Time-lapse magnetotelluric inversion method and system that fuse seismic constraints and physical guidance

Abstract

This invention relates to the field of geophysical exploration technology, specifically disclosing a time-shifted magnetotelluric inversion method and system that integrates seismic constraints and physical guidance. The method includes acquiring magnetotelluric observation data of a monitoring area at multiple consecutive times and seismic constraint data of the corresponding area; constructing a physical guidance loss function; establishing a deep neural network with an integrated attention mechanism, using the physical guidance loss function as the training target, and implementing physical feedback control on the network output using Maxwell's forward modeling equations; inputting the magnetotelluric observation data to be inverted and the seismic constraint data into the trained deep neural network, and outputting the time-shifted inversion result of the subsurface resistivity evolution over time. Through the synergistic effect of multiple constraints, the inversion result simultaneously satisfies the electromagnetic field propagation law, the continuity of time evolution, and the prior information of seismic structure, significantly improving the physical rationality and reliability of the inversion result.

Description

Technical Field

[0001] This invention relates to the field of geophysical exploration technology, specifically to a time-shifted magnetotelluric inversion method and system that integrates seismic constraints and physical guidance. Background Technology

[0002] Time-shifting magnetotelluric (MT) is an effective method for observing the changes in subsurface electrical structure over time. By collecting observation data from the same location at different times and analyzing the changes in data differences between these times, the resistivity changes of the subsurface medium can be derived. This method is based on natural field sources, and the observation equipment is lightweight, flexible, and cost-effective, making it widely applicable in the field of time-shifting inversion monitoring. Particularly in the development and utilization of geothermal resources, there is an urgent need for an effective method for monitoring geothermal reinjection to monitor the migration characteristics of reinjected geothermal tailwater in real time. Due to its sensitivity to changes in the electrical properties of pore fluids and its ability to probe deep, time-shifting MT has recently been applied to the emerging field of reservoir fluid monitoring.

[0003] However, existing time-shifted magnetotelluric inversion techniques have the following shortcomings:

[0004] First, magnetotelluric signals are unstable and irregular, making them susceptible to human-caused noise interference. This is particularly true when monitoring geothermal activity in densely populated residential areas, where electromagnetic interference is severe and the noise spectrum is rich. Existing processing methods typically discard data from periods with poor quality, leading to low accuracy in subsequent data processing and unreliable interpretation results.

[0005] Second, traditional inversion problems are not unique; the same observation data may yield different inversion results. Although the simultaneous inversion method in time-shift inversion can reduce the influence of noise, the inversion process requires multiple iterations, and each iteration requires a large number of forward modeling calculations. Simultaneous inversion of multiple time points consumes a lot of computational resources and is computationally inefficient.

[0006] Third, traditional inversion methods rely solely on electromagnetic data, failing to fully utilize other geophysical information such as seismic data. Seismic data offers high resolution for the layering interfaces and tectonic boundaries of subsurface media, while electromagnetic methods are sensitive to fluid saturation; the two are naturally complementary. Current technologies lack a scheme to effectively integrate seismic structural constraints into time-shifted magnetotelluric inversion.

[0007] Fourth, while deep learning methods have shown potential in geophysical inversion, purely data-driven neural networks are heavily dependent on the training set, making it impossible to guarantee the theoretical reliability of the results on the test set. Furthermore, the lack of standard datasets in the field of geophysics limits the application of deep learning. Simultaneously, existing deep learning inversion methods lack adherence to physical laws, making it difficult to guarantee the physical plausibility of the inversion results.

[0008] Therefore, there is an urgent need in this field for a time-shifted magnetotelluric inversion method that can integrate prior information about seismic structures, follow the laws of electromagnetic physics, and improve inversion accuracy and efficiency. Summary of the Invention

[0009] The purpose of this invention is to provide a time-shifted magnetotelluric inversion method and system that integrates seismic constraints and physical guidance to solve the problems mentioned in the background art.

