Geophysical electromagnetic method exploration data optimization method based on big data

Abstract

This application relates to a data optimization method for geophysical electromagnetic exploration based on big data. It employs a three-level filtering mechanism to plan node deployment for preliminary exploration. Based on the preliminary exploration results, a final score is calculated using a scoring formula to determine the geological structure. For complex geological scenarios, multi-band collaborative acquisition is used to identify and suppress interference. Cross-band data fusion is performed, and coarse-scale modeling is conducted on the fused data based on Occam inversion. Fine-scale optimization is then performed using GAN inversion combined with the coarse-scale modeling output and geological constraint parameters. The fine-scale optimization output is fused with gravity data, seismic data, and cryogenic data to output multi-source fused data. This multi-source fused data is then displayed through a 3D geographic information system platform. This application can effectively improve exploration accuracy and efficiency.

Description

Technical Field

[0001] This application relates to the field of electromagnetic exploration technology, and more specifically, to a method for optimizing geophysical electromagnetic exploration data based on big data. Background Technology

[0002] In the field of geological exploration, traditional single-band electromagnetic exploration technology faces bottlenecks such as strong ambiguity and insufficient adaptability to complex scenarios. Early methods often failed to balance different depth resolutions and penetration capabilities due to fixed frequency band acquisition. For example, while high-frequency bands could identify shallow details (such as underground cavities), they were difficult to penetrate thick overburden due to attenuation caused by low-resistivity lithology; while low-frequency bands could detect deep structures (such as concealed ore bodies), their resolution was limited, easily missing secondary anomalies such as thin veins. In terms of interference suppression, traditional filtering algorithms were insufficient in handling mixed interference (such as the superposition of power frequency narrowband and mobile communication broadband interference). Single notch filters could not cope with random impulse noise, resulting in limited signal-to-noise ratio improvement (usually only 5-8dB), and manual parameter adjustment was inefficient and difficult to adapt to real-time dynamic scenarios. At the level of multi-source data fusion, early technologies lacked cross-physics coupling models, and electromagnetic data was often interpreted independently from gravity, seismic, etc., failing to effectively utilize correlation features such as density-resistivity and wave velocity-resistivity. For example, electromagnetic inversion alone is insufficient to distinguish between high-density bedrock and metallic ore bodies, while gravity data can only reflect density differences but cannot clearly indicate water content, resulting in large target location errors (often exceeding 100 meters) and a drilling hit rate of less than 30%. Summary of the Invention

[0003] To overcome at least one deficiency in the prior art, this application provides a method for optimizing geophysical electromagnetic exploration data based on big data.

[0004] Firstly, a method for optimizing geophysical electromagnetic exploration data based on big data is provided, including: Collect geological prior data, filter preliminary frequency bands based on the geological prior data through a three-level filtering mechanism, and conduct preliminary exploration according to the planned node deployment based on the preliminary frequency bands; Based on the preliminary exploration results, the final score is calculated, and the geological structure of the exploration target is determined based on the final score; If the geological structure of the exploration target is a complex geological scene, multi-frequency band collaborative acquisition should be used for precise exploration; Cross-frequency band data fusion is performed on the precisely detected data, and coarse-scale modeling is performed on the fused data based on Occam inversion. GAN inversion is then used to optimize the coarse-scale modeling results for fine-scale optimization. The results of fine-scale optimization are fused with gravity data, seismic data, and geothermal data to obtain multi-source fused data. A 3D geographic information system platform is used to display multi-source fused data.

[0005] In one embodiment, a preliminary frequency band is selected based on prior geological data through a three-level filtering mechanism, including: Determine the depth of the exploration target. If the depth is very shallow, select the high-frequency band; if the depth is shallow to medium, select the medium-frequency band; if the depth is deep, select the low-frequency band. Obtain the lithological distribution layer from the geological prior data, and combine it with borehole core data to obtain the lithological combination characteristics of the exploration target, and adjust the selected frequency band according to the lithological combination characteristics; Obtain the fault density, stratigraphic dip angle, and intrusive body area ratio from the geological prior data, and calculate the structural complexity index; if the structural complexity index is greater than the set threshold, the area where the exploration target is located is a complex structural zone, and then reduce the frequency interval and increase the redundant frequency band.

[0006] In one embodiment, based on preliminary exploration results, a final score is calculated, and the geological structure of the exploration target is determined based on the final score, including: Based on the preliminary exploration results, the resistivity, gradient change, and noise energy ratio are determined. If the resistivity exhibits an abnormal and irregular shape, and the gradient change is greater than the gradient change threshold, and the noise energy ratio is greater than the noise energy ratio threshold, then the resistivity abnormal shape score, noise energy ratio score, and signal-to-noise ratio improvement score are calculated, and the final score is obtained. If the final score is greater than the score threshold, then the geological structure of the exploration target is a complex geological scene.

[0007] In one embodiment, multi-band collaborative acquisition is used for precise exploration, including: Multi-frequency synchronous transmission technology is adopted to transmit multiple frequency signals in the low-frequency band simultaneously; compressed sensing algorithm is used to collect key data points, and the underground structure outline is reconstructed based on the key data points using the L1 norm optimization algorithm; spatial domain downsampling is performed at fixed sampling intervals to identify potential geological anomaly areas. The frequency signal is adjusted to be in the mid-frequency band. For anomalous areas, phased array focusing technology is used to adjust the sensor's emission phase to form a beam focus. The sliding time window correlation analysis method is used to calculate the signal correlation within different time windows to identify stratigraphic interfaces and geological body boundaries. The frequency signal is adjusted to be in the high-frequency band, and a common offset acquisition mode is adopted. The signal is decomposed using wavelet packet decomposition technology to obtain the analysis results. Based on the analysis results and the time-frequency electromagnetic response feature library of typical geological bodies, shallow geological anomalies are identified by matched filtering algorithm to verify the exploration results in the mid-to-low frequency band.

[0008] In one embodiment, during the multi-band collaborative acquisition process, an acquisition period is set, and short-time Fourier transform and convolutional neural network are used to identify interference in the acquired data in real time. If the interference is narrowband interference, an adaptive notch filter is enabled, and the filter center frequency is dynamically adjusted according to the interference frequency. If the interference is broadband interference, the process is switched to an adjacent frequency band and the modulation method is adjusted. If the interference is random pulse interference, a Kalman filter algorithm is used to filter out noise through an iterative process of state prediction and measurement update.

[0009] In one embodiment, the convolutional neural network includes three parallel convolutional branches, a channel attention module, and a depthwise separable convolution; the three convolutional branches use three different sizes of convolutional kernels, namely 1×1, 3×3, and 5×5, to extract narrowband interference features, wideband interference features, and random impulse interference features, respectively.

[0010] In one embodiment, coarse-scale modeling is used, with borehole core data and seismic horizon depth as hard constraints. Constraint terms are added to the inversion objective function, and regularized inversion is adopted to construct the Occam inversion model. Fine-scale optimization employs a conditional generative adversarial network architecture. The generator takes the Occam inversion model and geological constraint parameters as input and outputs a high-resolution resistivity model. The discriminator takes real borehole data and the generator output data as input and uses a loss function to determine the authenticity of local features.

[0011] In one embodiment, a 3D geographic information system platform is used to display multi-source fused data, including: The resistivity distribution of the subsurface medium is displayed in the form of a 3D grid or isosurface, and the differences in electrical properties of the geological body are reflected by color coding. Resistivity slices at different depths are set. Gravity anomaly data are superimposed on the resistivity model as semi-transparent isosurfaces or colored shading maps to show density anomaly areas. The reflection interface interpreted by the seismicity is embedded in the 3D model as a red wireframe or curved surface, and the interface depth and attitude are marked. The subsurface temperature distribution is displayed in the form of isotherms or thermal cloud maps.

