Multi-modal holographic electromagnetic exploration data denoising method, device, equipment and medium

Abstract

The application discloses a kind of multi-modal holographic electromagnetic exploration data denoising method, device, equipment and medium, method includes: obtaining the standardization multi-modal data of holographic electromagnetic exploration of underground three-dimensional target body;According to polarization mode group, regional field block, parameter sequence hierarchical structure reorganizes data and is mapped into two-dimensional noisy image;Two-dimensional noisy image is input into pre-trained multi-modal attention deep learning network to extract features and carry out noise modeling to obtain two-dimensional noise estimation image;By difference operation and inverse mapping algorithm, the target multi-modal data after denoising is restored.The application effectively solves the problem of multi-modal data coupling feature fracture, while significantly improving the resolution and accuracy of subsequent three-dimensional inversion imaging in precise noise suppression.

Description

Technical Field

[0001] This invention relates to the field of electromagnetic exploration data processing technology, and in particular to a method, apparatus, equipment and medium for denoising multimodal holographic electromagnetic exploration data. Background Technology

[0002] Three-dimensional holographic electromagnetic exploration technology can achieve high-resolution imaging of underground three-dimensional targets by simultaneously acquiring multi-polarization modes, multi-regional field blocks, and various types of electromagnetic parameter data.

[0003] However, in actual exploration operations, the acquired multimodal data often contains complex noise. This includes environmental background noise (power frequency interference, natural electromagnetic noise), instrument electronic noise, human activity interference, and multimodal data coupling noise. This leads to a reduced signal-to-noise ratio, distortion of electromagnetic field and gradient information, and blurring of regional field characteristics. This not only affects the extraction of comprehensive information about the target but may also introduce false field source effects, interfering with the accuracy of subsequent imaging and inversion.

[0004] Existing technologies have significant limitations in denoising multimodal data from holographic electromagnetic exploration. On one hand, traditional filtering methods (such as wavelet thresholding and Kalman filtering) cannot adapt to the multi-polarization, multi-region, and multi-parameter coupling characteristics of 3D holographic electromagnetic data. Processing requires separate denoising of each mode, which disrupts polarization mode differences, regional field complementarity, and electromagnetic field-gradient coupling, easily leading to missing target information during imaging. On the other hand, existing deep learning methods mostly employ general image denoising networks, failing to design specific structures for the full-coverage observation characteristics, differential-integral hybrid excitation modes, and multi-regional field information differences inherent in 3D holographic electromagnetic exploration. This results in poor network adaptability to multimodal data, distortion of the field source features after denoising, and limited resolution improvement.

[0005] Chinese patent CN11681963B discloses a method for denoising artificial source electromagnetic exploration using synchronous array observation data. It constructs inter-station transfer functions by selecting high signal-to-noise ratio data segments and estimates and replaces noise at stations with strong interference. However, it only processes single-mode (artificial source) time / frequency domain signals, increasing the amplitude of the artificial source signal's dominant frequency by suppressing noise, and employing a traditional inter-station transfer function model based on the signal correlation of a single mode. This method is clearly incapable of handling highly coupled multi-mode data in holographic electromagnetic exploration, and even more so of adapting to the inherent physical correlations within holographic electromagnetic exploration multi-mode data. It easily disrupts polarization mode differences, regional field complementarity, and the coupling relationship between electromagnetic field components and gradient components, leading to broken multi-mode coupling characteristics, residual field source effects, or damage to high-resolution physical property information after denoising. It is difficult to completely preserve core exploration features while suppressing noise.

[0006] In view of the above, this application is hereby submitted. Summary of the Invention

[0007] This application provides a method, apparatus, device, and medium for denoising multimodal holographic electromagnetic exploration data, which solves the technical problem of poor denoising effect of three-dimensional holographic electromagnetic exploration multimodal data in the prior art and improves the denoising effect of three-dimensional holographic electromagnetic exploration multimodal data.

[0008] In a first aspect, this application provides a method for denoising multimodal holographic electromagnetic exploration data, including: Acquire standardized multimodal data of holographic electromagnetic exploration of underground three-dimensional targets; The standardized multimodal data is reorganized according to the hierarchical structure composed of polarization mode groups, regional field blocks, and parameter sequences to obtain the reorganized intermediate structure data. The reorganized intermediate structure data is mapped into a two-dimensional noisy image of a preset dimension by adopting the arrangement strategy of polarization mode group association, regional field block adjacency, and parameter sequence coupling. A two-dimensional noisy image is input into a pre-trained multimodal attention deep learning network. The multimodal attention deep learning network extracts multi-dimensional features and performs noise modeling to obtain a two-dimensional noise estimation image. The difference between the two-dimensional noisy image and the two-dimensional noise estimation image is calculated to obtain a clean two-dimensional image. The two-dimensional noise estimation image is then reversed and restored to obtain the holographic electromagnetic exploration denoised target multimodal data through an inverse mapping algorithm that matches the three-dimensional to two-dimensional fidelity reconstruction algorithm.

[0009] In some embodiments of this application, based on the aforementioned scheme, standardized multimodal data of holographic electromagnetic exploration of underground three-dimensional target bodies are obtained, including: Acquire raw multimodal data of holographic electromagnetic exploration of underground three-dimensional target bodies; the polarization mode group of the raw multimodal data includes at least one of TE polarization mode and TM polarization mode; the regional field blocks of the raw multimodal data include at least one of near-field, transition zone and far-field; the parameters of the raw multimodal data include at least one of electromagnetic field component data, electromagnetic field gradient component data, DC resistivity data, induced polarizability data, frequency domain response information of natural field source and frequency domain response information of artificial field source. Based on a pre-defined classification and standardization strategy, outlier removal and standardization are performed on the original multimodal data to obtain standardized multimodal data.

[0010] In some embodiments of this application, based on the aforementioned scheme, in the recombined intermediate structure data, the TE mode and TM mode are respectively regarded as independent polarization mode groups; in each polarization mode group, the near region, transition region, and far region are respectively regarded as independent regional field blocks; in each regional field block, the electromagnetic field component data, electromagnetic field gradient component data, DC resistivity data, excited polarizability data, natural field source frequency domain response information, and artificial field source frequency domain response information are arranged in a preset order to form a parameter sequence; In a two-dimensional noisy image, the near region, transition region, and far region of the regional field block are distributed adjacently in the two-dimensional image, and each electromagnetic field component data and the corresponding gradient component data form a local feature pair in the two-dimensional image.

[0011] In some embodiments of this application, based on the foregoing scheme, when the total length of the reconstructed intermediate structure data does not meet the complete condition for two-dimensional image mapping, the method further includes: Based on the frequency domain attenuation model and parametric coupling coefficient of the same polarization mode and the same region field, combined with the dispersion characteristic model of the natural field source MT, the radiation characteristic model of the artificial field source, and the spatial distribution model between resistivity and polarizability, fill data is generated. The padding data is added to the mapping process of the two-dimensional noisy image so that the physical characteristics of the padding data are consistent with the original data in the same group; wherein, the length of the padding data is less than the total length of the recombined intermediate structure data.