[0010] To achieve the above objectives, the present invention provides the following technical solution:

[0011] A time-shifted magnetotelluric inversion method integrating seismic constraints and physical guidance, the method comprising:

[0012] Acquire magnetotelluric observation data of the monitoring area at multiple consecutive times and corresponding seismic constraint data of the area;

[0013] Construct a physical guided loss function that includes data residual terms, physical equation constraint terms, time continuity constraint terms, and seismic structural similarity constraint terms;

[0014] A deep neural network with an integrated attention mechanism is established, with the physical guidance loss function as the training target, and physical feedback control is implemented on the network output using Maxwell's forward equations;

[0015] The magnetotelluric observation data and seismic constraint data to be inverted are input into the trained deep neural network, which outputs the time-shifted inversion results of the evolution of subsurface resistivity over time.

[0016] As a further embodiment of the present invention, the physical guidance loss function is:

[0017] ;

[0018] in, This is the error term between the predicted value of the deep learning model and the true label model; This refers to the data residuals based on the forward response and actual observation data; This is a spatial smoothing constraint term for the resistivity model; This is the time regularization term for the model at adjacent time points; For seismic structural constraints based on prior seismic information; µ1, µ2, µ3, µ4, and µ5 are the weighting coefficients for the corresponding terms.

[0019] As a further aspect of the present invention, the time regularization term employs a first-order difference operator to constrain the smooth change of the resistivity model at adjacent time points.

[0020] As a further embodiment of the present invention, the expression for the time regularization term is:

[0021] ;

[0022] Where T is the total number of observation times, m t Let be the resistivity model at time t.

[0023] As a further embodiment of the present invention, the seismic structural constraint term adopts the L2 norm form:

[0024] ;

[0025] in, d represents the predicted seismic response value generated based on the current resistivity model. s This is measured seismic structural constraint data.

[0026] As a further embodiment of the present invention, the deep neural network model is an architecture that integrates convolutional layers and self-attention mechanism modules, and its input channels include magnetotelluric TE mode data features, TM mode data features, and seismic structural profile features.

[0027] As a further embodiment of the present invention, the attention mechanism includes interconnected channel attention units and spatial attention units;

[0028] The channel attention unit is configured to extract feature channel weights through parallel average pooling and max pooling, and output them via a multilayer perceptron.

[0029] The spatial attention unit is configured to extract position weights in the spatial dimension of the feature map.

[0030] As a further aspect of the present invention, when training the deep neural network, a physical forward modeling algorithm is used to perform real-time forward modeling calculations on the resistivity model predicted by the network, and the calculated forward modeling error is fed back into the physical guidance loss function to update the network weight parameters.

[0031] The present invention also provides a time-shifted magnetotelluric inversion system that integrates seismic constraints and physical guidance, the system comprising:

[0032] The data acquisition module is used to acquire magnetotelluric observation data of the monitoring area at multiple consecutive times and the corresponding seismic constraint data of the area;

[0033] The loss function construction module is used to construct a physical guided loss function that includes data residual terms, physical equation constraint terms, time continuity constraint terms, and seismic structural similarity constraint terms.

[0034] The target determination module is used to build a deep neural network with an integrated attention mechanism, using the physical guidance loss function as the training target, and implementing physical feedback control on the network output using Maxwell's forward equations;

[0035] The training module is used to input the magnetotelluric observation data and seismic constraint data to be inverted into the trained deep neural network, and output the time-shifted inversion results of the subsurface resistivity evolution over time.

[0036] Compared with the prior art, the beneficial effects of the present invention are: through the synergistic effect of multiple constraints, the inversion results simultaneously satisfy the electromagnetic field propagation law, the continuity of time evolution and the prior information of seismic structure, which significantly improves the physical rationality and reliability of the inversion results.

[0037] This invention introduces seismic structural constraints into a physics-guided deep learning framework, leveraging the high-resolution characteristics of seismic data on subsurface media interfaces to compensate for the boundary insensitivity of electromagnetic methods. When local electromagnetic observation data is missing or interfered with by noise, the weight of the seismic constraint term is dynamically increased, effectively utilizing prior seismic information to compensate for the decrease in electromagnetic inversion resolution, thus improving the robustness of the method in complex noisy environments.