[0012] Secondly, a computer-readable storage medium is provided, which stores a computer program that, when executed by a processor, implements the aforementioned method for optimizing geophysical electromagnetic exploration data based on big data.

[0013] Thirdly, a computer program product is provided, including a computer program / instruction, which, when executed by a processor, implements the aforementioned method for optimizing geophysical electromagnetic exploration data based on big data.

[0014] Compared with existing technologies, this application has the following advantages: This application uses a three-level filtering mechanism to plan node deployment for preliminary exploration; based on the preliminary exploration results, it calculates the final score using a scoring formula to determine the geological structure; for complex geological scenarios, it employs multi-band collaborative acquisition to identify and suppress interference; it performs cross-band data fusion, and uses Occam inversion to perform coarse-scale modeling of the fused data, then uses GAN inversion combined with the coarse-scale modeling output and geological constraint parameters for fine-scale optimization; it fuses the fine-scale optimization output with gravity data, seismic data, and low-temperature data to output multi-source fused data; and it displays the multi-source fused data through a 3D geographic information system platform. This application can effectively improve exploration accuracy and efficiency. Attached Figure Description

[0015] This application can be better understood by referring to the description given below in conjunction with the accompanying drawings, which, together with the detailed description below, are incorporated in and form part of this specification. In the drawings: Figure 1 A flowchart of a method for optimizing geophysical electromagnetic exploration data based on big data is shown. Figure 2 A schematic diagram of a convolutional neural network is shown. Detailed Implementation

[0016] Exemplary embodiments of the present application will be described below with reference to the accompanying drawings. For clarity and brevity, not all features of the actual embodiments are described in the specification. However, it should be understood that many embodiment-specific decisions can be made in the development of any such actual embodiment to achieve the developer’s specific objectives, and these decisions may vary as the embodiments differ.

[0017] It should also be noted that, in order to avoid obscuring this application with unnecessary details, only the device structure closely related to the solution of this application is shown in the accompanying drawings, while other details that are not closely related to this application are omitted.

[0018] It should be understood that this application is not limited to the described embodiments by virtue of the following description with reference to the accompanying drawings. In this document, embodiments may be combined with each other, features may be substituted or borrowed between different embodiments, and one or more features may be omitted in one embodiment, where feasible.

[0019] This application provides a method for optimizing geophysical electromagnetic exploration data based on big data. Figure 1 A flowchart illustrating a data optimization method for geophysical electromagnetic exploration based on big data is shown. (See attached diagram) Figure 1 The method mainly includes the following steps: Step S1: Collect geological prior data, filter preliminary frequency bands based on the geological prior data through a three-level filtering mechanism, and conduct preliminary exploration according to the planned node deployment based on the preliminary frequency bands.

[0020] Step S2: Calculate the final score based on the preliminary exploration results, and determine the geological structure of the exploration target based on the final score.

[0021] Step S3: If the geological structure of the exploration target is a complex geological scene, use multi-band collaborative acquisition for precise exploration.

[0022] Step S4: Perform cross-frequency data fusion on the precisely detected data, and perform coarse-scale modeling on the fused data based on Occam inversion. Then, use GAN inversion to perform fine-scale optimization on the coarse-scale modeling results.

[0023] Step S5: The results of fine-scale optimization are fused with gravity data, seismic data, and geothermal data to obtain multi-source fused data.

[0024] Step S6: Use a 3D geographic information system platform to display the multi-source fused data.

[0025] In one embodiment, step S1, screening preliminary frequency bands based on prior geological data using a three-level filtering mechanism, includes: First, determine the depth of the exploration target. If the depth is very shallow (0-50 meters), select the high-frequency band to capture subtle geological changes such as shallow water pollution and underground cavities by utilizing the high resolution of high-frequency signals. If the depth is shallow to medium (50-300 meters), select the mid-frequency band to effectively identify geological bodies of medium depth while ensuring a certain penetration depth. If the depth is deep (above 300 meters), select the low-frequency band to penetrate thick strata and obtain information on deep geological structures by taking advantage of the large skin depth of low-frequency signals. Here, the frequency bands are divided into low-frequency band (0.1Hz-10kHz), mid-frequency band (10kHz-100kHz), and high-frequency band (100kHz-10MHz).

[0026] Then, the lithological distribution layer in the geological prior data is obtained, and combined with the borehole core data, the lithological combination characteristics of the exploration target are obtained, and the selected frequency band is adjusted according to the lithological combination characteristics. Here, lithological distribution layers (such as granite, sandstone, and clay) from a prior geological database are accessed, and combined with borehole core data to obtain the lithological combination characteristics of the target area (e.g., a region with clay overlay and sandstone interlayers). In areas with high resistivity lithology (such as granite and sandstone), the frequency band is increased at the same depth to enhance signal resolution and more clearly reveal the boundaries of geological bodies; in areas with low resistivity lithology (such as clay and shale), the frequency band is decreased to increase signal penetration and ensure that the target geological body can be detected. The system has a built-in electromagnetic response database for common lithologies, and automatically optimizes frequency band parameter settings by comparing measured data with database characteristics.

[0027] Then, the fault density, stratigraphic dip angle and intrusive body area ratio in the geological prior data are obtained, and the structural complexity index is calculated. If the structural complexity index is greater than the set threshold, the area where the exploration target is located is a complex structural zone, and the frequency interval is reduced and redundant frequency bands are added.

[0028] Structural Complexity Index (SCI) = (Fault Density / Maximum Fault Density in the Region) × 0.5 + (Structural Dip Angle / 90°) × 0.2 + (Intrusive Body Area Percentage / 100%) × 0.1. Fault density, structural dip angle, and intrusive body area percentage are all normalized to the 0-1 range to ensure dimensional consistency. A threshold is set (the threshold is user-defined, typically 0.6). When the SCI... When setting thresholds, areas are identified as complex structural zones (such as densely faulted zones and fold cores). Simple structural zones use a fixed 10kHz interval (e.g., 20kHz, 30kHz, 40kHz). For complex geological structural areas, such as fault zones and fold-developed areas, the system narrows the frequency interval to 5kHz (e.g., 20kHz, 25kHz, 30kHz). Redundant frequency bands are added for acquisition. During fault zone exploration, in addition to the primary selected frequency band, adjacent frequency bands within ±2kHz of the primary band are acquired to suppress false anomalies caused by tectonic interference at single frequencies, capture the complex electromagnetic response characteristics of fault zones, and improve the identification ability and accuracy of geological anomalies. The exploration results obtained through this three-level filtering mechanism constitute the preliminary exploration results.

[0029] In one embodiment, step S2, calculating the final score based on the preliminary exploration results, and determining the geological structure of the exploration target based on the final score, includes: Based on the preliminary exploration results, resistivity, gradient change, and noise energy proportion are determined. If the resistivity exhibits an abnormally irregular shape and the gradient change exceeds the gradient change threshold (e.g., 5), If the noise energy percentage is greater than the noise energy percentage threshold (e.g., 40%), then the resistivity anomaly morphology score, noise energy percentage score, and signal-to-noise ratio improvement score are calculated, and the final score is obtained. If the final score is greater than the score threshold, the geological structure of the exploration target is a complex geological scene. Otherwise, it is a simple scene.

[0030] Here, the final score = resistivity anomaly morphology score × corresponding weight + noise energy percentage score × corresponding weight + signal-to-noise ratio improvement score × corresponding weight.

[0031] The corresponding weights and thresholds are set according to the actual exploration area.

[0032] If the final score If the score threshold (usually set to 2 points) is used, it is initially judged to be a complex geological scene.

[0033] When performing calculations, the resistivity anomaly pattern with regular shape (layered, blocky) is scored as 0 points, the resistivity anomaly pattern with slight irregularity (local jaggedness) is scored as 1 point, the resistivity anomaly pattern with moderate irregularity (multiple fractures) is scored as 2 points, and the resistivity anomaly pattern with extremely irregularity (fragmented distribution) is scored as 3 points.