[0012] In some embodiments of this application, based on the aforementioned scheme, after inversely restoring the two-dimensional noise estimation image, the padding data is removed, and the remaining data is used as the target multimodal data.

[0013] In some embodiments of this application, based on the aforementioned scheme, the multimodal attention deep learning network includes a multimodal feature extraction layer, a field-source adaptation attention layer, a multiscale feature fusion layer, and a noise modeling layer connected in sequence; Multi-dimensional features are extracted and noise modeling is performed using a multimodal attention deep learning network to obtain a two-dimensional noise estimation image, including: Through the multimodal feature extraction layer, a first preset series of progressive dilated convolutions are used to extract at least one of polarization features, region features, parametric features, physical property features, and field source features from a two-dimensional noisy image. Among them, the multimodal feature extraction layer removes the batch normalization operation in the dilated convolution, retains the batch normalization operation in the residual block, and adopts an adaptive batch normalization strategy in the physical property parameter feature extraction branch and the field source type feature extraction branch. By utilizing the polarization attention module, region attention module, source type attention module, and physical property parameter attention module in the source adaptation attention layer, the feature weights extracted by the multimodal feature extraction layer are dynamically allocated and focused. Through a multi-scale feature fusion layer, the output features of each module in the field source adaptation attention layer are extracted by convolution at the corresponding scale and the features at different scales are fused. Through the noise modeling layer, the multimodal noise distribution is modeled using alternating low-parameter residual blocks and standard residual blocks based on the residual learning mechanism, and a two-dimensional noise estimation image is output.

[0014] In some embodiments of this application, based on the aforementioned scheme, the multimodal attention deep learning network is obtained through the following steps: Noisy multimodal data and clean multimodal data from holographic electromagnetic exploration were used as sample sets to train the initial deep learning network; The initial deep learning network is optimized by a joint loss function to obtain a multimodal attention deep learning network. The joint loss function is used to determine at least one of the following: magnitude mean square error loss, electromagnetic field and gradient data consistency loss, phase consistency loss, physical property parameter consistency loss, and field source feature consistency loss.

[0015] Secondly, this application provides a multimodal holographic electromagnetic exploration data denoising device, comprising: A standardized multimodal data acquisition module is used to acquire standardized multimodal data of holographic electromagnetic exploration of underground three-dimensional targets; The two-dimensional noisy image processing module is used to reorganize standardized multimodal data according to the hierarchical structure composed of polarization mode groups, regional field blocks, and parameter sequences to obtain reorganized intermediate structure data; and to map the reorganized intermediate structure data into a two-dimensional noisy image of a preset dimension by adopting the arrangement strategy of polarization mode group association, regional field block adjacency, and parameter sequence coupling. The two-dimensional noise estimation image processing module is used to input a two-dimensional noisy image into a pre-trained multimodal attention deep learning network, extract multi-dimensional features and perform noise modeling through the multimodal attention deep learning network to obtain a two-dimensional noise estimation image; The target multimodal data processing module is used to perform difference operations on the two-dimensional noisy image and the two-dimensional noise estimation image to obtain a clean two-dimensional image; through the inverse mapping algorithm that matches the three-dimensional to two-dimensional fidelity reconstruction algorithm, the two-dimensional noise estimation image is reversed to obtain the holographic electromagnetic exploration denoised target multimodal data.

[0016] Thirdly, this application provides an electronic device, comprising: processor; Memory used to store processor-executable instructions; The processor is configured to execute a multimodal holographic electromagnetic exploration data denoising method as provided in the first aspect.

[0017] Fourthly, this application provides a non-transitory computer-readable storage medium that, when the instructions in the storage medium are executed by the processor of an electronic device, enables the electronic device to perform a multimodal holographic electromagnetic exploration data denoising method as provided in the first aspect.

[0018] One or more technical solutions provided in the embodiments of this application have at least the following technical effects or advantages: 1. By constructing a hierarchical structure of polarization mode groups, regional field blocks, and parameter sequences, holographic electromagnetic data is recombined and mapped into a two-dimensional noisy image. Combined with a multimodal attention deep learning network for noise modeling and difference inverse mapping, the problem of multimodal data coupling feature breakage is effectively solved. While accurately suppressing noise, the resolution and accuracy of subsequent three-dimensional inversion imaging are significantly improved.

[0019] 2. By employing a classification-based standardization strategy to remove outliers and standardize the raw multimodal data from holographic electromagnetic exploration, and implementing differentiated normalization based on the dimensional differences and statistical distribution characteristics of different physical parameters, numerical-dominated gradient updates are effectively avoided, providing high-quality input data with clear physical meaning and stable numerical distribution for deep learning networks.

[0020] 3. In a two-dimensional noisy image, the TE and TM polarization modes are independently grouped, and the near, transition, and far regions are distributed adjacently. The electromagnetic field components and their corresponding gradient components form local feature pairs. Through explicit physical prior structure injection, the convolutional neural network can directly use the local receptive field to perceive electromagnetic differential constraints, which significantly enhances the network's ability to learn multimodal coupling features.

[0021] 4. When the length of the recombined data does not meet the mapping integrity requirement, the filling data that conforms to physical laws is generated based on the frequency domain attenuation model, the field source radiation characteristic model and the physical property spatial distribution model. This effectively avoids false high-frequency abrupt changes introduced by non-physical filling boundaries, ensures the overall physical consistency of the input two-dimensional image, and improves the stability and generalization ability of network training.

[0022] 5. After reverse restoration of the denoised data, the padding data is accurately removed, and only the effective part from the original observation and after denoising and enhancement is retained. This avoids the misleading influence of artificial data based on model prediction on subsequent inversion interpretation, and ensures the authenticity and geological interpretability of the final target multimodal data.

[0023] 6. By extracting multi-scale features through progressive dilated convolution and removing batch normalization in dilated convolution to preserve the absolute magnitude of the field source, and by using a field source-adaptive attention layer to dynamically assign weights to polarization, region, field source type and physical property parameters, high-fidelity extraction and adaptive focusing of physical features of holographic electromagnetic data are achieved, significantly improving the accuracy of noise modeling.

[0024] 7. By optimizing the network through the combined loss of mean square error of amplitude, electromagnetic field and gradient consistency, phase consistency, physical property parameter consistency, and field source characteristic consistency, the physical prior knowledge of electromagnetic exploration is transformed into differentiable mathematical constraints, which force the network to follow Maxwell's equations and field source radiation laws during the denoising process, thereby achieving a dual improvement in signal fidelity and noise suppression capability. Attached Figure Description

[0025] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0026] Figure 1 A flowchart illustrating a method for denoising multimodal holographic electromagnetic exploration data provided in this application embodiment; Figure 2 A schematic diagram of a three-dimensional holographic electromagnetic multimodal data structure provided in an embodiment of this application; Figure 3 A schematic diagram of the multimodal attention deep learning network structure provided in the embodiments of this application; Figure 4 A comparison curve of the noise reduction effects of different methods provided in the embodiments of this application; Figure 5 A comparison curve of the consistency of electromagnetic field and gradient data before and after denoising provided in the embodiments of this application; Figure 6 A schematic diagram of the structure of a multimodal holographic electromagnetic exploration data denoising device provided in an embodiment of this application; Figure 7 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0027] To better understand the above technical solutions, the following will provide a detailed explanation of the technical solutions in conjunction with the accompanying drawings and specific implementation methods.