[0038] By employing Maxwell's forward modeling equations to apply physical feedback control to the network output, the forward modeling error is fed back into the loss function in real time, guiding the network weight parameters to update in a direction consistent with the propagation laws of electromagnetic fields. This physical guidance mechanism enables the network to learn physical laws with a limited number of training samples, thus addressing the problem of insufficient data in the field of geophysics. Attached Figure Description

[0039] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention.

[0040] Figure 1 This is a flowchart of the time-shift inversion method that integrates seismic constraints and physical-guided deep learning, as provided in an embodiment of the present invention.

[0041] Figure 2 This is a flowchart of the frequency-preserving and amplitude-preserving noise reduction filtering method provided in an embodiment of the present invention;

[0042] Figure 3 A schematic diagram of a physically guided deep learning network architecture provided in an embodiment of the present invention;

[0043] Figure 4 The actual model diagram in the experiment is shown to verify the effectiveness of the time regularization and seismic structure dual constraint scheme proposed in this invention under the traditional iterative inversion framework;

[0044] Figure 5 To verify the effectiveness of the proposed time regularization and seismic structure dual-constraint scheme under the traditional iterative inversion framework, the traditional independent inversion diagram was used in the experiment;

[0045] Figure 6 To verify the effectiveness of the proposed time regularization and seismic structure dual-constraint scheme under the traditional iterative inversion framework, the time-constrained inversion diagram is shown in the experiment;

[0046] Figure 7 To verify the effectiveness of the proposed time regularization and seismic structure dual-constraint scheme under the traditional iterative inversion framework, the time-seismic dual-constraint inversion diagram was obtained in the experiment;

[0047] Figure 8 The experimental model diagram shows the inversion effect of the method of the present invention on a one-dimensional complex model;

[0048] Figure 9 The time-constrained inversion diagram of the experiment was used to verify the inversion effect of the method of the present invention on a one-dimensional complex model;

[0049] Figure 10 To verify the inversion effect of the method of the present invention on a one-dimensional complex model, the experimental time-seismic dual-constraint inversion diagram was obtained;

[0050] Figure 11 To verify the inversion effect of the method of the present invention on a one-dimensional complex model, a physical-guided deep learning time-shift inversion diagram was created.

[0051] Figure 12 To verify the inversion effect of the method of the present invention on a one-dimensional complex model, an experiment was conducted to integrate seismic constraints and physical-guided deep learning inversion graphs.

[0052] Figure 13 Figure (a) shows the actual model diagram of the inversion effect experiment of the method of the present invention on a two-dimensional complex model; Figure (b) shows the time-seismic dual-constraint inversion diagram of the inversion effect experiment of the method of the present invention on a two-dimensional complex model.

[0053] Figure 14 Figure (a) shows the physical-guided deep learning time-shift inversion diagram of the inversion effect experiment of the method of the present invention on a two-dimensional complex model; Figure (b) shows the inversion diagram of the fusion of seismic constraints and physical-guided deep learning of the method of the present invention on a two-dimensional complex model. Detailed Implementation

[0054] To make the technical problems to be solved, the technical solutions, and the beneficial effects of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the present invention and are not intended to limit the present invention.

[0055] like Figure 1 As shown in the embodiment of the present invention, the time-shifted magnetotelluric inversion method integrating seismic constraints and physical guidance includes:

[0056] Acquire magnetotelluric observation data of the monitoring area at multiple consecutive times and corresponding seismic constraint data of the area;

[0057] Construct a physical guided loss function that includes data residual terms, physical equation constraint terms, time continuity constraint terms, and seismic structural similarity constraint terms;

[0058] A deep neural network with an integrated attention mechanism is established, with the physical guidance loss function as the training target, and physical feedback control is implemented on the network output using Maxwell's forward equations;

[0059] The magnetotelluric observation data and seismic constraint data to be inverted are input into the trained deep neural network, which outputs the time-shifted inversion results of the evolution of subsurface resistivity over time.