[0034] When noise percentage 30%, noise energy percentage score is 0; 30% Noise ratio 40%, noise energy percentage score is 1 point; 34% Noise ratio 45%, noise energy percentage score is 2 points; noise percentage 45%, with a noise energy percentage score of 3.

[0035] The signal-to-noise ratio is improved after filtering. 15dB, signal-to-noise ratio improvement score is 0; 10dB Signal-to-noise ratio improvement 15dB, signal-to-noise ratio improvement score is 1 point; signal-to-noise ratio improvement 10dB, the signal-to-noise ratio improvement score is 2 points.

[0036] In one embodiment, step S3, employing multi-band collaborative acquisition for precise exploration, includes: First, multi-frequency synchronous transmission technology is used to simultaneously transmit multiple frequency signals in the low-frequency band; then, compressed sensing algorithm is used to collect key data points, and the outline of the underground structure is reconstructed based on the key data points using the L1 norm optimization algorithm. Spatial domain downsampling is performed at fixed sampling intervals to delineate potential geological anomaly areas. Here, a low-frequency coarse scan (0.1-1kHz) is first performed using multi-frequency synchronous transmission technology, simultaneously transmitting multiple frequency signals such as 0.1kHz, 0.3kHz, and 0.5kHz. Combined with compressed sensing algorithms, only key data points are collected (the sampling rate is reduced to 1 / 4 of that of traditional methods). The underground structure outline is then rapidly reconstructed using an L1 norm optimization algorithm (this step is a conventional technique well-known to those skilled in the art and will not be elaborated upon). In large-area deep exploration, spatial downsampling is performed at sampling intervals of 50-100 meters to quickly delineate potential geological anomaly areas, providing macroscopic guidance for subsequent exploration work.

[0037] Then, the frequency signal is adjusted to be in the mid-frequency band. For the abnormal area, phased array focusing technology is used to adjust the sensor's emission phase to form beam focusing. The sliding window correlation analysis method is used to calculate the signal correlation in different time windows to identify the stratigraphic interface and geological body boundary. Combined with the Bayesian inversion algorithm, the inversion results of the low-frequency band are used as prior information to constrain the mid-frequency band inversion process to obtain the mid-frequency band inversion results. Here, a mid-frequency band fine exploration (10-100kHz) is conducted, and suspected abnormal areas found in the low-frequency band are detected in a focused manner. Phased array focusing technology is used (this step is a conventional technique well known to those skilled in the art and will not be described in detail). By adjusting the transmission phase of multiple sensors, beam focusing (main lobe width <15°) is formed to enhance the target signal strength.

[0038] A sliding window correlation analysis method was employed to calculate the correlation of signals within different time windows (20ms, 50ms, and 100ms), accurately identifying stratigraphic interfaces and geological body boundaries. Combined with a Bayesian inversion algorithm, the inversion results in the low-frequency band were used as prior information to constrain the mid-frequency inversion process, ensuring that the boundary error of the resistivity model was less than 8%, thus achieving accurate characterization of the geological body morphology.

[0039] Then, the frequency signal is adjusted to be in the high-frequency band, and the common offset acquisition mode is adopted. The signal is decomposed using wavelet packet decomposition technology, and the attenuation characteristics and phase changes of each frequency band are analyzed to obtain the analysis results. Based on the analysis results and the time-frequency electromagnetic response feature library of typical geological bodies, shallow geological anomalies are identified by matched filtering algorithm to verify the exploration results in the mid-to-low frequency band.

[0040] Here, high-frequency band verification (100kHz-1MHz) is performed using a common-offset acquisition mode, maintaining a fixed transmit / receive distance (e.g., 10 meters) to eliminate the influence of geometric diffusion on the signal. Wavelet packet decomposition technology is used to decompose the signal into different frequency bands (e.g., 500-600kHz, 600-700kHz, etc.), and the attenuation characteristics and phase changes of each band are analyzed. A time-frequency electromagnetic response feature library of typical geological bodies is constructed, such as the high-frequency rapidly decaying oscillation signal of karst caves and the low-frequency phase anomaly characteristics of metallic ore bodies. A matched filtering algorithm is used to identify shallow geological anomalies, verify the exploration results in the mid-to-low frequency bands, and obtain detailed geological information such as the shallow oxidation zone of the ore body and shallow groundwater.

[0041] For example, exploration is being conducted in a mountainous metal ore area located in the southwest mountainous region. The geological conditions are complex, with densely developed fault zones (fault density 2.5 faults / km²), interbedded granite and shale, and dramatically undulating surface topography (slope > 40°). The exploration target is a concealed metal ore body buried at a depth of less than 500 meters, requiring electromagnetic methods to detect its location and morphology. First, depth-based filtering is performed, with the target depth set at 500 meters, initially selecting a low-frequency band (0.1Hz-10kHz). Next, lithological adaptation adjustments are made, with the surface consisting of low-resistivity shale (resistivity 20-100 Ω·cm). The deeper part is high-resistivity granite (resistivity > 800 Ω·cm). The ore body is located in the contact zone (presumably low-resistivity sulfides). Therefore, in the shale-covered area (shallow), the frequency band is lowered to the mid-low frequency band (5-15kHz) to enhance penetration and avoid signal attenuation by the low-resistivity layer; in the granite area (deep), the frequency band is appropriately increased to the mid-frequency band (20-50kHz) to improve resolution and identify the boundary between the ore body and the surrounding rock. Further structural complexity corrections are made: fault density = 2.5 faults / km² (the maximum regional fault density is 5 faults / km², normalized to 0.5), and stratigraphic dip angle = 50° (normalized to 50 / 90). 0.56), the intrusive body area ratio = 30% (normalized to 0.3), SCI = 0.5×0.5 + 0.56×0.2 + 0.3×0.1 = 0.392 < threshold 0.6, thus it is judged as a simple tectonic zone; if the maximum fracture density in the region is 3 fractures / km², the normalized fracture density is 2.5 / 3 0.83, SCI=0.83×0.5+0.56×0.2+0.3×0.1=0.557 A value of 0.6, close to the threshold, indicates a transitional structural zone, with the frequency interval narrowing to 7 kHz. Preliminary exploration results show an irregular low-resistivity anomaly (resistivity < 200 Ω·cm) at depth in the low-frequency band (3 kHz). The shape is serrated, and the gradient change is >8. / m, consistent with the fault zone strike. Time-frequency domain data: Noise energy accounts for 45% (mainly topographic scattering interference). After notch filtering and frequency band switching (from 3kHz to 5kHz), the signal-to-noise ratio improvement is only 8dB < 10dB. The judgment result meets the conditions of "irregular resistivity anomaly morphology + insufficient noise suppression effect", and it is judged to be a complex geological scene. Multi-frequency collaborative acquisition is initiated. Multi-frequency collaborative acquisition is carried out. 1kHz, 3kHz, and 5kHz signals are transmitted simultaneously. Combined with compressed sensing algorithm, the sampling rate is reduced to 1 / 4 of the traditional method (1 data point is collected every 50 meters). The model is reconstructed by L1 norm optimization algorithm, and an elliptical low resistivity anomaly area with a diameter of about 800 meters is delineated, which is speculated to be a ore body cluster area. The transmission phases of eight sensors were adjusted to form a focused beam with a main lobe width of 12°, enhancing the signal strength in the anomalous area (improving the signal-to-noise ratio by 15dB). Correlation of signals in 20ms and 50ms time windows was calculated, identifying three distinct stratigraphic interfaces. The interface at a depth of 550 meters coincided with the boundary of the low-resistivity anomaly. Using a low-frequency model as a priori, mid-frequency inversion was constrained. The resistivity model boundary error was 6.5% < 8%, confirming the ore body's strike as NW and its length as approximately 500 meters. Common-offset acquisition (transmitter-receiver distance 10 meters) eliminated geometric diffusion errors caused by topography. Wavelet packet decomposition was applied to the 500-600kHz and 600-700kHz frequency bands, revealing a high-frequency, rapidly attenuating signal in the shallow (200m) region, matching the characteristics of the sulfide oxidation zone (resistivity > 500kΩ). A three-dimensional electrical model of "deep low-resistivity ore body group - central fault zone - shallow oxidation zone" was constructed by cross-frequency band data fusion, locating the ore body center at a depth of 530 meters and a thickness of approximately 30 meters, with an 85% agreement rate with subsequent drilling results. The exploration cycle was shortened by 40% compared to traditional methods, and the borehole hit rate increased from 30% to 75%, verifying the effectiveness of the multi-frequency band intelligent decision-making system in complex scenarios.