[0028] Example 1 like Figure 1 As shown in the figure, this embodiment provides a method for denoising multimodal holographic electromagnetic exploration data, including the following steps.

[0029] Step S100: Obtain standardized multimodal data of holographic electromagnetic exploration of the underground three-dimensional target body.

[0030] Standardized multimodal data is obtained by preprocessing the raw data collected by the alternating coverage observation system across the entire region. Because three-dimensional holographic electromagnetic exploration simultaneously acquires various information such as DC resistivity, excited polarization, time domain, frequency domain, and information from natural and artificial field sources, and includes the three components of the electromagnetic field and the three components of the gradient, the dimensional differences between different parameters are significant, and various environmental interferences exist. Therefore, it is necessary to first perform categorized standardization processing and physical rationality verification on the raw data, eliminating outliers that exceed physical laws, thereby obtaining standardized multimodal data with stable numerical distribution and clear physical meaning.

[0031] Step S200: The standardized multimodal data is reorganized according to the hierarchical structure composed of polarization mode groups, regional field blocks, and parameter sequences to obtain the reorganized intermediate structure data; the reorganized intermediate structure data is mapped into a two-dimensional noisy image of a preset dimension by adopting the arrangement strategy of polarization mode group association, regional field block adjacency, and parameter sequence coupling.

[0032] The original 3D holographic electromagnetic data logically presents a multidimensional tensor structure. Direct flattening or random mapping would disrupt the complementarity between polarization modes, the continuity of near-field and far-field regions, and the constraints of Maxwell's equations between the electromagnetic field and the gradient. Therefore, this invention employs a specific hierarchical recombination mechanism. First, data is grouped according to polarization modes (e.g., TE / TM). Then, within each group, regional field blocks are divided according to regional fields (near-field, transition field, far-field). Finally, these regional field blocks are sorted by parameter type. Polarization mode groups refer to data groups divided based on the polarization characteristics of the electromagnetic field, including TE polarization mode groups and TM polarization mode groups. TE mode refers to a mode where the electric field vector is perpendicular to the exploration profile direction (i.e., transverse electric mode), while TM mode refers to a mode where the electric field vector is parallel to the exploration profile direction (i.e., transverse magnetic mode). Regional field blocks refer to data blocks divided based on the relative relationship between the distance from the observation point to the field source and the skin depth, including near-field, transition field, and far-field regions. The near-field region refers to the area where the observation point is close to the field source, and the electromagnetic field decays primarily geometrically. The far-field region refers to the area where the observation point is far from the field source, and the electromagnetic field decays primarily exponentially. The transition region is the area between the near-field and far-field regions. A parameter sequence refers to a one-dimensional sequence formed by arranging multiple physical parameters in a predetermined order within each field block. Parameter types include, but are not limited to, electromagnetic field component data, electromagnetic field gradient component data, DC resistivity data, induced polarizability data, frequency domain response information of natural field sources (such as MT impedance tensor elements), and frequency domain response information of artificial field sources (such as CSAMT apparent resistivity and phase).

[0033] Based on this, such as Figure 2As shown, a specific arrangement strategy maps these intermediate structural data into a two-dimensional image, so that physically strongly coupled parameters are represented as adjacent pixels or local texture features in the two-dimensional image space. This mapping algorithm involves determining the width and height of the two-dimensional image and mapping each element of the intermediate structural data to a two-dimensional pixel location according to physical relationships. This transformation enables convolutional neural networks to utilize their powerful spatial feature extraction capabilities to perceive and process multidimensional coupling features in electromagnetic exploration data as if they were processing natural image textures.

[0034] Step S300: Input the two-dimensional noisy image into the pre-trained multimodal attention deep learning network, extract multi-dimensional features and perform noise modeling through the multimodal attention deep learning network to obtain a two-dimensional noise estimation image.

[0035] The multimodal attention deep learning network, as a trained specialized model, is internally configured to adapt to the specific distribution of holographic electromagnetic data. Unlike general image denoising networks, this network focuses not only on the visual smoothness of the image but also on the physical fidelity of the electromagnetic signal. After receiving a noisy two-dimensional image, the network does not directly output a clean signal. Instead, it employs residual learning to extract noise components from the complex background, generating a two-dimensional noise estimation image of the same size as the input. This design avoids the over-smoothing effect that might be introduced when the network directly reconstructs the signal, thus preserving subtle geological anomalies that are easily misidentified as noise. It should be noted that this embodiment focuses on describing the overall input-output functionality of the network and the processing logic for noise modeling. Specific details regarding the network's internal hierarchical structure, the implementation details of the attention mechanism, and the batch normalization strategy will be further explained in subsequent embodiments.

[0036] Step S400: Perform a difference operation between the two-dimensional noisy image and the two-dimensional noise estimation image to obtain a clean two-dimensional image; use an inverse mapping algorithm that matches the three-dimensional to two-dimensional fidelity reconstruction algorithm to reverse restore the two-dimensional noise estimation image to obtain the holographic electromagnetic exploration denoised target multimodal data.

[0037] Difference operations are based on the physical assumption that noisy signal = clean signal + noise. By subtracting the estimated noise image from the original noisy image, the clean signal component is separated. However, the resulting clean 2D image is still image-domain data and lacks direct geophysical interpretability. Therefore, inverse mapping is required. Inverse mapping is the reverse process of mapping algorithms, which extracts the original data position corresponding to each pixel in the image to restore the 2D image to a one-dimensional or three-dimensional data sequence of the original structure. Inverse mapping is not only a shape transformation that restores a 2D matrix to a 3D tensor, but more importantly, it accurately restores the spatial positions in the image back to the corresponding polarization modes, regional field blocks, and parameter sequence indices, ensuring that the output target multimodal data is physically and semantically aligned with the input standardized multimodal data. This guarantees that the denoised data remains in the standard electromagnetic exploration data format and can be directly used for 3D inversion imaging or geological interpretation.

[0038] Example 2 Based on Example 1, this example further explains the specific process of acquiring standardized multimodal data of holographic electromagnetic exploration of underground three-dimensional target bodies, as well as the detailed strategies for hierarchical reconstruction and two-dimensional mapping.

[0039] Acquiring standardized multimodal data of holographic electromagnetic exploration of underground three-dimensional target bodies includes: acquiring raw multimodal data of holographic electromagnetic exploration of underground three-dimensional target bodies; and performing outlier removal and standardization processing on the raw multimodal data based on a preset classification standardization strategy to obtain standardized multimodal data.

