Radioactive measurement data processing method for improving uranium exploration efficiency

Abstract

The invention discloses a radioactive measurement data processing method for improving uranium exploration efficiency. The method comprises the following steps: standardized data acquisition: integrating a multi-parameter module, carrying out time-space synchronous acquisition of geological geophysical environment parameters, standardizing protocol alignment data, and supporting three-dimensional modeling; a three-dimensional coupling model is used for processing interference in a sub-module mode, LiDAR-DEM is used for correcting gamma attenuation in the terrain, an optical fiber thermopermeability instrument is used for correcting daughter errors in the hydrology, signal attenuation of a borehole is compensated through a drilling fluid chart, and a space-time continuous interference field is formed; performing dynamic equilibrium coefficient inversion: constructing equilibrium coefficient isoparametric mapping, and inputting rock core, logging and geochemical data into a random forest model; the transfer learning is trained by using historical data, and dynamic inversion and updating of a new area are carried out; intelligent data processing three-dimensional visualization: self-adaptive denoising and spectral shape matching reinforcement abnormity are carried out; constructing a three-dimensional model, and carrying out transfer learning to optimize the precision; the WebGL platform integrates multi-parameter display, virtual drilling functions and high-dimensional data intuitive interpretation.

Description

Technical Field

[0001] The present invention relates to the technical field of uranium exploration, and in particular to a method for processing radioactive measurement data for improving uranium exploration efficiency. Background Art

[0002] As a core strategic resource for nuclear energy development, the accuracy of uranium exploration directly impacts energy security and the achievement of carbon neutrality goals. Uranium deposits are diverse globally, and exploration requires a comprehensive approach using geological, geophysical, and geochemical methods. Radiometric measurement, due to uranium's naturally radioactive properties, is a key tool for locating ore bodies and assessing reserves.

[0003] However, traditional methods for processing radioactive measurement data suffer from a serious problem of insufficient correction for environmental interference. For example, in a sandstone-type uranium exploration project, traditional methods only corrected gamma signal attenuation through elevation data, ignoring nonlinear factors such as slope and surface vegetation. In hilly areas, actual measurements found that in areas with slopes >30°, the gamma signal attenuation rate was underestimated by 25%-40%, leading to misjudgment of ore body boundaries. When the surface is covered with gravel layers, traditional models fail to correct for the scattering effect and mistakenly classify background radiation as a mineralization signal, forming false anomalies. The exploration team deployed drill holes in the wrong area, resulting in an empty drill rate of 35%, with direct economic losses exceeding 10 million yuan. Therefore, a method for processing radioactive measurement data to improve the efficiency of uranium exploration is proposed. Summary of the Invention

[0004] The purpose of the present invention is to solve the shortcomings of the prior art and to propose a method for processing radioactive measurement data to improve the efficiency of uranium exploration.

[0005] In order to achieve the above object, the present invention adopts the following technical solutions:

[0006] A method for processing radioactivity measurement data for improving uranium exploration efficiency, comprising:

[0007] Standardized data acquisition: Deploy a multi-parameter synchronous measurement module, integrating a gamma spectrometer (NaI(Tl) / LaBr3 dual probes), a micro-meteorological station (temperature, humidity, air pressure / wind speed monitoring), a ground-penetrating radar (for detecting geological structures 50m underground), and a fiber-optic gyroscope (for recording instrument attitude angles), to synchronously collect geological, geophysical, and environmental parameters in both time and space. Through standardized equipment configuration and data protocols, the raw data is aligned in time and space, providing high-quality input for subsequent 3D modeling.

[0008] Modeling the environmental interference field: A three-dimensional coupled correction model of "geology-geophysics-environment" is constructed, and multi-source interference is processed in modules: terrain correction is based on DEM data obtained by airborne LiDAR, calculating the terrain correction index (TCI) and correcting gamma signal attenuation; hydrological correction uses distributed fiber-optic temperature and permeability meters to monitor groundwater flow, establish a hydrological-radioactivity coupling model, and correct for daughter retention errors; wellbore correction develops a drilling fluid absorption correction chart, and combines wellbore-density parameters to dynamically compensate for signal attenuation, ultimately forming a spatiotemporally continuous environmental interference field;

[0009] Dynamic equilibrium coefficient inversion: This involves mapping the equilibrium coefficient to geological parameters such as lithology, porosity, and permeability. A data-driven prediction model is developed using a random forest algorithm. Input parameters include core analysis data (U / Ra / Pb content), well logs (natural gamma ray / density / neutron porosity), and geochemical data (organic carbon / redox potential). A transfer learning network is introduced, and the model is trained using historical mining data to dynamically invert and update the equilibrium coefficient in new areas in both time and space.

[0010] Intelligent data processing and 3D visualization: Data purification and anomaly enhancement are achieved through an adaptive wavelet threshold denoising algorithm (dynamically adjusting the threshold based on the signal-to-noise ratio) and spectral shape matching enhancement technology (comparing with a standard uranium spectrum library to enhance the signal by 3-5 times). A machine learning model is then combined to construct a 3D geological model, and accuracy is continuously optimized through transfer learning. Finally, a WebGL interactive platform is developed, integrating multi-parameter collaborative display, virtual drilling, profile cutting, and UMAP dimensionality reduction visualization functions to enable intuitive interpretation of high-dimensional data and rapid identification of anomaly patterns.

[0011] The above technical solution further includes:

[0012] Furthermore, the temporal and spatial synchronous acquisition of geological, geophysical and environmental parameters comprises the following steps:

[0013] Device deployment:

[0014] Gamma spectrometer: Equipped with dual probes (NaI(Tl) / LaBr3), it simultaneously measures the total gamma count rate and energy spectrum data (including characteristic peaks of uranium, thorium, and potassium), with a sampling interval of 1 second to ensure high temporal resolution of radioactive signals. Micrometeorological station: Integrated with temperature, humidity, air pressure, and wind speed sensors, its sampling frequency is synchronized with the gamma spectrometer (1 second / time), recording the impact of environmental parameters on radioactive measurements (such as the effect of atmospheric turbulence on gamma-ray scattering). Ground penetrating radar: With a detection range of 50m, it uses stepped frequency continuous wave (SFCW) mode, triggering scans every 0.5m along the survey line to obtain a two-dimensional profile of the underground geological structure (rock interfaces and fault zones). Fiber optic gyroscope: Real-time recording of the instrument's three-dimensional attitude angles (pitch, roll, and heading), with a sampling interval of 1 second, is used to correct for changes in the gamma probe's effective area caused by instrument tilt.

[0015] Synchronous trigger:

[0016] All devices are aligned using GPS timestamps and synchronized sampling clocks using PPS (pulse per second) signals to ensure a time dimension error of less than 0.1 seconds. RTK-GPS (real-time kinematic differential positioning) is used for spatial positioning, with a horizontal accuracy of less than 2 cm and a vertical accuracy of less than 5 cm. This is forcibly linked to the position of the ground-penetrating radar survey line to unify spatial coordinates.

[0017] Multi-parameter fusion recording:

[0018] Data acquisition software (such as MultiSensorLogger) packages gamma spectrum, meteorological parameters, ground penetrating radar images, and attitude angle data in time series to generate structured data files (such as HDF5 format) containing metadata (device ID, sampling time, latitude and longitude, and elevation).