[0042] Furthermore, during the multi-band collaborative acquisition process, an acquisition period is set, and short-time Fourier transform and convolutional neural network are used to identify interference in the acquired data in real time. If the interference is narrowband interference, an adaptive notch filter is enabled, and the filter center frequency is dynamically adjusted according to the interference frequency. If the interference is broadband interference, the process is switched to an adjacent frequency band and the modulation method is adjusted. If the interference is random pulse interference, a Kalman filter algorithm is used to filter out noise through an iterative process of state prediction and measurement update.

[0043] Here, during data acquisition, the system performs comprehensive three-dimensional spectrum analysis (time domain, frequency domain, and time-frequency analysis) on the acquired data at 10-minute intervals, promptly identifying interference and dynamically adjusting the acquisition parameters. A combination of Short-Time Fourier Transform (STFT) and Convolutional Neural Network (CNN) is used for real-time interference identification of the acquired data. Addressing the insufficient handling capability of traditional filtering algorithms for mixed interference, this embodiment employs an improved Convolutional Neural Network (CNN) architecture for real-time interference identification. Figure 2 A schematic diagram of a convolutional neural network is shown, and the specific optimizations are as follows: Three parallel convolutional branches (1×1, 3×3, and 5×5 kernels) are added to the traditional CNN to extract sharp spectral features of narrowband interference (such as 50Hz power frequency), continuous spectral features of broadband interference (such as mobile communication signals), and instantaneous pulse features of random pulse interference (such as lightning noise), respectively, solving the problem of low accuracy of single convolutional kernels in identifying mixed interference. A channel attention module is added before the fully connected layer to enhance the focusing ability on strong interference by learning the weight coefficients of interference types (e.g., the weight of the power frequency interference channel is increased to 0.8) and reducing the interference of irrelevant noise. Depthwise separable convolution is used to replace traditional convolution, reducing the model parameters from 2.3M to 0.8M and doubling the inference speed, meeting the low latency requirements of real-time field data acquisition (end-to-end latency ≤200ms).

[0044] The CNN classifies the acquired data into different interference types every 10 minutes (with 92% accuracy), outputting the probability distribution of narrowband / wideband / random impulse interference and triggering corresponding filtering strategies. For narrowband interference, such as 50Hz power frequency interference, an adaptive notch filter is automatically activated, dynamically adjusting the filter center frequency (Q value adjustable up to 30) according to the interference frequency to effectively suppress interference. For wideband interference, such as mobile communication signal interference, the system automatically switches to adjacent frequency bands and adjusts the modulation method, such as switching from amplitude modulation (AM) to frequency modulation (FM), to reduce the impact of interference. In addition, for random impulse interference, a Kalman filter algorithm is used, effectively filtering out noise through an iterative process of state prediction and measurement updates, improving the signal-to-noise ratio by 12-15dB. The system optimizes the acquisition parameters in real time based on signal quality. In the high-frequency band (>500kHz), the sampling rate is automatically increased to above 1MHz to ensure no signal distortion; in the low-frequency band (<10kHz), the sampling rate is reduced to 10kHz to reduce data storage pressure. For suspected anomaly areas, the acquisition time is automatically extended from the standard 30 seconds to 2 minutes, and the number of signal superpositions is increased from 100 to 500, improving the signal-to-noise ratio and data reliability. Simultaneously, an automatic gain control (AGC) algorithm adjusts the signal gain in real time, increasing the gain by 3dB per level when the signal amplitude is <0.5V, and monitoring total harmonic distortion (THD). When THD >1%, the gain is reduced to prevent signal saturation distortion. After adopting an improved convolutional neural network (CNN) architecture, the interference identification accuracy is improved to 92%, a 17% improvement compared to the traditional CNN architecture. The optimized signal-to-noise ratio is improved by 12-15dB, an average improvement of 7dB compared to the traditional approach. The optimized effective data integrity is improved to 98%, a 23% improvement compared to the traditional approach. The optimized thin vein identification error is ≤10 meters, a 67% improvement compared to the traditional approach.

[0045] The improved convolutional neural network employs parallel convolutional branches with kernels of three different sizes: 1×1, 3×3, and 5×5. These branches target the feature scales of narrowband, wideband, and random impulse interference, respectively, to more comprehensively extract interference features from electromagnetic exploration data. A channel attention module is introduced to assign different weights to channels with different interference types, highlighting key features that have a greater impact on subsequent interference suppression and reducing irrelevant feature interference. Depthwise separable convolutions are used instead of traditional convolutions, significantly reducing the number of model parameters and improving computational efficiency while maintaining feature extraction effectiveness, making it more suitable for the real-time requirements of this scheme. Narrowband interference in electromagnetic exploration data typically has a narrow frequency range and relatively regular feature patterns; a 1×1 convolutional kernel can accurately capture its local features. Wideband interference covers a wide frequency range; a 3×3 convolutional kernel can effectively extract its multi-frequency combination features. Random impulse interference often has irregular burst characteristics; a 5×5 convolutional kernel can capture a wider range of feature correlations. Through parallel multi-scale convolution, the feature scales of different types of interference can be covered simultaneously, avoiding the limitations of a single convolutional kernel in extracting complex interference features. Different types of interference have varying importance in the feature channels of electromagnetic exploration data. For example, strong narrowband interference channels are more instructive for subsequent narrowband interference suppression. The channel attention module learns the weights of each channel, enhancing the features of important channels and appropriately suppressing the features of secondary channels, making the model more focused on processing key interference features. The spatial convolution and channel convolution of traditional convolution are separated. First, spatial convolution is performed independently on each channel using depthwise convolution to extract spatial features; then, pointwise convolution (1×1 convolution) is used to fuse the features of different channels. Compared with traditional convolution, this approach significantly reduces the number of parameters, significantly improves computation speed, and maintains good feature extraction capabilities, effectively balancing accuracy and efficiency when processing large-scale geophysical electromagnetic exploration data. The improved neural network interference identification accuracy was 82.3%, while the improved interference identification accuracy increased to 94.7%, an improvement of 12.4 percentage points. The number of model parameters decreased significantly from 1.285 million to 321,000. The processing time for a single batch of data decreased from 450ms to 120ms. After interference suppression, the signal-to-noise ratio of the data increased from 15.2dB to 28.9dB.