[0040] The polarization mode group of the original multimodal data includes at least one of the TE polarization mode and the TM polarization mode. The regional field blocks of the original multimodal data include at least one of the near-field, transition, and far-field regions. The parameters of the original multimodal data include at least one of the electromagnetic field component data, electromagnetic field gradient component data, DC resistivity data, induced polarizability data, natural field source frequency domain response information, and artificial field source frequency domain response information. Since the holographic electromagnetic exploration system simultaneously collects the above-mentioned multiple physical quantities, the parameters not only have huge differences in dimensions (for example, the electric field amplitude may be on the order of millivolts, while the resistivity is on the order of ohmmeters), but also have completely different statistical distribution characteristics. If the original data is directly input into a deep learning network, the parameters with larger values ​​will dominate the gradient update, causing the network to ignore weak but critical geological anomaly signals. Therefore, this embodiment performs outlier removal and standardization processing on the original multimodal data based on a preset classification and standardization strategy to obtain standardized multimodal data.

[0041] Different standardization methods were adopted for different types of parameters to adapt to their physical characteristics. For example, for electromagnetic field amplitude data, since it usually follows a normal distribution and relative changes are of interest, Z-Score standardization is preferred. , where x is the original electromagnetic field amplitude, μ is the mean amplitude, and σ is the standard deviation of the amplitude. That is, the mean and standard deviation are used to map the data to a zero-mean unit variance space to preserve the relative strength of the signal. For electromagnetic field gradient component data, since it is limited by the observation geometry and has a clear physical value range, it is preferable to use Min-Max normalization to linearly map it to a fixed interval to maintain the directionality and proportionality of the gradient. For DC resistivity data, considering that the resistivity of underground media often spans several orders of magnitude, logarithmic normalization can effectively compress the dynamic range and make high and low resistivity anomalies comparable in the network feature space. For excitation polarizability data, since it is itself a percentage or dimensionless ratio, it is usually normalized to the [0,1] interval. For phase data, special attention should be paid to its periodicity characteristics, and it should be normalized to the [-π,π] interval to avoid numerical abrupt changes caused by phase entanglement. In addition, before standardization, it is necessary to verify the physical rationality by combining the regional field characteristics and the differences in polarization mode response. For example, outliers that do not conform to the law of rapid decay of near field amplitude or the law of linear change of far field phase should be removed, so as to ensure the physical credibility of data quality from the source.

[0042] After completing the data standardization, the following further explains how to reorganize the standardized multimodal data according to the hierarchical structure composed of polarization mode groups, regional field blocks, and parameter sequences to obtain the reorganized intermediate structure data.

[0043] The original three-dimensional holographic electromagnetic data logically constitutes a multidimensional tensor. The reassembly is not a simple data stacking, but a structured rearrangement that strictly follows physical correlations. Specifically, in the reassembled intermediate structure data, the TE mode and TM mode are treated as independent polarization mode groups. This is because the two modes have different sensitivities to the electrical interface of the underground medium, and independent grouping is beneficial for the network to learn anisotropic characteristics separately. In each polarization mode group, the near region, transition region, and far region are treated as independent regional field blocks, and their adjacency relationship in the data sequence is maintained. This arrangement utilizes the continuity of the electromagnetic field as the transmission and reception distance changes, enabling the network to use the stable characteristics of the transition region to help correct noise interference in the near and far regions. In each regional field block, the electromagnetic field component data, electromagnetic field gradient component data, DC resistivity data, excited polarizability data, natural field source frequency domain response information, and artificial field source frequency domain response information are arranged in a preset order to form a parameter sequence.

[0044] This embodiment employs an arrangement strategy of polarization mode group association, regional field block adjacency, and parametric sequence coupling to map the recombined intermediate structure data into a two-dimensional noisy image of a preset dimension. This mapping process transforms physical coupling correlation into spatial proximity in a two-dimensional image. For example... Figure 2As shown, in the generated two-dimensional noisy image, the near region, transition region, and far region of the regional field block are distributed adjacently in the two-dimensional image, and each electromagnetic field component data and its corresponding gradient component data form a local feature pair in the two-dimensional image. For example, the electric field Ex component and its corresponding spatial gradient ( The kernels are arranged in adjacent rows or columns within the image matrix. The underlying mechanism of this design is that the basic operational units (convolutional kernels) of a convolutional neural network have local receptive fields. When two data points with a strong physical constraint relationship (such as the field-gradient relationship described by Maxwell's equations) are adjacent in the image space, a standard 3×3 or 5×5 convolutional kernel can simultaneously perceive both variables in a single forward propagation and directly extract their differential or integral correlation features. Conversely, if random shuffling or simple flattening is used for mapping, the field and gradient will be far apart in the image. The network needs extremely deep layers or a very large receptive field to reconstruct this connection, which not only increases the training difficulty but also easily leads to the loss of physical constraints during denoising. Therefore, the specific arrangement strategy in this embodiment essentially injects an explicit physical prior structure into the deep learning network, enabling it to implicitly perform feature extraction operations that conform to the physical laws of electromagnetic exploration while processing image texture, thereby significantly improving the physical fidelity of the denoising results.

[0045] During hierarchical restructuring and two-dimensional mapping, situations may arise where the data dimensions do not perfectly match the network input requirements. To address this issue, this embodiment also provides a physically-guided filling and culling mechanism.

[0046] Specifically, when the total length of the recombined intermediate structure data does not meet the complete condition for two-dimensional image mapping, the method further includes: generating filling data based on the frequency domain attenuation model and parametric coupling coefficient of the same polarization mode and the same region field, combined with the dispersion characteristic model of the natural field source MT, the radiation characteristic model of the artificial field source, and the spatial distribution model between resistivity and polarizability.

[0047] In 3D holographic electromagnetic exploration, the variations of various parameters with frequency or spatial location follow strict geophysical laws. For example, the electromagnetic field amplitude decays exponentially with increasing frequency, and the phase changes linearly with frequency. Simply using numerical padding introduces non-physical high-frequency abrupt changes or gradient breaks at the padding boundaries. These spurious features can be misidentified by deep learning networks as valid signals or specific noise patterns, leading to distorted denoising results. Therefore, by constructing a generative model that conforms to physical laws, the padding data becomes a natural extension of the original signal in the physical domain, thus ensuring the overall physical consistency of the input data.

[0048] Furthermore, the padding data is added to the mapping process of the two-dimensional noisy image so that the physical characteristics of the padding data are consistent with the original data in the same group; wherein, the length of the padding data is less than the total length of the recombined intermediate structure data.

[0049] In practice, the generation of fill data is not a simple application of a single formula, but rather a dynamic selection of the corresponding physical model based on the data type of the current location to be filled. For example, for electromagnetic field components belonging to the MT mode of natural field sources, a power-law decay model based on the dispersion characteristics of magnetotelluric signals is used for extrapolation; for artificial field source data, the theoretical response value is calculated based on the dipole radiation field model; and for physical properties such as DC resistivity and excited polarizability, interpolation prediction is performed using a spatial correlation model of adjacent measurement points. Simultaneously, a parametric coupling coefficient is introduced to ensure that the generated fill data not only conforms to physical laws but also maintains the correct proportional relationship and phase difference with other parameters in the same group, avoiding disruption of the inherent constraints between multimodal data due to fill. Regarding the fill length limitation, it is set to be less than the total length of the recombined intermediate structure data. This ensures that the fill data only plays an auxiliary role in filling dimensions and does not dominate the dataset. If the fill ratio is too high, the network training will primarily learn the characteristics of artificial physical models rather than the actual response of the underground medium, which will severely weaken the model's generalization ability and its suppression of real complex noise. It is generally preferable to keep the fill length within 10% of the total length in order to preserve the main position of the original observation information to the greatest extent possible while ensuring the normal operation of the network.