[0019] Furthermore, adjusting the alignment of the original data in time and space dimensions by standardizing the device configuration and data protocol includes the following steps:

[0020] Standardized equipment configuration:

[0021] The gamma spectrometer uses a standard source (such as 137Cs) for energy calibration, ensuring that the energy spectrum peak position accuracy is less than 1 keV. Micro-meteorological station sensors are cross-calibrated in a laboratory environment (for example, the temperature and humidity sensors are compared with primary standards, with an error of ≤2%). The spacing between ground-penetrating radar antennas is fixed at 0.2m to avoid detection depth deviations due to hardware differences. A unified data format (such as NetCDF or SegY) is used, defining field names, units, and coordinate systems (WGS84). The timestamp format is set to ISO8601 (UTC time zone), and the spatial coordinates use the UTM projection (with the central meridian adjusted according to the survey area).

[0022] Align the time dimension:

[0023] For devices with inconsistent sampling frequencies (e.g., 0.5m scanning interval for ground penetrating radar and 1 second / time for gamma spectrometer), all parameters are unified to a 1-second time resolution through linear interpolation. Time offset correction is performed on the lag parameters based on the device response time (e.g., sensor delay of micro-meteorological station ≤ 0.2 seconds) to ensure multi-parameter synchronization.

[0024] Align spatial dimensions:

[0025] Convert the RTK-GPS latitude and longitude coordinates to the local coordinate system of the survey area (such as UTM Zone 50N) to match the GPR survey line position; perform Kriging interpolation on discrete sampling points (such as gamma spectrometer point measurement data) to generate a regular grid with a spatial resolution of 1m×1m, which is then fused with the GPR continuous profile data;

[0026] Quality Control:

[0027] The Z-score method is used to identify abnormal data after temporal and spatial alignment (such as a sudden increase in gamma count rate without corresponding geological structure), mark and eliminate interference (such as false signals caused by instrument jitter), and compare the total count rate of the gamma spectrometer with the rock layer thickness revealed by the ground penetrating radar. If the difference exceeds 15%, the recalibration process is triggered.

[0028] Furthermore, the construction of the geological-geophysical-environmental three-dimensional coupled correction model and the processing of multi-source interference in modules include the following steps:

[0029] Terrain Correction:

[0030] Deploy a drone or helicopter equipped with LiDAR and plan a route along the survey area (altitude ≤ 500m, lateral overlap ≥ 60%). Laser pulses (wavelength 1064nm, pulse frequency 500kHz) are emitted to obtain raw point cloud data. This point cloud data is then denoised (excluding non-ground points such as vegetation and buildings) and filtered (using a progressively encrypted triangulated network filtering algorithm) to generate a high-precision digital elevation model (DEM). The terrain slope θ (unit: degrees) is extracted from the DEM:

[0031]

[0032] Where Δx and Δy are the horizontal position changes, and Δz is the vertical position change rate;

[0033] Measuring instrument height H (relative to the ground), combined with TCI = (1 + tanθ) 2 / H calculates the terrain correction index (TCI), which quantifies the effect of terrain relief on gamma-ray scattering (the steeper the slope and the lower the instrument, the more significant the scattering attenuation);

[0034] According to the experimental calibration results, an exponential relationship model between TCI and gamma count rate decay is established:

[0035] N corr =N raw ×e 0.3×TCI

[0036] Among them, N raw is the raw count rate, N corr is the corrected count rate; the model is applied to each sampling point in the survey area to generate gamma field data after terrain correction;

[0037] Hydrological correction:

[0038] Distributed fiber optic sensors (1m spacing, 0-50m depth coverage) are buried along the survey line, integrating temperature and seepage velocity monitoring functions. Through Raman scattering and Brillouin scattering technology, underground temperature field and seepage velocity field data are obtained in real time with a time resolution of 1 minute. Combining temperature field data (groundwater flow will change the local temperature gradient) with seepage velocity field data, the groundwater flow direction and velocity are inverted through the two-dimensional water flow equation (Darcy's law), and the uranium decay chain daughters (such as 226 Ra, 222 The migration velocity of Rn) in groundwater (calibrated by core leaching experiments), the migration velocity of the daughter body is linearly related to the groundwater flow rate;

[0039] Establishing a hydrological-radiological coupling model:

[0040] ΔN=N1×(1-e -λ×t )

[0041] Where ΔN is the gamma count rate deviation caused by daughter retention, N1 is the gamma count rate of the parent nuclide, λ is the decay constant of the daughter nuclide, and t is the time it takes for the daughter to separate from the parent nuclide. The gamma count rate deviation ΔN caused by daughter retention is calculated according to the model and the original data is corrected:

[0042] N corr =N raw -ΔN;

[0043] Wellbore Correction:

[0044] The laboratory simulates different borehole conditions: the drilling fluid containing barite (0%-50%) and clay (5%-20%) is prepared, and its absorption coefficient of gamma rays (μ, unit: cm -1 ); calculate and establish the relationship between the absorption rate and the drilling fluid composition; make a calibration chart: input the well diameter (D, unit: cm), the drilling fluid density (ρ, unit: g / cm 3 ), through the formula K=e μ×D / ρCalculate the signal attenuation compensation, where K is the compensation coefficient, μ is the absorption coefficient, D is the input wellbore diameter, and ρ is the drilling fluid density. During the logging process, the wellbore diameter D is measured in real time by a caliper (resolution 0.1 cm), and the density logging tool (resolution 0.01 g / cm 3 ) Measure the drilling fluid density ρ;

[0045] Call the calibration chart, query the compensation coefficient based on D and ρ, and correct the gamma count rate:

[0046] N corr =N raw ×K;

[0047] Three-dimensional coupled environmental interference field integration:

[0048] The gamma field after terrain correction, the daughter correction field after hydrological correction, and the signal compensation field after borehole correction are interpolated in the spatiotemporal dimension through Gaussian process regression (GPR) to generate a continuous environmental interference field (spatial resolution 1m, temporal resolution 1min);

[0049] A verification point is set above a known uranium ore body, and the gamma count rate before and after correction is compared with the actual uranium content (obtained through core analysis). The error is required to be less than 10%. If the error exceeds the standard, the GPR kernel function parameters (such as length scale and variance) are adjusted or additional monitoring data (such as increasing the LiDAR scanning frequency) are supplemented to regenerate the interference field.

[0050] Furthermore, the random forest algorithm is used to establish a data-driven prediction model, including the following steps:

[0051] Data preparation:

[0052] Integrate core analysis data, well log curves (such as resistivity, natural gamma, density), and geochemical data to build a multi-dimensional geological parameter database; handle missing values ​​(such as interpolation) and outliers (based on statistical methods or business understanding to eliminate), and perform normalization (such as Z-score standardization) to ensure consistent feature dimensions; divide the data into training and test sets in a ratio (such as 7:3) to ensure representative sample distribution;

[0053] Feature construction:

[0054] Through the feature importance evaluation provided by random forest (such as Gini importance or permutation-based importance), the features that contribute significantly to the prediction of the equilibrium coefficient (such as porosity, permeability, and lithology category) are screened;

[0055] Combine domain knowledge to construct interactive features (such as the product of porosity and permeability) or time series features (such as the time-varying attenuation rate of gamma signals);

[0056] Build the model:

[0057] Set key parameters for random forests, including the number of decision trees, maximum depth, and minimum number of sample splits; fit the random forest model using training data and preliminarily evaluate model performance using out-of-bag error; utilize multi-core CPUs to build decision trees in parallel to accelerate the training process;

[0058] Cross-validation:

[0059] The data is divided into 5 folds, 4 folds are used for training and 1 fold is used for validation each time, and the average RMSE and R are calculated. 2 To evaluate the stability of the model; traverse and optimize the key parameters and select the parameter combination that performs best on the validation set; terminate the training early when the performance of the validation set no longer improves to prevent overfitting.