[0046] For example, during the multi-band collaborative acquisition phase of a metal mine in a mountainous area (low-frequency coarse scanning completed, transitioning to mid-frequency fine exploration), sensors were deployed in a complex terrain area (slope > 45°), surrounded by a few high-voltage power lines (introducing 50Hz power frequency interference) and occasional mobile communication signal interference (frequency band 800-900MHz). Monitoring revealed periodic spikes in the signal waveform at certain times (period 20ms, corresponding to 50Hz power frequency interference), along with irregular pulse noise. The signal amplitude fluctuated significantly (0.3-0.8V) in suspected abnormal areas (ore body contact zone), requiring dynamic gain adjustment. FFT transformation revealed prominent harmonic energy at 50Hz and 100Hz (amplitude > -20dBm), indicating narrowband interference; continuous noise (broadband interference) existed in the 800-900MHz band, covering the adjacent area of ​​the mid-frequency band (20-50kHz). Time-frequency analysis (STFT) shows that the 50Hz interference is continuous, while mobile communication interference is intermittent (busy signal periods are 9:00-11:00 and 14:00-16:00 daily). Inputting the STFT into a CNN network (the pre-trained model has identified 10 common interference types) yields a combination of "50Hz power frequency interference + mobile communication broadband interference + random pulse interference" with a confidence level of 92%. Automatic adjustment of the center frequency to 50Hz, with a Q value of 25 (balancing suppression and signal fidelity), results in an attenuation of -30dB. The residual interference amplitude is <-50dBm, below the signal noise floor. After filtering, the time-domain waveform spikes disappear, the 50Hz peak in the frequency domain is essentially eliminated, and the signal-to-noise ratio improves by approximately 18dB. The original mid-frequency band was 30kHz. Upon detecting that the adjacent 28kHz band experienced less interference, the system automatically switched to 28kHz and adjusted the modulation method from AM to FM (improving anti-interference capability by 10dB). The sampling rate was increased from 50kHz to 60kHz (for high-frequency band adaptation), ensuring undistorted FM signal sampling. A state-space model was established to predict signal trends and update measured values. After five iterations, the impulse noise amplitude decreased from 0.5V to below 0.1V, and the signal-to-noise ratio improved by 15dB. In the ore body contact zone (a key area for mid-frequency band fine-tuning), the single-channel acquisition time was extended from 30 seconds to 2 minutes, and the number of signal superpositions was increased to 500 (from 100). After superposition, the noise standard deviation decreased from 0.2V to 0.05V, the target signal (ore body response) amplitude increased from 0.6V to 1.2V, and the signal-to-noise ratio improved from 3:1 to 24:1. When the signal amplitude is <0.5V (e.g., signals from deep, low-resistivity ore bodies), the gain is increased in three stages (+9dB) to stabilize the amplitude between 0.8-1.0V. When THD >1% (e.g., strong signal saturation), the gain is reduced by two stages (-6dB) to avoid harmonic distortion. Ultimately, the 50Hz power frequency interference suppression rate is >95%, the duration of mobile communication interference is reduced by 70%, random impulse noise residue is <5%, the data integrity rate of mid-frequency fine exploration is increased from 75% to 98%, and the proportion of effective data is significantly improved.Dynamic adjustment of sampling rate and gain improves the resolution of shallow high-frequency signals to 5 meters and the penetration depth of deep low-frequency signals to 800 meters. Long-term superposition acquisition of suspected anomaly areas provides high signal-to-noise ratio input for subsequent Bayesian inversion (the boundary error is reduced from 12% to 5.8%), ensuring the accuracy of ore body morphology characterization.

[0047] Furthermore, multi-band collaborative acquisition is used for precise exploration, with fixed and mobile nodes deployed collaboratively during acquisition.

[0048] Fixed nodes are primarily suitable for areas with complex geological conditions requiring long-term monitoring, such as known mineralized areas, active fault zones, and the geological stability monitoring around major infrastructure projects (nuclear power plants, dams). In complex geological scenarios, a dense grid is constructed with a spacing of 50-100 meters, and the sensors are securely fixed using concrete bases. They are equipped with high-performance electromagnetic shielding, lightning protection devices, and an independent solar power system (equipped with lithium battery storage, supporting 72 hours of continuous operation). The sensors feature a modular design, facilitating rapid on-site replacement of core components such as electrodes, magnetic rods, and data acquisition units. In relatively uniform plain areas, the node spacing is appropriately increased to 200-300 meters, and flexible layouts are implemented based on terrain features such as roads and rivers to ensure data coverage across the entire exploration area.

[0049] Mobile nodes include both UAV nodes and vehicle-mounted nodes. Multi-rotor or fixed-wing UAVs are used as the platform, equipped with lightweight electromagnetic sensors, suitable for rapid surveys of large areas and exploration of remote or dangerous areas (such as mountains, swamps, and earthquake-stricken areas). The UAVs are equipped with a dual-frequency GNSS positioning system (L1+L2 bands) and real-time dynamic carrier phase differential technology (RTK), maintaining centimeter-level positioning accuracy even in signal-blocked environments. Flight altitude is controlled between 50-200 meters, with flight path spacing of 200-500 meters, and a single flight can cover an area of ​​10-20 km². The UAVs communicate with the ground control center in real time via 4G / 5G communication links or data radios, transmitting collected data and flight status information, and supporting remote mission planning and parameter adjustments for efficient data acquisition. Electromagnetic acquisition equipment is integrated into all-terrain vehicles, suitable for rapid scanning of relatively flat areas and geological exploration along urban roads. The vehicle is equipped with a high-precision navigation system, a shock-absorbing platform, and a retractable electromagnetic sensor bracket, supporting both fixed-point measurement (data collection at 100-300 meter intervals) and continuous dynamic measurement modes. In continuous dynamic measurement mode, real-time positioning is achieved through an inertial navigation system and odometer, combined with GPS data for error correction, ensuring the spatial accuracy of the collected data. The onboard system can quickly switch between different types of electromagnetic measurement devices, such as controlled-source audio-frequency magnetotellurics (CSAMT) equipment and transient electromagnetic methods (TEM) equipment, to meet diverse exploration tasks.

[0050] In areas with strong electromagnetic interference, such as factories, substations, and communication base stations, multi-layer composite electromagnetic shielding technology is employed. For example, a shielding structure combining permalloy and copper mesh can achieve a shielding effectiveness of over 60dB, effectively reducing the intensity of external interference. Combined with frequency avoidance strategies, the system automatically scans the environmental electromagnetic spectrum and avoids the 50Hz / 60Hz power frequency and its harmonic bands (such as 49-51Hz and 99-101Hz) in real time. For interference bands that cannot be avoided, Generative Adversarial Networks (GANs) are used for data repair. The generator is trained with a large amount of interference-free data to learn the characteristic distribution of normal electromagnetic signals. Then, the interfered data is input into the generator to generate interference-free signal samples, replacing the original contaminated data. Actual tests show that the signal-to-noise ratio of the data repaired by GAN can be improved by more than 15dB, effectively restoring the authenticity and usability of the data.

[0051] In areas affected by lightning interference, a high-precision lightning monitoring and early warning system is deployed to monitor the number of cloud-to-ground flashes, lightning current intensity, and lightning location in real time. When lightning activity is detected (thunderstorm proximity index > 5), the system immediately and automatically suspends data acquisition, switches the sensors to protection mode, shuts down the transmitting circuit, and disconnects the antenna connection to prevent lightning damage to the equipment. After the lightning activity ends, data is automatically re-acquired based on timestamp information, and the re-acquired data is repaired using a pre-established lightning interference feature library (recording interference waveforms under different thunderstorm intensities). Wavelet transform is used to remove lightning pulse noise, and machine learning algorithms are combined to identify and recover valid signals that have been submerged by interference, ensuring the integrity and accuracy of the data.