[0050] After inversely reconstructing the 2D noise-estimated image, the padding data needs to be removed, and the remaining data is used as the target multimodal data. Although the padding data simulates real physical features as closely as possible during generation, it is essentially still a model-based prediction and does not contain any actual subsurface geological information. Retaining this data in the final output would not only increase the burden of storing and transmitting invalid data, but could also lead to it being mistakenly interpreted as a real anomaly in subsequent 3D inversion imaging, resulting in erroneous geological conclusions. Therefore, when executing the inverse transform algorithm, which is the inverse of the mapping process, the padding data segments are precisely stripped from the reconstructed 3D tensor based on pre-recorded padding location indices or length markers, retaining only the effective data portion derived from the original observations and enhanced with noise.

[0051] Example 3 Based on the foregoing embodiments, this embodiment further details the specific architecture and internal processing mechanism of the multimodal attention deep learning network.

[0052] like Figure 3As shown, the multimodal attention deep learning network comprises a multimodal feature extraction layer, a field-source adaptation attention layer, a multi-scale feature fusion layer, and a noise modeling layer connected in sequence. By extracting multi-dimensional features and performing noise modeling through the multimodal attention deep learning network, a two-dimensional noise estimation image is obtained, including: Through a multimodal feature extraction layer, a first preset series of progressively dilated convolutions is used to extract at least one of the following from a two-dimensional noisy image: polarization features, regional features, parametric features, physical property features, and field source features. In this embodiment, the first preset series is preferably a sequence of dilated convolution kernels with dilation rates of 1, 2, 4, and 8 in sequence. This progressive design has a clear physical orientation: the convolution kernel with a dilation rate of 1 focuses on capturing high-frequency details and local gradient changes in the near-field; the convolution kernels with dilation rates of 2 and 4 cover the medium-scale features of the transition zone and are used to correlate electromagnetic field components with their spatial derivatives; the convolution kernel with a dilation rate of 8 provides a sufficiently large receptive field to perceive the low-frequency trends and global background response of the far-field. Through this multi-scale parallel or serial extraction mechanism, the network can synchronously adapt to the large physical scale differences from the near field to the far field in the full-area alternating coverage observation system without reducing spatial resolution, avoiding the loss of weak geological anomaly information caused by traditional pooling operations.

[0053] In this multimodal feature extraction layer, batch normalization (BN) is removed from dilated convolutions but retained in residual blocks. An adaptive BN strategy is employed in the physical property feature extraction branch and the source type feature extraction branch. In general computer vision tasks, BN is typically used to accelerate convergence and prevent overfitting. However, in electromagnetic exploration data processing, BN forces the mean of the feature map back to zero and the variance back to one. While this stabilizes the numerical distribution, it inevitably smooths out the absolute magnitude information of the electromagnetic field amplitude and the radiation attenuation characteristics of the source. These absolute magnitudes are crucial for determining the distance and size of the target in inversion imaging. Therefore, BN is decisively removed from the dilated convolution layers that directly process the original signal features to preserve the physical fidelity of the source features; while BN is retained within the residual blocks, which are mainly used for feature transformation and nonlinear mapping, to maintain the training stability of the deep network. Furthermore, considering that physical properties such as DC resistivity and excitation polarizability often exhibit severe non-Gaussian or long-tailed distributions, and that the statistical characteristics of natural and artificial field sources differ significantly, the standard fixed parameter BN is difficult to adapt. Therefore, an adaptive batch normalization strategy is introduced in the corresponding feature extraction branch. This strategy dynamically adjusts the normalization parameter or affine transformation coefficient based on the real-time statistical characteristics of the input data, thereby ensuring training efficiency while avoiding distortion of physical property features caused by forced normalization.

[0054] After feature extraction, the polarization attention module, region attention module, source type attention module, and physical property parameter attention module in the source adaptation attention layer are used to dynamically allocate and focus the feature weights of the features extracted by the multimodal feature extraction layer. Unlike general attention mechanisms that only focus on salient regions of the image, each module here corresponds to a clear geophysical dimension: the polarization attention module is used to dynamically balance the response differences between TE and TM modes under different geological structures, preventing a certain mode from being ignored by the network due to low signal-to-noise ratio; the region attention module automatically adjusts the weights according to the energy distribution characteristics of the near, transition, and far regions, strengthening the feature expression of weak signal regions; the source type attention module distinguishes the spectral characteristics of natural and artificial fields, suppressing systematic interference introduced by specific source fields; and the physical property parameter attention module focuses on the feature channels corresponding to resistivity and polarizability anomalies. Through this multi-dimensional dynamic weighting, the network can automatically identify and focus on the most geologically significant information during processing, rather than blindly smoothing all high-frequency components.

[0055] Subsequently, a multi-scale feature fusion layer is used to extract the output features of each module in the field source adaptation attention layer by convolution at the corresponding scale, and the features at different scales are fused. This layer typically contains parallel processing paths with convolutional kernels of various sizes, such as 1×1, 3×3, and 5×5. 1×1 convolutions are used to integrate cross-channel physical properties and field source semantic information; 3×3 convolutions are used to extract local electromagnetic field and gradient coupling textures; and 5×5 convolutions are used to capture regional field continuity features over a larger area. The fused feature map not only contains rich multi-scale contextual information, but also eliminates redundant noise through the attention mechanism, providing a high-quality feature representation for subsequent noise modeling.

[0056] Finally, through the noise modeling layer, using alternating low-parameter residual blocks and standard residual blocks, the multimodal noise distribution is modeled based on the residual learning mechanism, outputting a two-dimensional noise estimation image. This embodiment adopts the residual learning paradigm, that is, allowing the network to directly learn noise rather than pure signals. This is because in holographic electromagnetic data, noise is often more statistically regular than the complex response of the underground medium, and is easier to fit. In terms of network structure, low-parameter residual blocks (such as reducing the number of channels or convolutional layers) are mainly used to capture high-frequency, random environmental background noise and instrument electronic noise. These types of noise have simple structures but are widely distributed; while standard residual blocks have stronger fitting capabilities and are used to model structured noise caused by multimodal coupling, residual field source effects, etc. The design of alternating connections between the two avoids the overfitting of weak geological signals that may be caused by all standard residual blocks (that is, misjudging effective signals as noise removal), and overcomes the problem of underfitting complex noise by all low-parameter blocks, thus achieving the best balance between accurate noise removal and complete signal preservation. The final output two-dimensional noise estimation image has the same size as the input two-dimensional noisy image, providing accurate noise components for the difference calculation in subsequent steps.

[0057] The training process of this network and the joint loss function used will be explained in detail below.

