process can be expressed as:
[0103] F φ (r) = t(p, d)
[0104] Where φ represents the parameters of the neural network, and its training loss function is:
[0105]
[0106] Among them, DA(F) φ (r), p, d) are the applied sequence alignment evaluation criteria, representing F φ (r) represents the optimal alignment scheme for p and d, and DA(·) can be any combination of sequence alignment criteria such as dynamic time programming, cross-correlation, and optimal transport distance.
[0107] The neural network architecture provided by this invention can be referenced. Figure 2 .
[0108] In this step, the predicted data p of the neural network model comes from exciting the wavelet at the source location, performing wavefield forward modeling in the current velocity model, and recording the forward wavefield at the detector location. When the current velocity model accurately reflects the actual subsurface conditions, the predicted data p has a high degree of matching with the actual field observation data d; conversely, when the current velocity model fails to accurately reflect the actual subsurface conditions, the predicted data p has a low degree of matching with the observed data d. In the full waveform inversion, it is necessary to accurately extract the travel time difference between p and d, and iteratively update the current velocity model to gradually eliminate this time difference, thereby making the current velocity model gradually approach the true model.
[0109] The role of the neural network is to nonlinearly estimate the time difference relationship between the forward-modeled predicted data p and the observed data d. Conventional methods for estimating this time difference relationship are usually linear and rely on the structural similarity between p and d. This neural network method can overcome the above problems. First, a neural network is initialized, taking a random vector as input and outputting the time difference between p and d. The time difference output by the initially constructed neural network cannot accurately reflect the time difference relationship between p and d. By backfeeding the set neural network objective function, the neural network is trained to gradually converge, and the output time difference relationship meets the requirements of the objective function, thus correctly inverting the time difference relationship between p and d.
[0110] In the travel time difference vector optimization module: when the travel time difference vector is greater than a preset target threshold, an FWI travel time matching objective function is constructed based on the travel time difference vector, an accompanying source for full waveform inversion is constructed based on the FWI travel time matching objective function, a backpropagation wavefield is obtained by backpropagating the accompanying source, and the current velocity model is updated based on the backpropagation wavefield and the forward modeling wavefield. Furthermore, the travel time difference vector is re-acquired based on the updated current velocity model and the preset neural network model. Step S2 is repeated until the travel time difference vector is not greater than the preset target threshold.
[0111] In some embodiments, updating the current velocity model based on the propagated wavefield and the forward wavefield includes:
[0112] Cross-correlation is performed on the returned wave field and the forward wave field to obtain the full waveform inversion gradient;
[0113] The current velocity model is updated based on the full waveform inversion gradient.
[0114] In some embodiments, the step of re-acquiring the travel time difference vector based on the updated current velocity model and the preset neural network model includes:
[0115] Based on the updated current velocity model, a new wavefield forward modeling is performed to obtain a single-channel prediction vector, and the travel time difference vector between the single-channel prediction vector and the single-channel observation vector is obtained again through the preset neural network model.
[0116] In step S2, the travel time difference t extracted by the neural network is reduced to a preset target threshold to obtain a velocity model that can reflect the true underground velocity.
[0117] Using the travel time difference *t* from the converged neural network output, a target function for full waveform inversion is constructed. Furthermore, an accompanying seismic source for full waveform inversion is constructed, and the propagated wavefield is obtained by backpropagating this source. This propagated wavefield is then cross-correlated with the forward propagation wavefield recorded in the first step to form the full waveform inversion gradient, which is used to update the current velocity model. This update can reduce the travel time difference between the predicted data *p* extracted by the neural network and the observed data *d*.
[0118] In the construction module: if the time difference vector is not greater than the preset target threshold, the final FWI time matching target function is constructed based on the time difference vector.
[0119] In some embodiments, the FWI time-matching objective function includes:
[0120]
[0121] Where L is the FWI travel time matching objective function, and t is the travel time difference vector.
[0122] If the travel time difference vector is not greater than the preset target threshold, the final FWI travel time matching objective function is constructed based on the travel time difference vector.
[0123] In some embodiments, the system further includes:
[0124] The inversion iteration module is used to perform inversion iteration based on the final FWI travel time matching objective function.