[0060] Model Evaluation Explanation:

[0061] Calculate RMSE, MAE, R on the test set 2 The results are compared with benchmark models (such as linear regression and support vector machine), and the contribution of each feature to the prediction is displayed through bar charts to assist geological interpretation; the nonlinear relationship between key features (such as porosity) and the equilibrium coefficient is analyzed to verify the rationality of the geological mechanism.

[0062] Furthermore, the introduction of the transfer learning network, the use of historical mining area data to train the model, and the dynamic inversion and spatiotemporal update of the new regional balance coefficient include the following steps:

[0063] Build a pre-trained model (source domain):

[0064] Collect high-quality data from historical mining areas (such as multi-year well logging records and core analysis data) to build a large-scale training set; use the random forest algorithm to train the basic model on historical data and save the model parameters (such as decision tree structure and feature importance); through feature importance analysis, identify key features that are common across regions (such as lithology and parent element content).

[0065] Transfer learning applications (target domain):

[0066] Use the pre-trained model as the initial parameters for the new regional model, freeze some underlying feature layers (such as the lithology identification layer), and adjust only the top prediction layer. Perform supervised training on the top layer of the model using a small sample of data from the new region, adjusting the weights to adapt to local geological characteristics. Combined with unlabeled data from the new region (such as real-time gamma signals), further optimize the model through self-supervised learning (such as contrastive learning).

[0067] Dynamic inversion:

[0068] Real-time measurement data from the new area (such as gamma spectrometer spatiotemporal data) is input into the migrated model to generate an initial balance coefficient prediction; based on geological movement models (such as fault activity and hydrological changes), the spatiotemporal evolution trend of the balance coefficient is predicted; combined with real-time measurement data, the prediction results are adjusted through particle resampling to reduce uncertainty; and using kriging interpolation or deep learning spatiotemporal models (such as ST-GCN) to fill in the prediction gaps in areas with sparse monitoring points and generate a continuous spatiotemporal distribution map.

[0069] Furthermore, the data purification and anomaly enhancement are performed by using the adaptive wavelet threshold denoising algorithm and spectral shape matching enhancement technology, including the following steps:

[0070] Adaptive wavelet threshold denoising:

[0071] Select a wavelet basis (such as Daubechies-4 or Symlet-5) based on data characteristics (such as the time-frequency characteristics of the gamma spectrum) to balance time domain resolution and frequency domain localization capabilities. Set the number of wavelet decomposition layers to 5 to ensure that low-frequency components (approximate coefficients) contain geological trend information and high-frequency components (detail coefficients) contain noise and weak anomalies. For each layer of detail coefficients, calculate the signal-to-noise ratio (SNR) of the current layer:

[0072]

[0073] Where P1 is the signal power and P2 is the noise power. The signal power is estimated by the approximation coefficient, and the noise power is estimated by the median absolute deviation (MAD) of the detail coefficient. The threshold is adjusted according to the SNR. When the SNR is high (>20dB), a soft threshold is used. Preserve more details; when the SNR is low (<10dB), use a hard threshold λ = 3σ to forcibly remove obvious noise; where σ is the noise standard deviation and N is the signal length; apply a dynamic threshold to each layer of detail coefficients to remove noise components below the threshold and retain abnormal signals above the threshold; reconstruct the denoised signal through inverse wavelet transform to ensure time domain continuity and frequency domain accuracy;

[0074] Enhanced spectral shape matching:

[0075] Collect gamma ray spectrum data of typical uranium ore bodies (including 238U, 235 U. 232 Th series has 16 characteristic peaks, such as 4.4keV (Pb X-ray), 18.6keV ( 238 U), 238.6keV( 232Th)) to establish a standard spectrum library; perform energy calibration and resolution correction on the spectrum library data to ensure that the characteristic peak position accuracy is ≤0.5keV; perform smoothing on the denoised gamma energy spectrum (such as Savitzky-Golay filtering) to suppress statistical fluctuation noise; calculate the correlation coefficient (Pearson or Spearman) between the spectrum to be measured and the standard spectrum, and identify the standard spectrum with the highest matching degree (such as the 238U dominant spectrum); based on the matching results, adjust the amplitude and baseline of the spectrum to be measured to align it with the standard spectrum; for characteristic peak areas with a matching degree greater than 0.8, use linear amplification (such as 2-5 times) to highlight weak abnormal signals while suppressing noise in non-matching areas.

[0076] Furthermore, the method of combining a machine learning model to construct a three-dimensional geological model and continuously optimizing the accuracy through transfer learning includes the following steps:

[0077] The machine learning model builds an initial 3D model:

[0078] De-noised gamma field data, environmental interference field data, core analysis data, and well logging curves were integrated to construct a multi-source feature matrix (such as gamma count rate, terrain correction index, and porosity). Continuous features (such as gamma values) were logarithmically transformed, and categorical features (such as lithology) were one-hot encoded to eliminate dimensional differences. A convolutional neural network (CNN) was selected as the basic model, with input being 2D profile data (such as gamma field slices) and output being 3D geological parameters (such as uranium content and lithology category).

[0079] The network structure is set as follows: convolution layer (3×3 convolution kernel, step size 1) + batch normalization layer + ReLU activation function × 4 layers, and finally a fully connected layer to output the prediction results; Adam optimizer is used, the learning rate is initially set to 0.001, decayed by 5% every 10 rounds, the training cycle is 50 rounds, and the batch size is 32; the prediction accuracy (such as R 2 =0.85, RMSE=1.2ppm), and the model's ability to identify faults and lithologic interfaces was checked through 3D visualization;

[0080] Transfer learning continuously optimizes accuracy:

[0081] Use high-quality data from historical mining areas (such as multi-year exploration data and high-precision well logging records) to train a general model, preserving the network weights and feature extraction layer parameters. Use the convolutional layer weights of the pre-trained model as the initial parameters of the new model, freeze the first two layers (to extract general geological features, such as fault patterns), and only adjust the last two layers (to adapt to the special geological phenomena in the new area).

[0082] Fine-tuning strategy:

[0083] Supervised fine-tuning: The top layer of the model is trained using a small sample of annotated data from the new area (e.g., drill hole data) with a learning rate of 0.0001 to avoid overfitting.

[0084] Unsupervised adaptation: Incorporating unlabeled data from new regions (e.g., real-time gamma fields) and optimizing feature representation through self-supervised learning (e.g., contrastive learning) can improve the model’s sensitivity to the geological characteristics of the new region.