[0060] like Figure 1 and Figure 2 As shown, in this embodiment, seismic constraint data serves as spatial prior information. Preprocessing the raw magnetotelluric time series data includes segmenting the raw data, performing spectrum estimation using short-time Fourier transform, and employing frequency- and amplitude-preserving filtering techniques to eliminate non-Gaussian distributed human noise interference. Specifically, the time-frequency electromagnetic data is processed independently frequency by frequency. New data at each frequency is output after frequency- and amplitude-preserving filtering and improved spectrum estimation, effectively suppressing complex noise in highly interfering environments such as geothermal fields.

[0061] In a preferred embodiment of the present invention, the physical guidance loss function is:

[0062] ;

[0063] in, This is the error term between the predicted value of the deep learning model and the true label model; This refers to the data residuals based on the forward response and actual observation data; This is a spatial smoothing constraint term for the resistivity model; This is the time regularization term for the model at adjacent time points; For seismic structural constraints based on prior seismic information; µ1, µ2, µ3, µ4, and µ5 are the weighting coefficients for the corresponding terms.

[0064] In this embodiment, Used to supervise the network's learning of the mapping relationship between data and the model; This is used to constrain the network output to conform to the electromagnetic field propagation law described by Maxwell's forward modeling equations; This is used to ensure the spatial continuity and smoothness of the inversion results; Used to constrain the smooth changes in the resistivity model between adjacent time points, ensuring the continuity of time-shifted inversion results over time; It is used to utilize the high-resolution characteristics of seismic data to identify the interfaces between layers of subsurface media, thus compensating for the insensitivity of electromagnetic methods to boundaries.

[0065] In a preferred embodiment of the present invention, the time regularization term uses a first-order difference operator to constrain the smooth change of the resistivity model at adjacent time points.

[0066] The expression for the time regularization term is:

[0067] ;

[0068] Where T is the total number of observation times, m t Let be the resistivity model at time t.

[0069] In this embodiment, by taking the root mean square of the sum of squares of the model differences between adjacent time points, the rate of change of resistivity in the time direction is effectively constrained, preventing drastic abrupt changes that do not conform to physical laws from occurring between adjacent time points.

[0070] In a preferred embodiment of the present invention, the seismic structural constraint term adopts the L2 norm form:

[0071] ;

[0072] in, d represents the predicted seismic response value generated based on the current resistivity model. s This is measured seismic structural constraint data.

[0073] In this embodiment, prior seismic information is incorporated into the inversion process by constraining the difference between the seismic response predicted by the model and the measured seismic data.

[0074] In a preferred embodiment of the present invention, the deep neural network model is an architecture that integrates convolutional layers and self-attention mechanism modules, and its input channels include magnetotelluric TE model data features, TM model data features, and seismic structural profile features.

[0075] In a preferred embodiment of the present invention, the attention mechanism includes interconnected channel attention units and spatial attention units;

[0076] The channel attention unit is configured to extract feature channel weights through parallel average pooling and max pooling, and output them via a multilayer perceptron.

[0077] The spatial attention unit is configured to extract position weights in the spatial dimension of the feature map.

[0078] In this embodiment, the channel attention unit enhances important electromagnetic feature channels. Specifically, the channel attention unit calculates the importance of each channel of the input data, and obtains two types of feature vectors by average pooling and max pooling, respectively. These vectors are then input into a multilayer perceptron and output as channel attention features after passing through two fully connected layers. This allows for the automatic identification and enhancement of electromagnetic feature channels that contribute significantly to the inversion task.

[0079] The spatial attention unit is used to extract location weights in the spatial dimension of the feature map, identifying and focusing on sensitive areas of electrical anomalies under seismic constraints. Specifically, the spatial attention unit performs average pooling and max pooling on the feature vectors in space, concatenates the two feature maps, performs convolution operations, and then obtains spatial attention features through an activation function. This guides the network to focus on the spatial locations where underground electrical anomalies may occur.