[0052] In areas of geomagnetic field variation, establish a dynamic geomagnetic field monitoring network to collect real-time data on the three components of the geomagnetic field (northward). East Vertical The data was sampled at a frequency of 1 Hz. This was combined with solar wind monitoring data (such as solar wind speed and interplanetary magnetic field strength) and geomagnetic activity index (…). An index was used to construct a geomagnetic field variation prediction model. A Kalman filter algorithm was employed to perform real-time correction on the collected electromagnetic data. The specific process is as follows: An ARIMA (Autoregressive Integral Moving Average) model was trained using historical geomagnetic field data to capture the periodic variation patterns of the geomagnetic field (such as diurnal and seasonal variations). External input terms were established by combining solar activity parameters (such as sunspot number) to improve the model's predictive ability for anomalous events such as geomagnetic storms. The real-time collected geomagnetic field data was compared with the output values ​​of the prediction model to calculate the magnetic field disturbance deviation Δ. Δ Δ Based on the principle of electromagnetic induction, the magnetic field disturbance is converted into an equivalent electric field disturbance term, which is then expressed by the formula... (in Angular frequency, The original electric field data is compensated for by the permeability (magnetic permeability), eliminating measurement errors caused by changes in the geomagnetic field. (This refers to the geomagnetic activity index.) When the value is greater than 3, the geomagnetic field correction frequency is automatically increased to 5 seconds / time, and the weight of the process noise covariance matrix of the Kalman filter is increased to enhance the tracking ability of rapidly changing interference.

[0053] In mountainous areas and other regions with significant topographic relief, UAVs equipped with LiDAR (Light Detection and Ranging) acquire 0.5-meter resolution topographic point cloud data. The signal propagation path is corrected through the following steps: Based on the FDTD (Finite Differential Time Domain) algorithm, a terrain-electromagnetic signal propagation model is established to simulate the effects of different terrain slopes (e.g., steep slopes >30°) on electromagnetic signals, including blocking, reflection, and scattering, predicting the signal attenuation coefficient of each acquisition node. Based on the simulation results, the altitude of fixed nodes is automatically adjusted (using electrically operated lifting supports) or temporary mobile nodes are deployed to ensure signal penetration between adjacent nodes is >90%. For example, in valley areas, nodes are deployed near ridgelines to reduce signal obstruction by utilizing higher terrain. The acquired data is then weighted by a terrain slope factor, using the following formula: ,in This is the measured electric field value. For terrain slope, This is an empirical correction factor (valued between 0.1 and 0.3, adjusted according to lithology). This is the corrected terrain slope factor.

[0054] In aquatic environments, a dual-frequency electromagnetic sensor (supporting both above-water and underwater modes) is employed: In shallow water areas (depth < 10 meters), non-contact inductive coupling technology is used, emitting electromagnetic fields through a coil on the water surface to avoid corrosion and signal attenuation caused by direct sensor contact with the water. In deep water areas (depth ≥ 10 meters), the underwater sensor is lowered to a height of 5-10 meters above the bottom via a cable. Combined with real-time monitoring of water depth and sensor attitude using sonar, the transmission power (maximum 500W) and sampling rate (100Hz) are automatically adjusted to compensate for the absorption of electromagnetic signals by the water (seawater conductivity is approximately 4S / m, and signal attenuation increases exponentially with increasing frequency).

[0055] In one embodiment, step S4, performing cross-frequency band data fusion on the precisely detected data, includes: Based on the BeiDou high-precision clock and inertial navigation system, data collected from different frequency bands are synchronized at the sub-millisecond level (time error < 0.5ms) and spatially aligned at the centimeter level. By constructing a unified spatiotemporal coordinate system, multi-frequency band data are mapped to the same three-dimensional spatial grid, ensuring data consistency in both time and space dimensions, laying the foundation for subsequent fusion analysis.

[0056] Through experimental measurements and theoretical calculations, transfer function models between different frequency bands were established to obtain the response differences between each band. Frequency domain correction was performed on the data for each frequency band, for example, for the low-frequency band. and high-frequency data Through calculation Obtain the transfer function Spectral correction is performed on low-frequency data to eliminate signal distortion caused by frequency band differences, enabling better fusion of data from different frequency bands.

[0057] Kriging interpolation is used to calculate the fusion weights of data from different frequency bands. The weight coefficients are dynamically adjusted based on data depth, signal-to-noise ratio, and geological reliability; for example, the weight of low-frequency data is calculated as depth / total depth. Multi-scale decomposition techniques, such as wavelet multi-resolution analysis, are employed to fuse information from various frequency bands at different scales. First, the macroscopic structural features of low-frequency data are preserved at a large scale, then detailed information from high-frequency data is incorporated at a small scale. Finally, a three-dimensional resistivity model that balances resolution and penetration depth is reconstructed, providing high-precision data support for geological interpretation.

[0058] In one embodiment, in step S4, coarse-scale modeling is performed, using borehole core data and seismic horizon depth as hard constraints. Constraint terms are added to the inversion objective function, and regularized inversion is adopted to construct the Occam inversion model.

[0059] In coarse-scale modeling (Occam inversion), borehole core data (such as resistivity logging curves) and seismic stratigraphic depths (such as reflection interface depths) are imported as hard constraints, and constraint terms are added to the inversion objective function to form geological prior constraints. : ; in: These are the prior model parameters. For the corresponding data standard deviation, Indicates the parameters of the inversion model. The number of prior data points.

[0060] Regularized inversion is employed, with smoothing factor weights ranging from 0.3 to 0.7 (increasing with depth) to ensure the stability of the deep model, thus constructing the Occam inversion model. The formula for the regularized inversion smoothing factor is as follows: , To vary with depth Weights of the changing smoothing factor To determine the maximum depth of the exploration target, the preconditioned conjugate gradient method (PCG) is used to solve the linearized inversion equation, optimizing computational efficiency by reducing the number of iterations to less than 50 and decreasing computation time by 60% compared to the traditional least squares method. During coarse-scale modeling, low-frequency coarse-scan electromagnetic data (electric field) is input. / ,magnetic field / / The amplitude, phase, and frequency response of the borehole core resistivity logging curve (e.g., the resistivity of a borehole at a depth of 500 meters is 150 Ω·cm). The data includes seismic interpretation depths (e.g., 300 meters from the top of the bedrock, 450-600 meters for fault zone F-01), preprocessed data (spatiotemporal alignment error <1ms / 5cm, signal-to-noise ratio improved by 15dB), and finally outputs a three-dimensional coarse-scale resistivity model (resolution 50-100 meters) to display deep low-resistivity anomalies (e.g., at a depth of 500-700 meters, resistivity <200 Nm). (The ellipsoidal region).

[0061] Fine-scale optimization employs a conditional generative adversarial network architecture. The generator takes the Occam inversion model and geological constraint parameters as input and outputs a high-resolution resistivity model. The discriminator takes real borehole data and the generator output data as input and uses a loss function to determine the authenticity of local features.

[0062] Generator: A lithological probability distribution vector (e.g., spatial distribution probability of granite and shale) and hard constraints on borehole core resistivity are additionally embedded in the generator's input layer. This allows the generator to prioritize fitting known geological features during training, avoiding the generation of meaningless false anomalies. Input is a coarse-scale Occam model and geological constraint parameters (e.g., lithological probability distribution), outputting a high-resolution resistivity model. The generator employs a U-Net structure (encoder-decoder symmetric network containing 8 downsampling / upsampling layers).

[0063] Discriminator: A PatchGAN discriminator and a global discriminator are jointly supervised. PatchGAN focuses on the detail authenticity of a 10×10 local area (such as the gradient of the ore body boundary), while the global discriminator verifies the rationality of the overall geological structure (such as the consistency of the fault zone strike), reducing the thin ore vein identification error from 30 meters to within 10 meters. Real borehole data and generator output data are input, and the authenticity of local features is determined by PatchGAN (loss function). The loss function is:

[0064] in: For real data, For noise vectors, Due to geological constraints, The value of the loss function. For the expected operation, The output of the discriminator, This is the output of the generator.

[0065] Training strategy: Use the Adam optimizer (learning rate 1e-4, =0.5), the generator is updated once every 5 training rounds, and after 2000 iterations, the RMSE between the model's predicted resistivity and the measured value is <8%. The final output is a high-resolution resistivity model (10-20 meters resolution), which identifies secondary anomalies (such as a low-resistivity vein with a thickness of 10 meters and a burial depth of 650 meters, and a resistivity of 80). ), and detailed reports on local structures (such as resistivity gradient within the fracture zone > 5). () / m, ore body boundary error <8%).