[0085] Dynamic Updates:

[0086] Real-time measurement data from the new area is input into the migrated model to generate an initial three-dimensional geological model. Based on geological movement models (such as fault activity and hydrological changes), the spatiotemporal evolution trends of geological parameters are predicted, and the prediction results are adjusted through particle resampling combined with real-time data. As data from new areas accumulates, new data is regularly added to the training set, and model parameters are updated through incremental learning, forming a closed loop of "prediction-verification-correction".

[0087] The present invention has the following beneficial effects:

[0088] In the present invention, a multi-parameter synchronous measurement module is deployed through standardized data acquisition to carry out spatiotemporal synchronous acquisition of geological, geophysical and environmental parameters, and the data are aligned through standardized protocols to provide high-quality input for subsequent processing; a three-dimensional coupling correction model is constructed when modeling the environmental interference field, and multi-source interference is processed in modules to form a spatiotemporal continuous interference field and input into a random forest model to correct environmental deviations; dynamic balance coefficient inversion uses random forest and transfer learning to establish a mapping relationship between the balance coefficient and lithology / porosity / permeability, and perform dynamic inversion of new areas; intelligent data processing three-dimensional visualization purifies data through adaptive wavelet denoising and spectral matching, combines machine learning to build a three-dimensional model and optimizes accuracy through transfer learning, and finally integrates multi-parameter display and UMAP dimensionality reduction functions through the WebGL platform, reversely guides the data acquisition module to adjust parameters, and forms a "collection-correction-prediction-optimization" closed loop, which effectively solves the problem of insufficient environmental interference correction in traditional methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0089] Figure 1 This is a system block diagram of a method for processing radioactive measurement data to improve uranium exploration efficiency, proposed by the present invention. DETAILED DESCRIPTION

[0090] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0091] See also Figure 1 As shown, the present invention is a method for processing radioactivity measurement data to improve uranium exploration efficiency, comprising:

[0092] A method for processing radioactivity measurement data for improving uranium exploration efficiency, comprising:

[0093] Standardized data acquisition: Deploy a multi-parameter synchronous measurement module, integrating a gamma spectrometer (NaI(Tl) / LaBr3 dual probes), a micro-meteorological station (temperature, humidity, air pressure / wind speed monitoring), a ground-penetrating radar (for detecting geological structures 50m underground), and a fiber-optic gyroscope (for recording instrument attitude angles), to synchronously collect geological, geophysical, and environmental parameters in both time and space. Through standardized equipment configuration and data protocols, the raw data is aligned in time and space, providing high-quality input for subsequent 3D modeling.

[0094] Modeling the environmental interference field: A three-dimensional coupled correction model of "geology-geophysics-environment" is constructed, and multi-source interference is processed in modules: terrain correction is based on DEM data obtained by airborne LiDAR, calculating the terrain correction index (TCI) and correcting gamma signal attenuation; hydrological correction uses distributed fiber-optic temperature and permeability meters to monitor groundwater flow, establish a hydrological-radioactivity coupling model, and correct for daughter retention errors; wellbore correction develops a drilling fluid absorption correction chart, and combines wellbore-density parameters to dynamically compensate for signal attenuation, ultimately forming a spatiotemporally continuous environmental interference field;

[0095] Dynamic equilibrium coefficient inversion: This involves mapping the equilibrium coefficient to geological parameters such as lithology, porosity, and permeability. A data-driven prediction model is developed using a random forest algorithm. Input parameters include core analysis data (U / Ra / Pb content), well logs (natural gamma ray / density / neutron porosity), and geochemical data (organic carbon / redox potential). A transfer learning network is introduced, and the model is trained using historical mining data to dynamically invert and update the equilibrium coefficient in new areas in both time and space.

[0096] Intelligent data processing and 3D visualization: Data purification and anomaly enhancement are achieved through an adaptive wavelet threshold denoising algorithm (dynamically adjusting the threshold based on the signal-to-noise ratio) and spectral shape matching enhancement technology (comparing with a standard uranium spectrum library to enhance the signal by 3-5 times). A machine learning model is then combined to construct a 3D geological model, and accuracy is continuously optimized through transfer learning. Finally, a WebGL interactive platform is developed, integrating multi-parameter collaborative display, virtual drilling, profile cutting, and UMAP dimensionality reduction visualization functions to enable intuitive interpretation of high-dimensional data and rapid identification of anomaly patterns.

[0097] In one embodiment, the temporal and spatial synchronous acquisition of geological, geophysical and environmental parameters comprises the following steps:

[0098] Device deployment:

[0099] Gamma spectrometer: Equipped with dual probes (NaI(Tl) / LaBr3), it simultaneously measures the total gamma count rate and energy spectrum data (including characteristic peaks of uranium, thorium, and potassium), with a sampling interval of 1 second to ensure high temporal resolution of radioactive signals. Micrometeorological station: Integrated with temperature, humidity, air pressure, and wind speed sensors, its sampling frequency is synchronized with the gamma spectrometer (1 second / time), recording the impact of environmental parameters on radioactive measurements (such as the effect of atmospheric turbulence on gamma-ray scattering). Ground penetrating radar: With a detection range of 50m, it uses stepped frequency continuous wave (SFCW) mode, triggering scans every 0.5m along the survey line to obtain a two-dimensional profile of the underground geological structure (rock interfaces and fault zones). Fiber optic gyroscope: Real-time recording of the instrument's three-dimensional attitude angles (pitch, roll, and heading), with a sampling interval of 1 second, is used to correct for changes in the gamma probe's effective area caused by instrument tilt.

[0100] Synchronous trigger:

[0101] All devices are aligned via GPS timestamps and use PPS (pulse per second) signals to synchronize sampling clocks, ensuring a time dimension error of less than 0.1 seconds. RTK-GPS (real-time kinematic differential positioning) is used for spatial positioning, with a horizontal accuracy of less than 2 cm and a vertical accuracy of less than 5 cm. This is forcibly associated with the position of the ground-penetrating radar survey line to establish a spatial coordinate system.

[0102] Multi-parameter fusion recording:

[0103] Data acquisition software (such as MultiSensorLogger) packages gamma spectrum, meteorological parameters, ground penetrating radar images, and attitude angle data in time series to generate structured data files (such as HDF5 format) containing metadata (device ID, sampling time, latitude and longitude, and elevation).

[0104] In one embodiment, adjusting the alignment of raw data in time and space dimensions by standardizing device configuration and data protocols includes the following steps:

[0105] Standardized equipment configuration:

[0106] The gamma spectrometer uses a standard source (such as 137Cs) for energy calibration, ensuring that the energy spectrum peak position accuracy is less than 1 keV. Micro-meteorological station sensors are cross-calibrated in a laboratory environment (for example, the temperature and humidity sensors are compared with primary standards, with an error of ≤2%). The spacing between ground-penetrating radar antennas is fixed at 0.2m to avoid detection depth deviations due to hardware differences. A unified data format (such as NetCDF or SegY) is used, defining field names, units, and coordinate systems (WGS84). The timestamp format is set to ISO8601 (UTC time zone), and the spatial coordinates use the UTM projection (with the central meridian adjusted according to the survey area).