[0080] The network structure also employs multi-scale convolutional kernels (including 3×3 and 5×5 convolutional kernels) to extract features from different receptive fields in parallel, and uses batch regularization and dropout layers between each layer to increase the stability of the network during training and alleviate the overfitting problem.

[0081] like Figure 3 As shown, in a preferred embodiment of the present invention, when training the deep neural network, the resistivity model predicted by the network is calculated in real time using a physical forward modeling algorithm, and the calculated forward modeling error is fed back into the physical guidance loss function to update the network weight parameters.

[0082] In this embodiment, the calculated forward modeling error is fed back into the physical guidance loss function, which can guide the network weight parameters to be updated in a direction that conforms to the laws of Maxwell's equations, thereby achieving high-precision training on small sample datasets.

[0083] The training process includes:

[0084] Initialize the connection weight parameters of the deep neural network model;

[0085] The constructed magnetotelluric dataset and seismic constraint data are input into the deep neural network model;

[0086] The deep neural network model performs forward propagation on the input data to obtain the predicted resistivity model;

[0087] The forward model of the predicted resistivity was calculated using Maxwell's forward modeling equations to obtain the forward response;

[0088] The difference between the forward model response and the actual observed data (i.e. (items) and the difference between the predictive model and the real model (i.e.) Substitute the items (e.g., Item 1, Item 2, Item 3, Item 4, Item 5, Item 6, Item 7, Item 8, Item 9, Item 10, Item 11, Item 12, Item 13, Item 14, Item 15, Item 16, Item 17, Item 18, Item 19 ...

[0089] The network weights are updated by backpropagation using gradient descent, so that the network output gradually conforms to the physical laws of electromagnetic field propagation.

[0090] Repeat the above process until the termination condition is met.

[0091] The magnetotelluric observation data and seismic constraint data to be processed are input into a trained deep neural network model, and the output is a time-shift inversion model of the evolution of subsurface resistivity over time.

[0092] Specifically, the system receives real-time electromagnetic field time-domain data, performs the same preprocessing as the training data (including denoising, normalization, TE / TM mode separation, and seismic constraint data registration), inputs the preprocessed data into the optimal deep neural network model, and directly outputs the inverted resistivity model after forward propagation calculation. Finally, it outputs resistivity change maps at multiple time points for dynamic monitoring and analysis.

[0093] This invention, for the first time, constructs a multi-dimensional constrained physical guided loss function in time-shifted magnetotelluric inversion, comprising data residual terms, physical equation constraint terms, temporal continuity constraint terms, and seismic structural similarity constraint terms. Through the synergistic effect of multiple constraints, the inversion results simultaneously satisfy the electromagnetic field propagation law, temporal evolution continuity, and prior information on seismic structures, significantly improving the physical rationality and reliability of the inversion results.

[0094] This invention introduces seismic structural constraints into a physics-guided deep learning framework, leveraging the high-resolution characteristics of seismic data on subsurface media interfaces to compensate for the boundary insensitivity of electromagnetic methods. When local electromagnetic observation data is missing or interfered with by noise, the weight of the seismic constraint term is dynamically increased, effectively utilizing prior seismic information to compensate for the decrease in electromagnetic inversion resolution, thus improving the robustness of the method in complex noisy environments.

[0095] This invention designs a deep neural network model integrating channel attention and spatial attention. The channel attention mechanism automatically identifies and enhances electromagnetic feature channels that contribute significantly to the inversion task, while the spatial attention mechanism guides the network to focus on sensitive regions of electrical anomalies. The introduction of the attention mechanism significantly improves the clarity of the inversion model boundaries and the accuracy of anomaly morphology recognition.

[0096] This invention employs Maxwell's forward modeling equations to provide physical feedback control over the network output, feeding the forward modeling error back into the loss function in real time and guiding the network weight parameters to update in a direction consistent with the propagation laws of electromagnetic fields. This physical guidance mechanism enables the network to learn physical laws even with a limited number of training samples, thus addressing the problem of insufficient data in the field of geophysics.