[0058] The multimodal attention deep learning network is obtained through the following steps: the initial deep learning network is trained using noisy multimodal data and clean multimodal data from holographic electromagnetic exploration as sample sets; the initial deep learning network is optimized by a joint loss function to obtain the multimodal attention deep learning network.

[0059] In this embodiment, the sample set does not rely solely on limited measured data. Instead, it generates pure multimodal data covering various typical geological structures such as sedimentary rocks, igneous rocks, and metamorphic rocks through high-precision three-dimensional electromagnetic forward modeling. This data is then superimposed with measured environmental noise and instrument noise of different signal-to-noise ratios and types, thereby constructing a massive and precisely labeled combination of noisy and pure data pairs. This ensures that the network can learn the noise distribution patterns under various geological scenarios, from simple to complex, avoiding overfitting problems caused by a single training sample. This allows the model to maintain robust denoising performance even when facing unknown geological conditions in actual exploration.

[0060] To overcome the shortcomings of traditional deep learning denoising methods that focus only on image visual quality while neglecting geophysical properties, this invention designs a dedicated joint loss function. This joint loss function determines at least one of the following: magnitude mean square error loss, electromagnetic field and gradient data consistency loss, phase consistency loss, physical property parameter consistency loss, and field source feature consistency loss. Essentially, this joint loss function transforms the prior physical knowledge of electromagnetic exploration into differentiable mathematical constraints, forcing the network to not only pursue numerical approximation during optimization but also adhere to the intrinsic physical laws of the electromagnetic field.

[0061] In a preferred embodiment, the complete expression for the joint loss function is: ; in, The amplitude mean square error loss is used to measure the overall deviation in amplitude between the denoised data and the real data, ensuring the basic fidelity of the signal energy. ,in A is the denoised amplitude, and A is the true amplitude. The total number of data samples; To compensate for the consistency loss between electromagnetic field and gradient data, Where x represents the spatial coordinate direction, This represents the gradient operator along the corresponding spatial coordinate direction. This represents the gradient response of the denoised data. It represents the gradient response of the real data; by calculating the difference between the gradient operator response of the denoised data and the directly observed gradient data, the network is forced to actively maintain the physical consistency of the field and gradient during denoising. For phase consistency loss, ,in, This represents the phase value after denoising. This represents the true phase value. Given that phase information is extremely sensitive to underground electrical interfaces and is easily affected by noise, resulting in periodic wrapping or abrupt changes, this loss is specifically constrained for the phase components to ensure that the denoised phase curve is continuous, smooth, and free of spurious jumps. For the loss of consistency of physical property parameters, ,in, This represents the resistivity parameter after noise reduction. This represents the actual resistivity parameter. This represents the excitation polarization parameter after denoising. This represents the actual excitation polarizability parameter; this loss is used to constrain the value range and spatial distribution of physical properties such as DC resistivity and excitation polarizability, and to prevent the network from generating non-physical values ​​such as negative resistivity or polarizability that exceeds the reasonable range. For the consistency loss of field-source features, , The source feature vector of the original real data. The field source feature vector of the denoised data is used to preserve the spectral characteristics and radiation patterns of natural and artificial field sources, and to prevent the network from confusing the response characteristics of different field sources or introducing false field source effects during the denoising process. , , , These are the weighting coefficients for each type of loss, used to balance magnitude accuracy, gradient fidelity, and phase consistency.

[0062] Through the collaborative optimization of the aforementioned multi-dimensional joint loss function, the network is trained into a specialized model that understands both image processing and electromagnetic physics. In terms of specific training configuration, this embodiment preferably uses the Adam optimizer, with an initial learning rate set to 1×10⁻⁶. -3 The system employs a cosine annealing strategy for attenuation, with a batch size of 128 and a total training epoch of no less than 220 epochs. It should be understood that the above hyperparameters are merely a preferred example. In practical applications, those skilled in the art can adaptively adjust them according to hardware computing power and data scale, as long as network convergence and all indicators of the loss function reach the preset thresholds.

[0063] Example 4 To further verify the effectiveness of the method proposed in this invention in practical engineering applications, this embodiment takes a typical three-dimensional holographic electromagnetic exploration project in a metal mining area as an example.

[0064] In this application scenario, the exploration target is a copper-nickel sulfide ore body buried at a depth of approximately 300 to 800 meters. The geological environment is complex, with strong human-induced electromagnetic interference at the surface and significant anisotropic characteristics in the subsurface medium. A full-area alternating coverage observation system was used to collect holographic multimodal data, including TE / TM dual-polarization modes, near / transition / far-area multi-region field information, and electromagnetic field three-components and gradient three-components. The multimodal holographic electromagnetic exploration data denoising method of this invention was used to process this batch of data, and a comparative analysis was conducted with commonly used wavelet thresholding, Kalman filtering, and general convolutional neural network (CNN) denoising methods without physical constraints to visually demonstrate the technical advantages of this invention.

[0065] like Figure 4 As shown, the comparison curves illustrate the denoising effect of the electromagnetic component (Ex) of the electric field in the transition region under TE mode. Figure 4As shown in the upper part, from the amplitude response perspective, the signal-to-noise ratio (SNR) of the original noisy data is only 16 dB, with noise overwhelming the weak anomalous signals of the deep ore body. After processing by the method of this invention, the SNR is improved to 40 dB, effectively suppressing random noise and fully preserving the high-frequency details reflecting the ore body boundary. In contrast, although the wavelet thresholding method and Kalman filtering method improve the SNR to 28 dB and 26 dB respectively, they exhibit obvious over-smoothing in the high-frequency band, resulting in the loss of some valuable geological information. Although the general CNN method improves the SNR to 30 dB, due to the lack of targeted adaptation to the physical distribution of electromagnetic data, some structured interference still remains in the strong noise frequency band.

[0066] Regarding the preservation of phase information, such as Figure 4 As shown in the lower half, the phase curves of the original data are chaotic. The phase error after general CNN processing is approximately 0.59 rad, and that of wavelet thresholding is 0.65 rad. This is mainly because these methods fail to establish a physical coupling constraint between phase and amplitude, leading to artificial phase distortion introduced during denoising. In contrast, this invention forces a constraint on the network output through a phase consistency loss term in the joint loss function, reducing the phase error of the denoised data to 0.17 rad, with the curve shape highly consistent with the actual geological model. This superior phase fidelity directly ensures the accuracy of target depth determination in subsequent frequency domain inversion, avoiding depth interpretation bias caused by phase distortion.

[0067] Furthermore, such as Figure 5 As shown, the electromagnetic field components (Ex) before and after denoising are illustrated with their corresponding gradient components (Ex). The consistency between the electromagnetic field and gradient data is compared. In three-dimensional holographic electromagnetic exploration, the electromagnetic field and gradient data satisfy a strict differential relationship described by Maxwell's equations. This physical coupling is a key criterion for distinguishing real geological anomalies from random noise or spurious source effects. Before denoising, the Ex component is affected by multiple noise interferences. The gradient components exhibit significant waveform differences, with a correlation coefficient of only 0.62, indicating severe damage to the internal physical structure of the data. Direct use for inversion is highly susceptible to generating spurious constructs. After processing using the method of this invention, the trends of the two curves achieve a high degree of overlap, significantly improving smoothness, and the correlation coefficient jumps to 0.98. This qualitative change is not a simple result of numerical fitting, but rather benefits from the electromagnetic field and gradient data consistency loss introduced in this invention during the network training phase. This forces the deep learning network to actively repair and maintain the physical correlation between the field and the gradient while suppressing noise. In contrast, traditional methods, due to the independent processing of modalities, often disrupt this natural coupling relationship, resulting in visually smoother data after denoising, but no longer consistent in physical properties.