[0107] Align the time dimension:

[0108] For devices with inconsistent sampling frequencies (e.g., 0.5m scanning interval for ground penetrating radar and 1 second / time for gamma spectrometer), all parameters are unified to a 1-second time resolution through linear interpolation. Time offset correction is performed on the lag parameters based on the device response time (e.g., sensor delay of micro-meteorological station ≤ 0.2 seconds) to ensure multi-parameter synchronization.

[0109] Align spatial dimensions:

[0110] Convert the RTK-GPS latitude and longitude coordinates to the local coordinate system of the survey area (such as UTM Zone 50N) to match the GPR survey line position; perform Kriging interpolation on discrete sampling points (such as gamma spectrometer point measurement data) to generate a regular grid with a spatial resolution of 1m×1m, which is then fused with the GPR continuous profile data;

[0111] Quality Control:

[0112] The Z-score method is used to identify abnormal data after temporal and spatial alignment (such as a sudden increase in gamma count rate without corresponding geological structure), mark and eliminate interference (such as false signals caused by instrument jitter), and compare the total count rate of the gamma spectrometer with the rock layer thickness revealed by the ground penetrating radar. If the difference exceeds 15%, the recalibration process is triggered.

[0113] In one embodiment, the construction of a three-dimensional geological-geophysical-environmental coupled correction model and processing of multi-source interference in modules includes the following steps:

[0114] Terrain Correction:

[0115] Deploy a drone or helicopter equipped with LiDAR and plan a route along the survey area (altitude ≤ 500m, lateral overlap ≥ 60%). Laser pulses (wavelength 1064nm, pulse frequency 500kHz) are emitted to obtain raw point cloud data. This point cloud data is then denoised (excluding non-ground points such as vegetation and buildings) and filtered (using a progressively encrypted triangulated network filtering algorithm) to generate a high-precision digital elevation model (DEM). The terrain slope θ (unit: degrees) is extracted from the DEM:

[0116]

[0117] Where Δx and Δy are the horizontal position changes, and Δz is the vertical position change rate;

[0118] Measuring instrument height H (relative to the ground), combined with TCI = (1 + tanθ) 2 / H calculates the terrain correction index (TCI), which quantifies the effect of terrain relief on gamma-ray scattering (the steeper the slope and the lower the instrument, the more significant the scattering attenuation);

[0119] According to the experimental calibration results, an exponential relationship model between TCI and gamma count rate decay is established:

[0120] N corr =N raw ×e 0.3×TCI

[0121] Among them, N raw is the raw count rate, N corr is the corrected count rate; the model is applied to each sampling point in the survey area to generate gamma field data after terrain correction;

[0122] Hydrological correction:

[0123] Distributed fiber optic sensors (1m spacing, 0-50m depth coverage) are buried along the survey line, integrating temperature and seepage velocity monitoring functions. Through Raman scattering and Brillouin scattering technology, underground temperature field and seepage velocity field data are obtained in real time with a time resolution of 1 minute. Combining temperature field data (groundwater flow will change the local temperature gradient) with seepage velocity field data, the groundwater flow direction and velocity are inverted through the two-dimensional water flow equation (Darcy's law), and the uranium decay chain products (such as 226Ra, 222 The migration velocity of Rn) in groundwater (calibrated by core leaching experiments), the migration velocity of the daughter body is linearly related to the groundwater flow rate;

[0124] Establishing a hydrological-radiological coupling model:

[0125] ΔN=N1×(1-e -λ ×t)

[0126] Where ΔN is the gamma count rate deviation caused by daughter retention, N1 is the gamma count rate of the parent nuclide, λ is the decay constant of the daughter nuclide, and t is the time it takes for the daughter to separate from the parent nuclide. The gamma count rate deviation ΔN caused by daughter retention is calculated according to the model and the original data is corrected:

[0127] N corr =N raw -ΔN;

[0128] Wellbore Correction:

[0129] The laboratory simulates different borehole conditions: the drilling fluid containing barite (0%-50%) and clay (5%-20%) is prepared, and its absorption coefficient of gamma rays (μ, unit: cm -1 ); calculate and establish the relationship between the absorption rate and the drilling fluid composition; make a calibration chart: input the well diameter (D, unit: cm), the drilling fluid density (ρ, unit: g / cm 3), the signal attenuation compensation is calculated by the formula K = eμ × D / ρ, where K is the compensation coefficient, μ is the absorption coefficient, D is the input wellbore diameter, and ρ is the drilling fluid density. During the logging process, the wellbore diameter D is measured in real time by a caliper (resolution 0.1 cm), and the density logging tool (resolution 0.01 g / cm 3 ) Measure the drilling fluid density ρ;

[0130] Call the calibration chart, query the compensation coefficient based on D and ρ, and correct the gamma count rate:

[0131] N corr =N raw ×K;

[0132] Three-dimensional coupled environmental interference field integration:

[0133] The gamma field after terrain correction, the daughter correction field after hydrological correction, and the signal compensation field after borehole correction are interpolated in the spatiotemporal dimension through Gaussian process regression (GPR) to generate a continuous environmental interference field (spatial resolution 1m, temporal resolution 1min);

[0134] A verification point is set above a known uranium ore body, and the gamma count rate before and after correction is compared with the actual uranium content (obtained through core analysis). The error is required to be less than 10%. If the error exceeds the standard, the GPR kernel function parameters (such as length scale and variance) are adjusted or additional monitoring data (such as increasing the LiDAR scanning frequency) are supplemented to regenerate the interference field.

[0135] In one embodiment, the use of the random forest algorithm to establish a data-driven prediction model includes the following steps:

[0136] Data preparation:

[0137] Integrate core analysis data, well log curves (such as resistivity, natural gamma, density), and geochemical data to build a multi-dimensional geological parameter database; handle missing values ​​(such as interpolation) and outliers (based on statistical methods or business understanding to eliminate), and perform normalization (such as Z-score standardization) to ensure consistent feature dimensions; divide the data into training and test sets in a ratio (such as 7:3) to ensure representative sample distribution;

[0138] Feature construction:

[0139] Through the feature importance evaluation provided by random forest (such as Gini importance or permutation-based importance), the features that contribute significantly to the prediction of the equilibrium coefficient (such as porosity, permeability, and lithology category) are screened;

[0140] Combine domain knowledge to construct interactive features (such as the product of porosity and permeability) or time series features (such as the time-varying attenuation rate of gamma signals);

[0141] Build the model:

[0142] Set key parameters for random forests, including the number of decision trees, maximum depth, and minimum number of sample splits; fit the random forest model using training data and preliminarily evaluate model performance using out-of-bag error; utilize multi-core CPUs to build decision trees in parallel to accelerate the training process;

[0143] Cross-validation:

[0144] The data is divided into 5 folds, 4 folds are used for training and 1 fold is used for validation each time, and the average RMSE and R are calculated. 2 To evaluate the stability of the model; traverse and optimize the key parameters and select the parameter combination that performs best on the validation set; terminate the training early when the performance of the validation set no longer improves to prevent overfitting.

[0145] Model Evaluation Explanation:

[0146] Calculate RMSE, MAE, R on the test set 2 The results are compared with benchmark models (such as linear regression and support vector machine), and the contribution of each feature to the prediction is displayed through bar charts to assist geological interpretation; the nonlinear relationship between key features (such as porosity) and the equilibrium coefficient is analyzed to verify the rationality of the geological mechanism.