[0097] During training or inference, this invention can dynamically adjust the weighting coefficients of the seismic constraint terms based on the signal-to-noise ratio of the input data, thereby achieving an adaptive response to data quality and improving the versatility and robustness of the method in different monitoring scenarios.

[0098] To verify the effectiveness of the proposed time regularization and seismic structure dual-constraint scheme under the traditional iterative inversion framework, the study focuses on simulating scenarios of discontinuous time data loss due to instrument failure during geothermal reinjection or deep thermal reservoir development.

[0099] Monitoring Model Construction: A deep one-dimensional reservoir monitoring model was constructed to simulate the diffusion evolution of low-resistivity fluids during water injection and heat extraction. The model observation depth was set at 5000m, and the vertical direction was divided into 60 layers at logarithmic intervals to simulate the fine structure of the deep thermal reservoir. The background resistivity was 100Ω·m. Four monitoring time points were set. Initially, a low-resistivity anomaly of 1Ω·m was present at a depth of 1586m-1655m; as time progressed, the anomaly simulated fluid diffusion, expanding 250m upwards and downwards in each step. High-resolution reflection interfaces obtained from seismic exploration were utilized.

[0100] Observational data acquisition and joint denoising: in 10 -3 ~10 2 Sixty frequency points were selected within the Hz band, and the forward response at four time points was obtained using a standard one-dimensional forward modeling algorithm. 5% Gaussian noise was added to the observation data. To simulate an actual monitoring scenario, the full-band observation data at time point 3 was artificially removed to simulate a station failure at that time. The frequency-preserving and amplitude-preserving denoising filtering technique described in this invention was applied to preprocess the signals at the remaining time points to improve the consistency of the signal-to-noise ratio benchmark.

[0101] Comparative Experiment Design: To quantitatively evaluate the advancement of the method of this invention, the following comparative experiments were designed: traditional independent inversion; time-constrained only; and dual-constraint inversion of this invention.

[0102] Inversion Results Analysis and Technical Conclusions: Experimental Comparison Results are as follows Figures 4 to 7As shown, traditional independent inversion completely fails at time 3. Due to the introduction of a time regularization term, the observation data from adjacent time points maintain continuity and can still reflect anomalies at times of data loss, but the anomaly boundaries are relatively blurred, and false anomalies appear in the background. This indicates that time constraints can partially compensate for data loss, but their ability to constrain anomaly boundaries is limited. This invention adds seismic structural constraints to the time constraints, which not only recovers anomaly information at times of data loss but also provides clear anomaly boundaries, achieving the highest degree of agreement with the actual model. This demonstrates that the method of this invention has good robustness under data loss conditions, and the prior information on seismic structures can effectively compensate for the deficiencies in electromagnetic observation data.

[0103] Verify the inversion effect of the method of the present invention on a one-dimensional complex model.

[0104] Model Construction and Denoising Preprocessing: A one-dimensional complex layered model was constructed with a background resistivity of 100 Ω·m. The anomaly simulated a dynamic thickening and low-resistivity process, where the burial depth diffused bidirectionally from approximately 800 meters upwards and downwards to 250-1500 meters, while the resistivity gradually decreased from approximately 20 Ω·m to 1 Ω·m. 5000 sets of sample data were randomly generated, and the frequency-preserving and amplitude-preserving denoising filtering method of this invention was used for preprocessing.

[0105] Physical guidance network architecture design and training: A one-dimensional convolutional neural network was constructed with an input layer dimension of 8×2×50 (8 time points, two channels for apparent resistivity and phase, 50 frequency points) and an output layer dimension of 8×30 (8 time points, 30 layers of resistivity). The physical guidance loss function of this invention was used for training. The seismic guidance branch receives the structure identified by the seismic identification and provides geometric guidance for the spatial distribution of anomalies. The training consisted of 256 epochs, a batch size of 64, an Adam optimizer, and a learning rate of 0.001.

[0106] Comparative experimental designs: time-constrained inversion; dual-constraint inversion; physical-guided deep learning time-shift inversion; and the time-shift inversion of this invention that integrates seismic constraints and physical-guided deep learning.