[0068] In summary, by transforming abstract algorithm optimization metrics into industry-standard exploration performance metrics, this invention demonstrates that it can not only significantly improve the signal-to-noise ratio of data but also effectively eliminate spurious field source effects caused by noise, providing a solid data foundation for subsequent high-resolution imaging of three-dimensional anisotropic resistivity and polarizability.

[0069] Example 5 Based on the same inventive concept, this embodiment provides as follows: Figure 6 This invention illustrates a multimodal holographic electromagnetic exploration data denoising device. This device is a concrete product embodiment of the aforementioned method embodiments, designed to achieve fully automated processing of holographic electromagnetic exploration data from acquisition and preprocessing to high-precision denoising output through a modular architecture design.

[0070] The device includes a standardized multimodal data acquisition module 61, a two-dimensional noisy image processing module 62, a two-dimensional noise estimation image processing module 63, and a target multimodal data processing module 64. It should be understood that each module in this embodiment can be a functional unit implemented through software code, a hardware entity integrating dedicated circuitry, or an embedded processing unit combining software and hardware, as long as it can achieve the following data flow and processing functions.

[0071] The standardized multimodal data acquisition module 61 is used to acquire standardized multimodal data from holographic electromagnetic exploration of underground three-dimensional targets. As the data entry point for the entire device, this module not only reads raw multimodal data from the alternating coverage observation system or storage medium across the entire area, but also incorporates data cleaning and standardization logic units. During execution, this module, based on a preset categorized standardization strategy, performs outlier removal and numerical normalization on the input TE / TM polarization modes, near / transition / far-field blocks, and various electromagnetic parameters, eliminating dimensional differences between different physical quantities. The output standardized multimodal data meets the input requirements of subsequent deep learning networks in terms of numerical distribution, while maintaining the integrity and validity of the original exploration data in terms of physical semantics, providing a high-quality data baseline for subsequent processing.

[0072] The two-dimensional noisy image processing module 62 is used to reorganize standardized multimodal data according to a hierarchical structure composed of polarization mode groups, regional field blocks, and parameter sequences to obtain reorganized intermediate structure data. It employs an arrangement strategy of polarization mode group association, regional field block adjacency, and parameter sequence coupling to map the reorganized intermediate structure data into a two-dimensional noisy image of a preset dimension. Internally, it first uses a reorganization subunit to rearrange discrete three-dimensional tensor data into an ordered intermediate sequence according to the hierarchical logic of polarization mode groups, regional field blocks, and parameter sequences. Then, a mapping subunit converts this intermediate sequence into a two-dimensional matrix form. During this data flow process, the module ensures that parameters with strong physical correlations (such as electric field components and their gradient components) are spatially adjacent in the generated two-dimensional noisy image. Furthermore, when the data length does not meet the mapping conditions, the module can also be configured to perform a physical model-based padding operation to generate padding data that conforms to frequency domain attenuation characteristics to fill in the dimensions, thereby ensuring the regularity of the output image. This structural transformation enables subsequent neural networks to directly utilize convolution operations to perceive the inherent physical laws of electromagnetic data.

[0073] The two-dimensional noise estimation image processing module 63 is used to input a two-dimensional noisy image into a pre-trained multimodal attention deep learning network. The multimodal attention deep learning network extracts multi-dimensional features and models noise to obtain a two-dimensional noise estimation image. Internally, it loads the model parameters of the multimodal attention deep learning network provided in the aforementioned embodiments. In terms of data flow, this module receives image data output from the two-dimensional noisy image processing module and separates the noise component from the complex noisy background through hierarchical operations such as progressive dilated convolution, field-source adapted attention mechanism, and multi-scale feature fusion. Unlike general image processing modules, this module dynamically adapts to the field-source characteristics and physical property distribution of the holographic electromagnetic data during processing through a specific batch normalization strategy and attention weight allocation, ensuring that the generated two-dimensional noise estimation image contains only noise information while preserving geological anomaly response features to the greatest extent. The output of this module is consistent with the input two-dimensional noisy image in dimensionality, providing accurate noise components for subsequent interpolation operations.

[0074] The target multimodal data processing module 64 performs a difference operation between the two-dimensional noisy image and the two-dimensional noise-estimated image to obtain a clean two-dimensional image. Then, using an inverse mapping algorithm matched with the three-dimensional to two-dimensional fidelity reconstruction algorithm, it reverse-engineers the two-dimensional noise-estimated image to obtain the holographic electromagnetic exploration denoised target multimodal data. It should be noted that the inverse mapping object here typically refers to the clean two-dimensional image obtained after the difference operation in the actual processing logic, or to auxiliary restoration processing of the noise-estimated image under a specific residual learning paradigm, losslessly converting the image domain processing result back to the physical data domain. For example, this module first performs pixel-level subtraction to remove noise, and then calls a mapping algorithm that is strictly inverse of the two-dimensional noisy image processing module to accurately backfill each pixel in the two-dimensional matrix to the corresponding polarization mode, region field block, and parameter sequence index position. If padding data is introduced in the preceding steps, this module will also perform a culling operation based on pre-stored padding markers to ensure that the final output target multimodal data does not contain any artificially constructed false information.

[0075] Through the modular architecture described above, this invention encapsulates complex electromagnetic exploration data denoising methods into a standardized data processing product. The modules interact through clearly defined data interfaces, ensuring both logical coherence of the processing flow and high system flexibility. For example, in practical engineering applications, the two-dimensional noise estimation image processing module 63 can be independently deployed on a GPU accelerator card to improve inference speed, while other modules run on a general-purpose CPU for data scheduling, thereby achieving optimized allocation of hardware and software resources.

[0076] Example 6 Based on the same inventive concept, this embodiment provides as follows: Figure 7 An electronic device is shown, which is used to implement the multimodal holographic electromagnetic exploration data denoising method of the present invention at the hardware entity level.

[0077] The electronic device includes a processor 71 and a memory 72 for storing processor-executable instructions; wherein the processor 71 is configured to execute a multimodal holographic electromagnetic exploration data denoising method as provided in the foregoing embodiments.

[0078] Furthermore, the present invention also provides a non-transitory computer-readable storage medium, wherein when the instructions in the storage medium are executed by the processor 71 of the electronic device, the electronic device is able to execute a multimodal holographic electromagnetic exploration data denoising method as provided in the foregoing embodiments.

[0079] Since the electronic device described in this embodiment is an electronic device used to implement the information processing method in the embodiments of this application, those skilled in the art can understand the specific implementation methods and various variations of the electronic device in this embodiment based on the information processing method described in the embodiments of this application. Therefore, how the electronic device implements the method in the embodiments of this application will not be described in detail here. Any electronic device used by those skilled in the art to implement the information processing method in the embodiments of this application falls within the scope of protection of this application.