[0147] In one embodiment, the introduction of the transfer learning network, the use of historical mining area data to train the model, and the dynamic inversion and spatiotemporal update of the new regional balance coefficient include the following steps:

[0148] Build a pre-trained model (source domain):

[0149] Collect high-quality data from historical mining areas (such as multi-year well logging records and core analysis data) to build a large-scale training set; use the random forest algorithm to train the basic model on historical data and save the model parameters (such as decision tree structure and feature importance); through feature importance analysis, identify key features that are common across regions (such as lithology and parent element content).

[0150] Transfer learning applications (target domain):

[0151] Use the pre-trained model as the initial parameters for the new regional model, freeze some underlying feature layers (such as the lithology identification layer), and adjust only the top prediction layer. Perform supervised training on the top layer of the model using a small sample of data from the new region, adjusting the weights to adapt to local geological characteristics. Combined with unlabeled data from the new region (such as real-time gamma signals), further optimize the model through self-supervised learning (such as contrastive learning).

[0152] Dynamic inversion:

[0153] Real-time measurement data from the new area (such as gamma spectrometer spatiotemporal data) is input into the migrated model to generate an initial balance coefficient prediction; based on geological movement models (such as fault activity and hydrological changes), the spatiotemporal evolution trend of the balance coefficient is predicted; combined with real-time measurement data, the prediction results are adjusted through particle resampling to reduce uncertainty; and using kriging interpolation or deep learning spatiotemporal models (such as ST-GCN) to fill in the prediction gaps in areas with sparse monitoring points and generate a continuous spatiotemporal distribution map.

[0154] In one embodiment, the data purification and anomaly enhancement using the adaptive wavelet threshold denoising algorithm and spectral shape matching enhancement technology includes the following steps:

[0155] Adaptive wavelet threshold denoising:

[0156] Select a wavelet basis (such as Daubechies-4 or Symlet-5) based on data characteristics (such as the time-frequency characteristics of the gamma spectrum) to balance time domain resolution and frequency domain localization capabilities. Set the number of wavelet decomposition layers to 5 to ensure that low-frequency components (approximate coefficients) contain geological trend information and high-frequency components (detail coefficients) contain noise and weak anomalies. For each layer of detail coefficients, calculate the signal-to-noise ratio (SNR) of the current layer:

[0157]

[0158] Where P1 is the signal power and P2 is the noise power. The signal power is estimated by the approximation coefficient, and the noise power is estimated by the median absolute deviation (MAD) of the detail coefficient. The threshold is adjusted according to the SNR. When the SNR is high (>20dB), a soft threshold is used. Preserve more details; when the SNR is low (<10dB), use a hard threshold λ = 3σ to forcibly remove obvious noise; where σ is the noise standard deviation and N is the signal length; apply a dynamic threshold to each layer of detail coefficients to remove noise components below the threshold and retain abnormal signals above the threshold; reconstruct the denoised signal through inverse wavelet transform to ensure time domain continuity and frequency domain accuracy;

[0159] Enhanced spectral shape matching:

[0160] Collect gamma ray spectrum data of typical uranium ore bodies (including 238U, 235 U. 232 Th series has 16 characteristic peaks, such as 4.4keV (Pb X-ray), 18.6keV ( 238 U), 238.6keV( 232Th)) to establish a standard spectrum library; perform energy calibration and resolution correction on the spectrum library data to ensure that the characteristic peak position accuracy is ≤0.5keV; perform smoothing on the denoised gamma energy spectrum (such as Savitzky-Golay filtering) to suppress statistical fluctuation noise; calculate the correlation coefficient (Pearson or Spearman) between the spectrum to be measured and the standard spectrum, and identify the standard spectrum with the highest matching degree (such as the 238U dominant spectrum); based on the matching results, adjust the amplitude and baseline of the spectrum to be measured to align it with the standard spectrum; for characteristic peak areas with a matching degree greater than 0.8, use linear amplification (such as 2-5 times) to highlight weak abnormal signals while suppressing noise in non-matching areas.

[0161] In one embodiment, the method of combining a machine learning model to construct a three-dimensional geological model and continuously optimizing the accuracy through transfer learning includes the following steps:

[0162] The machine learning model builds an initial 3D model:

[0163] De-noised gamma field data, environmental interference field data, core analysis data, and well logging curves were integrated to construct a multi-source feature matrix (such as gamma count rate, terrain correction index, and porosity). Continuous features (such as gamma values) were logarithmically transformed, and categorical features (such as lithology) were one-hot encoded to eliminate dimensional differences. A convolutional neural network (CNN) was selected as the basic model, with input being 2D profile data (such as gamma field slices) and output being 3D geological parameters (such as uranium content and lithology category).

[0164] The network structure is set as follows: convolution layer (3×3 convolution kernel, step size 1) + batch normalization layer + ReLU activation function × 4 layers, and finally a fully connected layer to output the prediction results; Adam optimizer is used, the learning rate is initially set to 0.001, decayed by 5% every 10 rounds, the training cycle is 50 rounds, and the batch size is 32; the prediction accuracy (such as R 2 =0.85, RMSE=1.2ppm), and the model's ability to identify faults and lithologic interfaces was checked through 3D visualization;

[0165] Transfer learning continuously optimizes accuracy:

[0166] Use high-quality data from historical mining areas (such as multi-year exploration data and high-precision well logging records) to train a general model, preserving the network weights and feature extraction layer parameters. Use the convolutional layer weights of the pre-trained model as the initial parameters of the new model, freeze the first two layers (to extract general geological features, such as fault patterns), and only adjust the last two layers (to adapt to the special geological phenomena in the new area).

[0167] Fine-tuning strategy:

[0168] Supervised fine-tuning: The top layer of the model is trained using a small sample of annotated data from the new area (e.g., drill hole data) with a learning rate of 0.0001 to avoid overfitting.

[0169] Unsupervised adaptation: Incorporating unlabeled data from new regions (e.g., real-time gamma fields) and optimizing feature representation through self-supervised learning (e.g., contrastive learning) can improve the model’s sensitivity to the geological characteristics of the new region.

[0170] Dynamic Updates:

[0171] Real-time measurement data from the new area is input into the migrated model to generate an initial three-dimensional geological model. Based on geological movement models (such as fault activity and hydrological changes), the spatiotemporal evolution trends of geological parameters are predicted, and the prediction results are adjusted through particle resampling combined with real-time data. As data from new areas accumulates, new data is regularly added to the training set, and model parameters are updated through incremental learning, forming a closed loop of "prediction-verification-correction".