[0125] It should be noted that the time difference data alignment results extracted by the system disclosed in this embodiment can be used as a reference. Figure 3 Wherein: the red and blue curves are the predicted earthquake records and the observed earthquake records, respectively, and the black dashed line represents the travel time correspondence extracted from the two data.
[0126] It is worth mentioning that all modules involved in this embodiment are logical modules. In practical applications, a logical unit can be a physical unit, a part of a physical unit, or a combination of multiple physical units. Furthermore, to highlight the innovative aspects of this application, this embodiment does not introduce units that are not closely related to solving the technical problems proposed in this application; however, this does not mean that other units are absent in this embodiment.
[0127] Example 4:
[0128] Another embodiment of this application relates to an electronic device, comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the neural network-based full waveform inversion travel-time matching objective function construction method in the above embodiments:
[0129] Step S1: Excite the wavelet at the source location, perform forward wave field modeling in the current velocity model, record the forward wave field through a detector, and obtain a single-channel prediction vector; and obtain the travel time difference vector between the single-channel prediction vector and the single-channel observation vector through a preset neural network model;
[0130] Step S2: If the travel time difference vector is greater than a preset target threshold, construct an FWI travel time matching objective function based on the travel time difference vector, construct a full-waveform inversion associated source based on the FWI travel time matching objective function, obtain the backpropagation wavefield by backpropagating the associated source, update the current velocity model based on the backpropagation wavefield and the forward modeling wavefield, and re-acquire the travel time difference vector based on the updated current velocity model and the preset neural network model; repeat step S2 until the travel time difference vector is not greater than the preset target threshold.
[0131] Step S3: If the time difference vector is not greater than the preset target threshold, construct the final FWI time matching target function based on the time difference vector.
[0132] The memory and processor are connected via a bus, which can include any number of interconnecting buses and bridges, connecting various circuits of one or more processors and memories. The bus can also connect various other circuits, such as peripheral devices, voltage regulators, and power management circuits, which are well known in the art and will not be described further herein. The bus interface provides an interface between the bus and the transceiver. The transceiver can be a single element or multiple elements, such as multiple receivers and transmitters, providing a unit for communicating with various other devices over a transmission medium. Data processed by the processor is transmitted over the wireless medium via an antenna, which further receives data and transmits it to the processor.
[0133] The processor manages the bus and general processing, and also provides various functions, including timing, peripheral interfaces, voltage regulation, power management, and other control functions. Memory is used to store data used by the processor during operation.
[0134] Example 5:
[0135] Another embodiment of this application relates to a computer-readable storage medium storing a computer program. When executed by a processor, the computer program implements the neural network-based full waveform inversion time-matching objective function construction method in the above-described method embodiments.
[0136] Step S1: Excite the wavelet at the source location, perform forward wave field modeling in the current velocity model, record the forward wave field through a detector, and obtain a single-channel prediction vector; and obtain the travel time difference vector between the single-channel prediction vector and the single-channel observation vector through a preset neural network model;
[0137] Step S2: If the travel time difference vector is greater than a preset target threshold, construct an FWI travel time matching objective function based on the travel time difference vector, construct a full-waveform inversion associated source based on the FWI travel time matching objective function, obtain the backpropagation wavefield by backpropagating the associated source, update the current velocity model based on the backpropagation wavefield and the forward modeling wavefield, and re-acquire the travel time difference vector based on the updated current velocity model and the preset neural network model; repeat step S2 until the travel time difference vector is not greater than the preset target threshold.
[0138] Step S3: If the time difference vector is not greater than the preset target threshold, construct the final FWI time matching target function based on the time difference vector.
[0139] That is, those skilled in the art will understand that all or part of the steps in the methods of the above embodiments can be implemented by a program instructing related hardware. This program is stored in a storage medium and includes several instructions to cause a device (which may be a microcontroller, chip, etc.) or processor to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.
[0140] In some embodiments of this application, a computer program product is also provided, including a computer program that, when executed by a processor, implements the steps of the method described in the above embodiments:
[0141] Step S1: Excite the wavelet at the source location, perform forward wave field modeling in the current velocity model, record the forward wave field through a detector, and obtain a single-channel prediction vector; and obtain the travel time difference vector between the single-channel prediction vector and the single-channel observation vector through a preset neural network model;
[0142] Step S2: If the travel time difference vector is greater than a preset target threshold, construct an FWI travel time matching objective function based on the travel time difference vector, construct a full-waveform inversion associated source based on the FWI travel time matching objective function, obtain the backpropagation wavefield by backpropagating the associated source, update the current velocity model based on the backpropagation wavefield and the forward modeling wavefield, and re-acquire the travel time difference vector based on the updated current velocity model and the preset neural network model; repeat step S2 until the travel time difference vector is not greater than the preset target threshold.