[0107] Experimental Results Analysis and Conclusions: Figures 8 to 12 The paper demonstrates the reconstruction effects of various methods on the reservoir diffusion process at eight time points. Time-constrained inversion showed significantly poor reconstruction results; while dual-constraint inversion improved accuracy, the anomalies exhibited severe "fogging," and the first two methods took tens of minutes; physical-guided deep learning time-shift inversion achieved rapid imaging, but multiple false anomalies appeared in the results; the resistivity model output by the method of this invention had the sharpest boundaries, perfectly reconstructing complex morphologies, and providing accurate resistance recovery. This indicates that the method of this invention, through the synergistic effect of seismic structural constraints and physical-guided deep learning, can achieve optimal inversion results on one-dimensional complex models.

[0108] Verify the inversion effect of the method of the present invention on complex two-dimensional models.

[0109] Model Construction and Denoising Preprocessing: The model's horizontal range is -2500m to 2500m, and its vertical depth is 2500m. The mesh is 51×50, with a unit mesh size of 100m×50m. The background resistivity is 100Ω·m. A high-resistivity bedrock barrier is set at a depth of 2000m to 2400m. Three continuous monitoring times are set to simulate the process of fracturing fluid (1Ω·m) gradually diffusing vertically from shallow to deep, accompanied by lateral contraction. A two-dimensional seismic structure matrix is ​​constructed as the spatial geometric constraint input. The frequency-preserving and amplitude-preserving denoising filtering method of this invention is used for preprocessing.

[0110] Design and Training of a Two-Dimensional Physically Guided Convolutional Neural Network: A two-dimensional convolutional neural network was constructed, with an input layer dimension of 3×3×50×51 (3 time points, 3 channels for TE / TM / seismic transmission, 50 frequency points, and 51 receiver points). An attention module was introduced, including channel attention units and spatial attention units. The convolutional layers used 3×3 and 5×5 convolutional kernels in parallel, with a dropout ratio of 0.2. The output layer dimension was 3×5000 (3 time points, 5000 grid parameters). The physics-guided loss function of this invention was used for training, and the Adam optimizer was used for 60 rounds of iterative optimization to ensure that the inversion model conforms to the physical laws of Maxwell's equations while possessing sharp seismic interface constraint characteristics.

[0111] For two-dimensional laterally non-uniform models, the following comparison groups are set up: dual-constraint inversion; physical-guided deep learning time-shift inversion; and the time-shift inversion of this invention that integrates seismic constraints and physical-guided deep learning.

[0112] Experimental Results Analysis and Technical Conclusions: Figure 13 and Figure 14 The paper demonstrates the effectiveness of various methods in characterizing the evolution of three consecutive fracturing processes. In the inversion profile obtained through dual-constraint inversion, the anomaly is blurred, and the resistivity is severely interfered with by the high-resistivity bedrock. While the physical-guided deep learning time-shift inversion imaging is fast, it exhibits obvious oscillating spurious anomalies at deep interfaces, resulting in discontinuous boundary characterization. The two-dimensional model output from the time-shift inversion of this invention, which integrates seismic constraints and physical-guided deep learning, has the highest boundary clarity and can accurately characterize the dynamic changes of the anomaly. This indicates that the method of this invention, through the synergistic effect of seismic structural constraints and physical-guided deep learning, can achieve optimal inversion results on complex two-dimensional models.

[0113] This invention also provides a time-shifted magnetotelluric inversion system that integrates seismic constraints and physical guidance, the system comprising:

[0114] The data acquisition module is used to acquire magnetotelluric observation data of the monitoring area at multiple consecutive times and the corresponding seismic constraint data of the area;

[0115] The loss function construction module is used to construct a physical guided loss function that includes data residual terms, physical equation constraint terms, time continuity constraint terms, and seismic structural similarity constraint terms.

[0116] The target determination module is used to build a deep neural network with an integrated attention mechanism, using the physical guidance loss function as the training target, and implementing physical feedback control on the network output using Maxwell's forward equations;

[0117] The training module is used to input the magnetotelluric observation data and seismic constraint data to be inverted into the trained deep neural network, and output the time-shifted inversion results of the subsurface resistivity evolution over time.