[0172] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A method for processing radioactivity measurement data to improve uranium exploration efficiency, characterized in that: include: Standardized data acquisition: Deploy a multi-parameter synchronous measurement module, integrating a gamma spectrometer, micro-meteorological station, ground-penetrating radar, and fiber-optic gyroscope to synchronously collect geological, geophysical, and environmental parameters in both time and space. Through standardized equipment configuration and data protocols, adjust the alignment of raw data in time and space, providing high-quality input for subsequent 3D modeling. Modeling the environmental interference field: A three-dimensional coupled geological-geophysical-environmental correction model is constructed, and multi-source interference is processed in modules. Terrain correction uses DEM data acquired by airborne LiDAR to calculate the terrain correction index and correct for gamma signal attenuation. Hydrological correction uses distributed fiber-optic thermo-permeameters to monitor groundwater flow, establish a hydrological-radioactivity coupling model, and correct for daughter retention errors. Borehole correction develops a drilling fluid absorption correction chart, combining wellbore-density parameters to dynamically compensate for signal attenuation, ultimately forming a spatiotemporally continuous environmental interference field. This is then input into a random forest model to correct for environmental biases in the balance coefficient prediction. Dynamic balance coefficient inversion: Constructs a mapping relationship between the balance coefficient and geological parameters such as lithology, porosity, and permeability. A data-driven prediction model is established using a random forest algorithm. Input parameters include core analysis data, well logging curves, and geochemical data. A transfer learning network is introduced, and the model is trained using historical mining data to perform dynamic inversion and spatiotemporal updates of the balance coefficient in new areas. Intelligent data processing and 3D visualization: Data purification and anomaly enhancement are performed through an adaptive wavelet threshold denoising algorithm and spectral shape matching enhancement technology. Subsequently, a 3D geological model is constructed by combining machine learning models, and accuracy is continuously optimized through transfer learning. Finally, a WebGL interactive platform is developed, integrating multi-parameter collaborative display, virtual drilling, profile cutting, and UMAP dimensionality reduction visualization functions to intuitively interpret high-dimensional data and quickly identify anomaly patterns, and reversely guide the data acquisition module to adjust measurement parameters.

2. The method for processing radioactivity measurement data for improving uranium mine exploration efficiency according to claim 1, characterized in that: The time-space synchronous acquisition of geological, geophysical and environmental parameters comprises the following steps: Device deployment: Gamma spectrometer: Equipped with dual probes, it simultaneously measures the total gamma count rate and energy spectrum data, with a sampling interval of 1 second to improve the high temporal resolution of radioactive signals. Micrometeorological station: Integrated with temperature, humidity, air pressure, and wind speed sensors, its sampling frequency is synchronized with the gamma spectrometer to record the impact of environmental parameters on radioactive measurements. Ground penetrating radar: With a detection range of 50m, it uses a stepped-frequency continuous wave mode, triggering a scan every 0.5m along the survey line to obtain a two-dimensional profile including rock interfaces and underground geological structures in fault zones. Fiber optic gyroscope: Real-time recording of the instrument's three-dimensional attitude angle, with a sampling interval of 1 second, to correct for changes in the gamma probe's effective area caused by instrument tilt. Synchronous trigger: All devices are aligned via GPS timestamps, using PPS signals to synchronize sampling clocks. RTK-GPS is used for spatial positioning, achieving a horizontal accuracy of less than 2cm and a vertical accuracy of less than 5cm. This is forcibly associated with the GPR line position for unified spatial coordinates. Multi-parameter fusion recording: The data acquisition software packages the gamma spectrum, meteorological parameters, ground penetrating radar images, and attitude angle data in time series to generate a structured data file containing device ID, sampling time, latitude and longitude, and elevation metadata.

3. The method for processing radioactivity measurement data for improving uranium exploration efficiency according to claim 1, characterized in that: The method of adjusting the alignment of raw data in time and space dimensions by standardizing device configuration and data protocol includes the following steps: Standardized equipment configuration: The gamma spectrometer uses a standard source for energy calibration, ensuring an energy spectrum peak accuracy of less than 1 keV. The micrometeorological station sensors are cross-calibrated in a laboratory environment. The ground-penetrating radar antenna spacing is fixed at 0.2 m to avoid depth deviations due to hardware differences. The data format is unified, with field names, units, and coordinate systems defined. The timestamp format is set to ISO 8601 in the UTC time zone, and the spatial coordinates use the UTM projection. Aligning time and space dimensions: For devices with inconsistent sampling frequencies, all parameters are unified to a 1-second time resolution through linear interpolation. Time offset correction is performed on lag parameters based on device response time to improve multi-parameter synchronization. The RTK-GPS latitude and longitude coordinates are converted to the local coordinate system of the survey area to match the GPR line position. Kriging interpolation is performed on discrete sampling points to generate a regular grid with a spatial resolution of 1m×1m, which is then fused with the GPR continuous profile data. Quality Control: The Z-score method is used to identify abnormal data after temporal and spatial alignment, mark and eliminate interference, and compare the total count rate of the gamma spectrometer with the rock layer thickness revealed by the ground penetrating radar. If the difference exceeds 15%, the recalibration process is triggered.

4. The method for processing radioactivity measurement data for improving uranium exploration efficiency according to claim 1, characterized in that: The construction of the geological-geophysical-environmental three-dimensional coupled correction model and the processing of multi-source interference in modules include the following steps: Terrain Correction: Deploy a drone or helicopter equipped with LiDAR, plan a route along the survey area, emit laser pulses, obtain raw point cloud data, denoise and filter the point cloud data, and generate a high-precision digital elevation model (DEM). Extract the terrain slope θ from the DEM: Where Δx and Δy are the horizontal position changes, and Δz is the vertical position change rate; Measuring instrument height H, combined with TCI = (1 + tanθ) 2 / H calculates the terrain correction index (TCI), which quantifies the effect of terrain undulation on gamma-ray scattering. The steeper the slope and the lower the instrument, the more significant the scattering attenuation. According to the experimental calibration results, an exponential relationship model between TCI and gamma count rate decay is established: N corr =N raw ×e 0.3×TCI Among them, N raw is the raw count rate, N corr is the corrected count rate; the model is applied to each sampling point in the survey area to generate gamma field data after terrain correction; Hydrological correction: Distributed fiber optic sensors are buried along the survey line, integrating temperature and seepage velocity monitoring functions. Using Raman scattering and Brillouin scattering techniques, they acquire real-time underground temperature and seepage velocity data with a time resolution of 1 minute. Combining these temperature and seepage velocity data, they invert groundwater flow direction and velocity using a two-dimensional water flow equation, measuring the migration velocity of uranium decay chain daughters in groundwater. The calculated daughter migration velocity is linearly related to groundwater flow velocity. Establishing a hydrological-radiological coupling model: ΔN=N1×(1-e -λ ×t) Where ΔN is the gamma count rate deviation caused by daughter retention, N1 is the gamma count rate of the parent nuclide, λ is the decay constant of the daughter nuclide, and t is the time it takes for the daughter to separate from the parent nuclide. The gamma count rate deviation ΔN caused by daughter retention is calculated according to the model and the original data is corrected: N corr =N raw -ΔN; Wellbore Correction: The laboratory simulates different borehole conditions: prepares drilling fluid containing barite and clay, and measures its absorption coefficient of gamma rays; calculates and establishes the relationship between absorption rate and drilling fluid composition; makes a calibration chart: inputs well diameter and drilling fluid density, and uses the formula K=e μ×D / ρ Calculate the signal attenuation compensation, where K is the compensation coefficient, μ is the absorption coefficient, D is the input wellbore diameter, and ρ is the drilling fluid density. During logging, the wellbore diameter D is measured in real time using a caliper, and the drilling fluid density ρ is measured using a density logging tool. Call the calibration chart, query the compensation coefficient based on D and ρ, and correct the gamma count rate: N corr =N raw ×K; Three-dimensional coupled environmental interference field integration: The gamma field after terrain correction, the daughter correction field after hydrological correction, and the signal compensation field after borehole correction are interpolated in the time and space dimensions through Gaussian process regression to generate a continuous environmental interference field. A verification point is set above the known uranium ore body to compare the gamma counting rate before and after correction with the actual uranium content, with the error required to be less than 10%. If the error exceeds the standard, the GPR kernel function parameters are adjusted or the monitoring data is supplemented to regenerate the interference field.