[0066] In one embodiment, step S5, fusing the results of fine-scale optimization with gravity data, seismic data, and geothermal data, includes: Resample all data to the same grid (3D) or the same profile location (2D) to ensure spatial alignment.

[0067] For gravity data, it may be necessary to convert observed gravity anomalies into density models (through inversion or simple empirical conversion) in order to compare or fuse them with resistivity models.

[0068] For seismic data, if a reflection interface is available, extract the layer depth; if a wave velocity model is available, use it directly.

[0069] For geothermal data, it may be necessary to interpolate from point data to a grid and take into account the background field of geothermal gradient.

[0070] Specific fusion methods can include: machine learning-based fusion. This involves training a multi-input, single-output model using machine learning methods (such as neural networks), taking multi-source data as input, and predicting a comprehensive attribute model. This requires a large amount of training data (known real-world subsurface models). Another approach is to use unsupervised learning (such as clustering) to classify multi-source data into different lithological or fluid categories.

[0071] In other embodiments, structural constraint-based fusion can also be employed. Subsurface structures reflected by different geophysical data often exhibit consistency. Fusion can be performed by extracting structural information (such as gradients and edges) from various data sources. For example: Structural boundaries were extracted from resistivity model, density model, and wave velocity model respectively (using edge detection algorithm).

[0072] These boundary information are weighted and fused to obtain a comprehensive structural boundary model.

[0073] Then, using this comprehensive structural boundary as a constraint, joint inversion can be performed or it can be used as a reference for geological interpretation.

[0074] In one embodiment, step S6, displaying the multi-source fused data using a 3D geographic information system platform, includes: Using a 3D Geographic Information System (3DGIS) platform, the processed multi-source data is presented in an intuitive way, supporting basic spatial query and analysis functions. Specific content is shown below: Resistivity volume model: Displays the resistivity distribution of underground media in the form of a three-dimensional grid or isosurface. It uses color coding (e.g., blue represents low resistivity and red represents high resistivity) to intuitively reflect the differences in electrical properties of geological bodies. It allows for layered viewing of resistivity slices at different depths (e.g., displaying a horizontal slice every 10 meters).

[0075] Gravity anomaly data: These are overlaid on the resistivity model as semi-transparent isosurfaces or colored shading maps to highlight areas of density anomalies (such as high gravity values ​​caused by ore bodies).

[0076] Seismic interface data: Reflection interfaces interpreted from seismic events (such as the top surface of bedrock and fault interfaces) are embedded into the 3D model as red wireframes or curved surfaces, and the interface depth and attitude (dip angle, dip direction) are marked. Geothermal field data: The underground temperature distribution is displayed in the form of isotherms or thermal cloud maps, superimposed on the resistivity model to help identify geothermal reservoirs (such as low resistivity + high temperature anomaly areas).

[0077] Its functions include: Arbitrary sectioning: Supports generating horizontal, vertical, or oblique sections by dragging the mouse or inputting coordinates, displaying real-time 2D curves or contour maps of parameters such as resistivity and gravity gradient at the section, and marking the depth and characteristics of key geological interfaces (such as aquifers and fault zones). Attribute querying: Clicking on any location or geological body in the 3D model will pop up an attribute window displaying the coordinates, resistivity value, gravity value, and other geoscientific parameters of that point, as well as the corresponding borehole or seismic interpretation results.

[0078] For multi-temporal data (such as exploration results from different years), it supports switching between 3D models of different periods by sliding the time axis, displaying the change areas of parameters such as resistivity in the form of difference cloud maps, and marking the magnitude and trend of change (such as the diffusion path of groundwater pollution).

[0079] Automatically or manually mark areas of geological anomalies discovered during exploration (such as low resistivity anomalies, areas with high gravity values) with yellow or orange spheres / ellipsoids, accompanied by a brief description (such as "suspected water-bearing body, depth 150-200 meters").

[0080] It supports batch import of borehole locations and displays the lithology and resistivity logging curves revealed by the boreholes in the form of bar charts, which can be superimposed on the 3D model to assist in verification.

[0081] Based on multi-source data such as resistivity, gravity, and seismic data, combined with a geological knowledge base, high-potential exploration target areas are automatically identified and marked with red polygons or grids, along with information such as target area number, area, and main anomaly characteristics (e.g., "Target area A: Low resistivity anomaly range 500m×300m, presumed to be a concealed ore body").

[0082] Generate a target area priority list, sorted by indicators such as anomaly intensity and geological consistency, and support export to Excel spreadsheet.

[0083] This system integrates 3D models, anomaly analysis results, and exploration recommendations to automatically generate a standardized report, including: Text description: Project overview, data processing flow, and major geological findings; Attached charts and graphs: Resistivity contour maps of key strata, gravity anomaly planar maps, and interpretation diagrams of typical cross-sections; 3D model link: Embedded interactive WebGL format 3D scene, allowing direct viewing in a browser. The report supports PDF and Word formats, with key charts and graphs automatically numbered, and data citations consistent with analysis conclusions.

[0084] This application provides a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, it implements the above-described method for optimizing geophysical electromagnetic exploration data based on big data.

[0085] This application provides a computer program product, including a computer program / instruction, which, when executed by a processor, implements the aforementioned method for optimizing geophysical electromagnetic exploration data based on big data.

[0086] In summary, this application has the following technical effects: 1. A multi-band progressive acquisition mode, employing low-frequency (0.1Hz-10kHz) coarse scanning, mid-frequency (10-100kHz) fine exploration, and high-frequency (100kHz-1MHz) verification, systematically solves the core contradiction in traditional single-frequency exploration where "penetration depth and resolution cannot be simultaneously achieved." The low-frequency band utilizes the large skin depth to rapidly delineate the macroscopic outline of deep geological anomalies (such as the distribution range of concealed ore bodies and fault zones) with sampling intervals of 50-100 meters, providing macroscopic spatial positioning for exploration. The mid-frequency band, through phased array focusing technology (main lobe width <15°) and Bayesian inversion constraints, controls the resistivity model boundary error to within 8%, accurately depicting the morphological details of geological bodies (such as stratigraphic interfaces, ore body strike and thickness). The high-frequency band employs co-offset acquisition and wavelet packet decomposition technology to identify shallow, fine geological features (such as shallow oxidation zones within ore bodies and shallow groundwater distribution), verifying the reliability of the mid- and low-frequency band exploration results. This technology achieves full-depth coverage of "macro-control of deep structure - fine interpretation of anomalies in the middle and shallow layers - dynamic verification of shallow features", which improves the anomaly identification rate in complex geological scenarios by 60% and shortens the exploration cycle by 30%–40% compared with traditional single-band methods.

[0087] 2. Based on a real-time interference identification algorithm using Short-Time Fourier Transform (STFT) and Convolutional Neural Network (CNN), and in conjunction with multiple technology modules such as adaptive notch filtering, intelligent frequency band switching, and Kalman filtering, this system achieves accurate classification and dynamic suppression of narrowband interference (e.g., 50Hz power frequency), broadband interference (e.g., mobile communication signals), and random impulse interference (e.g., lightning noise). For narrowband interference, the adaptive notch filter can dynamically adjust the center frequency according to the interference frequency (Q value can be adjusted up to 30), with an attenuation of -30dB, effectively preserving the target signal. For broadband interference, the system automatically switches to an adjacent clean frequency band (frequency error < 2kHz) and adjusts the modulation method (e.g., AM→FM), improving anti-interference capability by 10dB. For random impulse interference, the Kalman filter, through an iterative mechanism of state prediction and measurement update (typically 5 iterations), reduces noise amplitude by more than 80% and improves the signal-to-noise ratio by 12–15dB. This technology increases the effective data integrity rate from 75% to 98% under complex interference environments, and improves the average data signal-to-noise ratio by 15dB, providing highly reliable input data for subsequent inversion and significantly reducing the risk of abnormal misjudgment or missed detection caused by interference.