[0080] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0081] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0082] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0083] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1The steps of the function specified in one or more boxes.

[0084] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0085] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for denoising multimodal holographic electromagnetic exploration data, characterized in that, include: Acquire standardized multimodal data of holographic electromagnetic exploration of underground three-dimensional targets; The standardized multimodal data is reorganized according to the hierarchical structure composed of polarization mode groups, regional field blocks, and parameter sequences to obtain reorganized intermediate structure data; the reorganized intermediate structure data is mapped into a two-dimensional noisy image of a preset dimension by adopting the arrangement strategy of polarization mode group association, regional field block adjacency, and parameter sequence coupling. The two-dimensional noisy image is input into a pre-trained multimodal attention deep learning network. The multimodal attention deep learning network extracts multi-dimensional features and performs noise modeling to obtain a two-dimensional noise estimation image. The two-dimensional noisy image and the two-dimensional noise estimation image are interpolated to obtain a clean two-dimensional image. The two-dimensional noise estimation image is then reverse-reconstructed using an inverse mapping algorithm that matches the three-dimensional to two-dimensional fidelity reconstruction algorithm to obtain the holographic electromagnetic exploration denoised target multimodal data.

2. The method as described in claim 1, characterized in that, The standardized multimodal data obtained from the holographic electromagnetic exploration of the underground three-dimensional target body includes: The original multimodal data of holographic electromagnetic exploration of underground three-dimensional target bodies are obtained; the polarization mode group of the original multimodal data includes at least one of TE polarization mode and TM polarization mode; the regional field block of the original multimodal data includes at least one of near-field, transition zone and far-field; the parameters of the original multimodal data include at least one of electromagnetic field component data, electromagnetic field gradient component data, DC resistivity data, induced polarizability data, natural field source frequency domain response information and artificial field source frequency domain response information. Based on a preset classification and standardization strategy, outlier removal and standardization processing are performed on the original multimodal data to obtain standardized multimodal data.

3. The method as described in claim 1, characterized in that, In the recombined intermediate structure data, the TE mode and TM mode are respectively treated as independent polarization mode groups; in each polarization mode group, the near region, transition region, and far region are respectively treated as independent regional field blocks; in each regional field block, the electromagnetic field component data, electromagnetic field gradient component data, DC resistivity data, excited polarizability data, natural field source frequency domain response information, and artificial field source frequency domain response information are arranged in a preset order to form a parameter sequence; In the two-dimensional noisy image, the near region, transition region, and far region of the regional field block are distributed adjacently in the two-dimensional image, and each electromagnetic field component data and the corresponding gradient component data form a local feature pair in the two-dimensional image.

4. The method as described in claim 1 or 3, characterized in that, If the total length of the reconstructed intermediate structure data does not meet the complete condition for two-dimensional image mapping, the method further includes: Based on the frequency domain attenuation model and parametric coupling coefficient of the same polarization mode and the same region field, combined with the dispersion characteristic model of the natural field source MT, the radiation characteristic model of the artificial field source, and the spatial distribution model between resistivity and polarizability, fill data is generated. The padding data is added to the mapping process of the two-dimensional noisy image so that the physical characteristics of the padding data are consistent with the original data in the same group; wherein the length of the padding data is less than the total length of the recombined intermediate structure data.

5. The method as described in claim 4, characterized in that, After inversely restoring the two-dimensional noise-estimated image, the padding data is removed, and the remaining data is used as the target multimodal data.

6. The method as described in claim 1, characterized in that, The multimodal attention deep learning network includes a multimodal feature extraction layer, a field-source adaptation attention layer, a multi-scale feature fusion layer, and a noise modeling layer connected in sequence. The step of extracting multi-dimensional features and performing noise modeling through the multimodal attention deep learning network to obtain a two-dimensional noise estimation image includes: Through the multimodal feature extraction layer, at least one of polarization features, region features, parametric features, physical property features, and field source features in the two-dimensional noisy image is extracted using a first preset series of progressive dilated convolutions. The multimodal feature extraction layer removes batch normalization operations in the dilated convolutions, retains batch normalization operations in the residual blocks, and adopts an adaptive batch normalization strategy in the physical property parameter feature extraction branch and the field source type feature extraction branch. The polarization attention module, region attention module, source type attention module, and physical property parameter attention module in the source adaptation attention layer are used to dynamically allocate and focus the feature weights of the features extracted by the multimodal feature extraction layer. Through the multi-scale feature fusion layer, the output features of each module in the field source adaptation attention layer are extracted by convolution at the corresponding scale and the features at different scales are fused. The noise modeling layer utilizes alternating low-parameter residual blocks and standard residual blocks to model the multimodal noise distribution based on a residual learning mechanism, and outputs the two-dimensional noise estimation image.

7. The method as described in claim 1, characterized in that, The multimodal attention deep learning network is obtained through the following steps: Noisy multimodal data and clean multimodal data from holographic electromagnetic exploration were used as sample sets to train the initial deep learning network; The multimodal attention deep learning network is obtained by optimizing the initial deep learning network through a joint loss function; the joint loss function is used to determine at least one of the following: magnitude mean square error loss, electromagnetic field and gradient data consistency loss, phase consistency loss, physical property parameter consistency loss, and field source feature consistency loss.

8. A multimodal holographic electromagnetic exploration data denoising device, characterized in that, include: A standardized multimodal data acquisition module is used to acquire standardized multimodal data of holographic electromagnetic exploration of underground three-dimensional targets; The two-dimensional noisy image processing module is used to reorganize the standardized multimodal data according to the hierarchical structure composed of polarization mode groups, regional field blocks, and parameter sequences to obtain reorganized intermediate structure data; and to map the reorganized intermediate structure data into a two-dimensional noisy image of a preset dimension by adopting an arrangement strategy of polarization mode group association, regional field block adjacency, and parameter sequence coupling. The two-dimensional noise estimation image processing module is used to input the two-dimensional noisy image into a pre-trained multimodal attention deep learning network, extract multi-dimensional features and perform noise modeling through the multimodal attention deep learning network to obtain a two-dimensional noise estimation image; The target multimodal data processing module is used to perform a difference operation between the two-dimensional noisy image and the two-dimensional noise estimation image to obtain a clean two-dimensional image; and to reverse restore the two-dimensional noise estimation image to obtain the holographic electromagnetic exploration denoised target multimodal data through an inverse mapping algorithm that matches the three-dimensional to two-dimensional fidelity reconstruction algorithm.

9. An electronic device, characterized in that, include: processor; Memory used to store the processor's executable instructions; The processor is configured to execute a multimodal holographic electromagnetic exploration data denoising method as described in any one of claims 1 to 7.

10. A non-transitory computer-readable storage medium, characterized in that, When the instructions in the storage medium are executed by the processor of the electronic device, the electronic device is able to perform a method for denoising multimodal holographic electromagnetic exploration data as described in any one of claims 1 to 7.

Images