5. The method for processing radioactivity measurement data for improving uranium exploration efficiency according to claim 1, characterized in that: The random forest algorithm is used to establish a data-driven prediction model, including the following steps: Data preparation: Integrate core analysis data, well log curves including resistivity, natural gamma, density, and geochemical data to build a multi-dimensional geological parameter database; handle missing values ​​and outliers, and perform normalization to ensure consistent feature dimensions; divide the data into training and test sets proportionally to improve the representativeness of the sample distribution; Feature construction: Through the feature importance evaluation provided by random forest, features including porosity, permeability, and lithology that contribute significantly to the prediction of the equilibrium coefficient are screened; interactive features or time series features are constructed in combination with domain knowledge; Build the model: Set key parameters for random forests, including the number of decision trees, maximum depth, and minimum number of sample splits; fit the random forest model using training data and preliminarily evaluate model performance using out-of-bag error; utilize multi-core CPUs to build decision trees in parallel to accelerate the training process; Cross-validation: The data is divided into 5 folds, 4 folds are used for training and 1 fold is used for validation each time, and the average RMSE and R are calculated. 2 To evaluate model stability; traverse and optimize key parameters and select the parameter combination that performs best on the validation set; terminate training early when the validation set performance no longer improves to prevent overfitting; Model Evaluation Explanation: Calculate RMSE, MAE, R on the test set 2 The contribution of each feature to the prediction is shown in a bar chart and compared with the benchmark model to assist geological interpretation. The nonlinear relationship between key features and the balance coefficient is analyzed to verify the rationality of the geological mechanism.

6. The method for processing radioactivity measurement data for improving uranium exploration efficiency according to claim 1, characterized in that: The introduction of the transfer learning network, the use of historical mining area data to train the model, and the dynamic inversion and spatiotemporal update of the new regional balance coefficient include the following steps: Build a pre-trained model: Collect high-quality data from historical mining areas, including years of well logging and core analysis data, to build a large-scale training set. Use the random forest algorithm to train a basic model on the historical data and save the model parameters. Through feature importance analysis, identify key features of lithology and parent element content that are common across regions. Transfer learning applications: The pre-trained model is used as the initial parameters of the model for the new region, freezing some of the underlying feature layers and adjusting only the top prediction layer. The top layer of the model is trained in a supervised manner using a small sample of data from the new region, adjusting the weights to adapt to the local geological characteristics. The model is further optimized through self-supervised learning, combined with unlabeled data from the new region. Dynamic inversion: The real-time measurement data of the new area is input into the migrated model to generate an initial balance coefficient prediction; based on the geological movement model, the spatiotemporal evolution trend of the balance coefficient is predicted; combined with the real-time measurement data, the prediction results are adjusted through particle resampling to reduce uncertainty; using Kriging interpolation or deep learning spatiotemporal models, the prediction gaps in the sparse monitoring point areas are filled to generate a continuous spatiotemporal distribution map.

7. The method for processing radioactivity measurement data for improving uranium exploration efficiency according to claim 1, characterized in that: The data purification and anomaly enhancement are performed by using the adaptive wavelet threshold denoising algorithm and spectral shape matching enhancement technology, including the following steps: Adaptive wavelet threshold denoising: The wavelet basis is selected based on the data characteristics, balancing the time domain resolution and frequency domain localization capability. The number of wavelet decomposition layers is set to 5, so that the low-frequency components contain geological trend information and the high-frequency components contain noise and weak anomalies. For each layer detail coefficient, the signal-to-noise ratio of the current layer is calculated: Among them, P1 is the signal power and P2 is the noise power. The signal power is estimated by the approximate coefficient and the noise power is estimated by the median absolute deviation of the detail coefficient. The threshold is adjusted according to the SNR. When the SNR is high, the soft threshold is used. Preserve more details; when the SNR is low, use a hard threshold λ = 3σ to force the removal of obvious noise; where σ is the noise standard deviation and N is the signal length; apply a dynamic threshold to each layer of detail coefficients to remove noise components below the threshold and retain abnormal signals above the threshold; reconstruct the denoised signal through inverse wavelet transform to improve time domain continuity and frequency domain accuracy; Enhanced spectral shape matching: Gamma energy spectrum data of typical uranium ore bodies, 18.6keV, and 238.6keV were collected to establish a standard spectrum library; the spectrum library data was energy scaled and resolution calibrated to make the characteristic peak position accuracy less than 0.5keV; the denoised gamma energy spectrum was smoothed to suppress statistical fluctuation noise; the correlation coefficient between the spectrum to be measured and the standard spectrum was calculated to identify the standard spectrum with the highest matching degree; based on the matching results, the amplitude and baseline of the spectrum to be measured were adjusted to align it with the standard spectrum; for characteristic peak areas with a matching degree greater than 0.8, linear amplification was used to highlight weak abnormal signals while suppressing noise in non-matching areas.

8. The method for processing radioactivity measurement data for improving uranium exploration efficiency according to claim 1, characterized in that: The method of combining a machine learning model to construct a 3D geological model and continuously optimizing its accuracy through transfer learning includes the following steps: The machine learning model builds an initial 3D model: The denoised gamma field data, environmental interference field data, core analysis data, and well logging curves were integrated to construct a multi-source feature matrix. Continuous features were logarithmically transformed, and categorical features were one-hot encoded to eliminate dimensional differences. A convolutional neural network was selected as the basic model, with two-dimensional profile data as input and three-dimensional geological parameters as output. The network structure was set up as follows: a convolutional layer, a batch normalization layer, and a ReLU activation function, followed by a fully connected layer to output the prediction results. The Adam optimizer was used, with an initial learning rate of 0.001, decaying by 5% every 10 epochs, a training cycle of 50 epochs, and a batch size of 32. The prediction accuracy was calculated on the validation set, and the model's ability to identify faults and lithologic interfaces was verified through 3D visualization. Transfer learning continuously optimizes accuracy: Use high-quality data from historical mining areas to train a general model, preserving network weights and feature extraction layer parameters. Use the convolutional layer weights of the pre-trained model as the initial parameters of the new model, freeze the first two layers, and only adjust the last two layers. Fine-tuning strategy: Supervised fine-tuning: The top layer of the model is trained using a small sample of labeled data from the new region, with a learning rate of 0.0001 to avoid overfitting. Unsupervised adaptation: Incorporating unlabeled data from new areas, optimizing feature representation through self-supervised learning, and improving the model’s sensitivity to the geological characteristics of the new area. Dynamic Updates: Real-time measurement data from the new area is input into the migrated model to generate an initial 3D geological model. Based on the geological movement model, the temporal and spatial evolution trends of geological parameters are predicted, and the prediction results are adjusted through particle resampling combined with real-time data. As data from new areas accumulates, new data is regularly added to the training set, and model parameters are updated through incremental learning, forming a closed loop of "prediction-verification-correction".

Images