[0143] Step S3: If the time difference vector is not greater than the preset target threshold, construct the final FWI time matching target function based on the time difference vector.
[0144] Those skilled in the art will understand that the above embodiments are specific embodiments for implementing this application, and in practical applications, various changes can be made to them in form and detail without departing from the spirit and scope of this application.
Claims
1. A method for constructing a full-waveform inversion travel-time matching objective function based on neural networks, characterized in that, The method includes: Step S1: Excite the wavelet at the source location, perform forward wave field modeling in the current velocity model, record the forward wave field through a detector, and obtain a single-channel prediction vector; and obtain the travel time difference vector between the single-channel prediction vector and the single-channel observation vector through a preset neural network model; Step S2: If the travel time difference vector is greater than a preset target threshold, construct an FWI travel time matching objective function based on the travel time difference vector, construct a full-waveform inversion associated source based on the FWI travel time matching objective function, obtain the backpropagation wavefield by backpropagating the associated source, update the current velocity model based on the backpropagation wavefield and the forward modeling wavefield, and re-acquire the travel time difference vector based on the updated current velocity model and the preset neural network model; repeat step S2 until the travel time difference vector is not greater than the preset target threshold. Step S3: If the time difference vector is not greater than the preset target threshold, construct the final FWI time matching target function based on the time difference vector.
2. The method according to claim 1, characterized in that, The method further includes: Step S4: Perform inversion iteration based on the final FWI time-matching objective function.
3. The method according to claim 1, characterized in that, The FWI time-matching objective function includes: Where L is the FWI travel time matching objective function, and t is the travel time difference vector.
4. The method according to claim 1, characterized in that, The step of updating the current velocity model based on the returned wavefield and the forward wavefield includes: Cross-correlation is performed on the returned wave field and the forward wave field to obtain the full waveform inversion gradient; The current velocity model is updated based on the full waveform inversion gradient.
5. The method according to claim 1, characterized in that, The step of re-acquiring the travel time difference vector based on the updated current velocity model and the preset neural network model includes: Based on the updated current velocity model, a new wavefield forward modeling is performed to obtain a single-channel prediction vector, and the travel time difference vector between the single-channel prediction vector and the single-channel observation vector is obtained again through the preset neural network model.
6. A system for constructing a full-waveform inversion travel-time matching objective function based on a neural network, characterized in that, include: The travel time difference vector determination module is used to excite the wavelet at the source location, perform wavefield forward modeling in the current velocity model, record the forward modeled wavefield through a detector, and obtain a single-channel prediction vector; and to obtain the travel time difference vector between the single-channel prediction vector and the single-channel observation vector through a preset neural network model. The travel time difference vector optimization module is used to construct an FWI travel time matching objective function based on the travel time difference vector when the travel time difference vector is greater than a preset target threshold; construct an accompanying source for full waveform inversion based on the FWI travel time matching objective function; obtain a backpropagation wavefield by backpropagating the accompanying source; update the current velocity model based on the backpropagation wavefield and the forward modeling wavefield; and re-acquire the travel time difference vector based on the updated current velocity model and the preset neural network model; and control the travel time difference vector optimization module to repeat the process multiple times until the travel time difference vector is not greater than the preset target threshold. The construction module is used to construct the final FWI travel time matching objective function based on the travel time difference vector, provided that the travel time difference vector is not greater than the preset target threshold.
7. The system according to claim 6, characterized in that, The system also includes: The inversion iteration module is used to perform inversion iteration based on the final FWI travel time matching objective function.
8. An electronic device, characterized in that, include: At least one processor; as well as, A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor, which, when executed by the at least one processor, enables the at least one processor to perform the neural network-based full waveform inversion walk-time matching objective function construction method as described in any one of claims 1 to 5.
9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the steps of the method according to any one of claims 1 to 5.
10. A computer program product, comprising a computer program, characterized in that, When executed by a processor, the computer program implements the steps of the method according to any one of claims 1 to 5.