[0118] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A time-shifted magnetotelluric inversion method integrating seismic constraints and physical guidance, characterized in that, The method includes: Acquire magnetotelluric observation data of the monitoring area at multiple consecutive times and corresponding seismic constraint data of the area; Construct a physical guided loss function that includes data residual terms, physical equation constraint terms, time continuity constraint terms, and seismic structural similarity constraint terms; A deep neural network with an integrated attention mechanism is established, with the physical guidance loss function as the training target, and physical feedback control is implemented on the network output using Maxwell's forward equations; The magnetotelluric observation data and seismic constraint data to be inverted are input into the trained deep neural network, which outputs the time-shifted inversion results of the subsurface resistivity evolution over time.

2. The time-shifted magnetotelluric inversion method integrating seismic constraints and physical guidance according to claim 1, characterized in that, The physical guidance loss function is: ； in, This is the error term between the predicted value of the deep learning model and the true label model; This refers to the data residuals based on the forward response and actual observation data; This is a spatial smoothing constraint term for the resistivity model; This is the time regularization term for the model at adjacent time points; For seismic structural constraints based on prior seismic information; µ1, µ2, µ3, µ4, and µ5 are the weighting coefficients for the corresponding terms.

3. The time-shifted magnetotelluric inversion method integrating seismic constraints and physical guidance according to claim 2, characterized in that, The time regularization term uses a first-order difference operator to constrain the smooth change of the resistivity model at adjacent time points.

4. The time-shifted magnetotelluric inversion method integrating seismic constraints and physical guidance according to claim 3, characterized in that, The expression for the time regularization term is: ； Where T is the total number of observation times, m t Let be the resistivity model at time t.

5. The time-shifted magnetotelluric inversion method integrating seismic constraints and physical guidance according to claim 2, characterized in that, The seismic structural constraint terms are in L2 norm form: ； in, d represents the predicted seismic response value generated based on the current resistivity model. s This is measured seismic structural constraint data.

6. The time-shifted magnetotelluric inversion method integrating seismic constraints and physical guidance according to claim 1, characterized in that, The deep neural network model is an architecture that integrates convolutional layers and self-attention mechanism modules. Its input channels include magnetotelluric TE model data features, TM model data features, and seismic structural profile features.

7. The time-shifted magnetotelluric inversion method integrating seismic constraints and physical guidance according to claim 1 or 6, characterized in that, The attention mechanism includes interconnected channel attention units and spatial attention units; The channel attention unit is configured to extract feature channel weights through parallel average pooling and max pooling, and output them via a multilayer perceptron. The spatial attention unit is configured to extract position weights in the spatial dimension of the feature map.

8. The time-shifted magnetotelluric inversion method integrating seismic constraints and physical guidance according to claim 1, characterized in that, When training the deep neural network, the resistivity model predicted by the network is calculated in real time using a physical forward modeling algorithm, and the calculated forward modeling error is fed back into the physical guidance loss function to update the network weight parameters.

9. A time-shifted magnetotelluric inversion system integrating seismic constraints and physical guidance, used to implement the time-shifted magnetotelluric inversion method integrating seismic constraints and physical guidance as described in any one of claims 1-8, characterized in that, The system includes: The data acquisition module is used to acquire magnetotelluric observation data of the monitoring area at multiple consecutive times and the corresponding seismic constraint data of the area; The loss function construction module is used to construct a physical guided loss function that includes data residual terms, physical equation constraint terms, time continuity constraint terms, and seismic structural similarity constraint terms. The target determination module is used to build a deep neural network with an integrated attention mechanism, using the physical guidance loss function as the training target, and implementing physical feedback control on the network output using Maxwell's forward equations; The training module is used to input the magnetotelluric observation data and seismic constraint data to be inverted into the trained deep neural network, and output the time-shifted inversion results of the subsurface resistivity evolution over time.

Images