[0088] 3. By combining coarse- and fine-scale joint modeling of Occam inversion and conditional generative adversarial networks (cGAN), and integrating cross-physics field coupling models such as electromagnetic-gravity, electromagnetic-seismic, and electromagnetic-geothermal, a full-process inversion system of "macro-constraints—detail optimization—cross-domain verification" was constructed, significantly reducing the ambiguity of single methods and improving the accuracy of geological interpretation. Occam inversion, by introducing hard constraints such as borehole core resistivity and seismic reflection interface depth, constructs a coarse-scale model with a resolution of 50–100 meters, solving the framework problem of deep geological structures; cGAN inversion utilizes the U-Net neural network architecture to extract high-frequency detailed features (such as 10–20 meter-level thin veins and internal structures of fault zones), improving the model resolution to 10–20 meters, with a root mean square error (RMSE) of <8% between predicted and measured resistivity; multi-source data fusion, through physics field coupling models (such as density-resistivity empirical formula models), constructs a "macro-constraints—detail optimization" system. —The “cross-domain verification” full-process inversion system significantly reduces the ambiguity of single methods and improves the accuracy of geological interpretation. Occam inversion introduces hard constraints such as borehole core resistivity and seismic reflection interface depth to construct a coarse-scale model with a resolution of 50–100 meters, solving the framework problem of deep geological structures. cGAN inversion uses the U-Net neural network architecture to extract high-frequency details (wave velocity-resistivity probability model), achieving cross-domain data cross-validation, controlling the ore body boundary error within 20 meters (electromagnetic-gravity combined) and the stratigraphic interface depth error < ±8 meters (electromagnetic-seismic combined). This technology improves the comprehensive interpretation accuracy of complex geological bodies from 50%–60% of traditional single methods to over 80%, increases the drilling hit rate to 75%–80%, and improves the inversion calculation efficiency by 60% (iterations < 50 times), providing efficient and accurate technical support for deep resource exploration and geothermal reservoir assessment.

[0089] The above descriptions are merely various embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for optimizing geophysical electromagnetic exploration data based on big data, characterized in that, include: Collect geological prior data, filter preliminary frequency bands based on the geological prior data through a three-level filtering mechanism, and conduct preliminary exploration according to the planned node deployment based on the preliminary frequency bands; Based on the preliminary exploration results, the final score is calculated, and the geological structure of the exploration target is determined based on the final score; If the geological structure of the exploration target is a complex geological scene, multi-frequency band collaborative acquisition should be used for precise exploration; Cross-frequency band data fusion is performed on the precisely detected data, and coarse-scale modeling is performed on the fused data based on Occam inversion. GAN inversion is then used to optimize the coarse-scale modeling results for fine-scale optimization. The results of fine-scale optimization are fused with gravity data, seismic data, and geothermal data to obtain multi-source fused data. The multi-source fusion data is displayed using a three-dimensional geographic information system platform.

2. The method as described in claim 1, characterized in that, in, Based on the aforementioned geological prior data, a preliminary frequency band is selected through a three-level filtering mechanism, including: Determine the depth of the exploration target. If the depth is very shallow, select the high-frequency band; if the depth is shallow to medium, select the medium-frequency band; if the depth is deep, select the low-frequency band. Obtain the lithological distribution layer from the geological prior data, and combine it with borehole core data to obtain the lithological combination characteristics of the exploration target, and adjust the selected frequency band according to the lithological combination characteristics; Obtain the fracture density, stratigraphic dip angle, and intrusive body area ratio from the geological prior data, and calculate the structural complexity index; if the structural complexity index is greater than a set threshold, the area where the exploration target is located is a complex structural zone, and then reduce the frequency interval and increase the redundant frequency band.

3. The method as described in claim 1, characterized in that, in, Based on the preliminary exploration results, a final score is calculated, and the geological structure of the exploration target is determined based on the final score, including: Based on the preliminary exploration results, resistivity, gradient change, and noise energy ratio are determined. If the resistivity exhibits an abnormal and irregular shape, and the gradient change is greater than the gradient change threshold, and the noise energy ratio is greater than the noise energy ratio threshold, then the resistivity abnormal shape score, noise energy ratio score, and signal-to-noise ratio improvement score are calculated, and the final score is obtained. If the final score is greater than the score threshold, then the geological structure of the exploration target is a complex geological scene.

4. The method as described in claim 1, characterized in that, in, Precise exploration is achieved through multi-band collaborative acquisition, including: Multi-frequency synchronous transmission technology is adopted to simultaneously transmit multiple frequency signals in the low-frequency band; compressed sensing algorithm is used to collect key data points, and the underground structure outline is reconstructed based on the key data points using the L1 norm optimization algorithm; spatial domain downsampling is performed at fixed sampling intervals to determine potential geological anomaly areas. The frequency signal is adjusted to be in the mid-frequency band. For anomalous areas, phased array focusing technology is used to adjust the sensor's emission phase to form a beam focus. The sliding time window correlation analysis method is used to calculate the signal correlation within different time windows to identify stratigraphic interfaces and geological body boundaries. The frequency signal is adjusted to be in the high-frequency band, and a common offset acquisition mode is adopted. The signal is decomposed using wavelet packet decomposition technology to obtain the analysis results. Based on the analysis results and the time-frequency electromagnetic response feature library of typical geological bodies, shallow geological anomalies are identified by matched filtering algorithm to verify the exploration results in the mid-to-low frequency band.

5. The method as described in claim 1, characterized in that, During the multi-band collaborative acquisition process, the acquisition period is set, and short-time Fourier transform and convolutional neural network are used to identify interference in the acquired data in real time. If the interference is narrowband interference, an adaptive notch filter is enabled, and the filter center frequency is dynamically adjusted according to the interference frequency. If the interference is broadband interference, the process is switched to an adjacent frequency band and the modulation method is adjusted. If the interference is random pulse interference, the Kalman filter algorithm is used to filter out noise through an iterative process of state prediction and measurement update.

6. The method as described in claim 5, characterized in that, The convolutional neural network includes three parallel convolutional branches, a channel attention module, and a depthwise separable convolution; the three convolutional branches use three different sizes of convolutional kernels, namely 1×1, 3×3, and 5×5, to extract narrowband interference features, wideband interference features, and random impulse interference features, respectively.

7. The method as described in claim 1, characterized in that, The coarse-scale modeling uses borehole core data and seismic stratigraphic depth as hard constraints, adds constraint terms to the inversion objective function, and employs regularized inversion to construct the Occam inversion model. The fine-scale optimization adopts a conditional generative adversarial network architecture. The generator is input with the Occam inversion model and geological constraint parameters and outputs a high-resolution resistivity model. The discriminator is input with real borehole data and generator output data, and the authenticity of local features is judged by a loss function.

8. The method as described in claim 1, characterized in that, in, The multi-source fused data is displayed using a 3D geographic information system platform, including: The resistivity distribution of the subsurface medium is displayed in the form of a 3D grid or isosurface, and the differences in electrical properties of the geological body are reflected by color coding. Resistivity slices at different depths are set. Gravity anomaly data are superimposed on the resistivity model as semi-transparent isosurfaces or colored shading maps to show density anomaly areas. The reflection interface interpreted by the seismicity is embedded in the 3D model as a red wireframe or curved surface, and the interface depth and attitude are marked. The subsurface temperature distribution is displayed in the form of isotherms or thermal cloud maps.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the geophysical electromagnetic exploration data optimization method based on big data as described in any one of claims 1-8.

10. A computer program product, characterized in that, Includes a computer program / instruction, which, when executed by a processor, implements the geophysical electromagnetic exploration data optimization method based on big data as described in any one of claims 1-8.

Images