A multi-element geochemical anomaly intelligent identification system based on convolutional autoencoder

Abstract

The application provides a multi-element geochemical anomaly intelligent identification system based on a convolutional autoencoder, and relates to the technical field of geochemical data intelligent processing and anomaly identification. The multi-element geochemical anomaly intelligent identification system based on the convolutional autoencoder comprises a data acquisition and preprocessing module, which is used for acquiring original multi-element geochemical data, and generating a standardized geochemical data set through data cleaning, spatial coordinate verification and data format standardization processing. Local background statistics are calculated through spatial neighborhood statistics, and a local adaptive normalization method is applied to effectively eliminate the influence of spatial non-stationarity of geochemical data, significantly improve the identification accuracy of local weak anomalies, and simultaneously capture spatial features of different scales by using a multi-branch convolutional neural network structure combined with a conventional convolutional layer and a hollow convolutional layer.

Description

Technical Field

[0001] This invention relates to the field of intelligent processing and anomaly recognition technology for geochemical data, specifically to an intelligent recognition system for multi-element geochemical anomalies based on a convolutional autoencoder. Background Technology

[0002] The field of intelligent processing and anomaly identification of geochemical data focuses on using artificial intelligence, machine learning, and big data analytics to automate the processing, pattern recognition, and anomaly mining of geochemical data. Its core lies in overcoming the limitations of traditional geochemical anomaly identification methods (such as multivariate statistics and geological cumulative indices) when dealing with massive, high-dimensional, and spatially non-stationary data. Through intelligent algorithms, it automatically learns the complex relationships between elements, spatial distribution patterns, and boundary features between the background and anomalies, enabling the extraction of weak or hidden mineralization anomaly signals from complex background noise. This provides high-precision and high-efficiency decision support for mineral exploration, environmental assessment, and geological process research. One such system is a multi-element geochemical anomaly intelligent identification system based on a convolutional autoencoder. This system is specifically designed for processing multi-element geochemical data. Using a convolutional autoencoder as its core architecture, it integrates spatial perception normalization, multi-scale feature encoding, elemental cross-correlation attention mechanisms, and interpretable parsing modules to achieve multi-element collaborative anomaly detection and causal analysis.

[0003] Traditional geochemical anomaly identification methods rely on global statistical assumptions, neglecting local features within a spatially heterogeneous background. This leads to weak and gradual anomalies being masked by strong background noise. For example, multivariate statistical methods assume data follows a global distribution, making it difficult to adapt to the spatial non-stationarity of complex geological environments, resulting in missed detections of local mineralization anomalies. Elemental combination analysis heavily relies on expert experience and presuppositions, lacking adaptive modeling of complex inter-element relationships, potentially overlooking the indicative significance of atypical symbiotic combinations. Existing deep learning models are mostly black-box structures, resulting in a lack of interpretability for anomaly identification results. Geologists struggle to verify the reliability of the results, reducing the credibility of exploration decisions. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a multi-element geochemical anomaly intelligent identification system based on a convolutional autoencoder. This system solves the problem that traditional geochemical anomaly identification methods rely on global statistical assumptions and ignore local features of spatially heterogeneous backgrounds, causing weak and slow anomalies to be submerged in strong background noise. For example, multivariate statistical methods assume that the data follows a global distribution, which is difficult to adapt to the spatial non-stationarity of complex geological environments, resulting in the missed detection of local mineralization anomalies.

[0005] To achieve the above objectives, the present invention is implemented through the following technical solution: a multi-element geochemical anomaly intelligent identification system based on a convolutional autoencoder, comprising the following modules: a data acquisition and preprocessing module, used to acquire raw multi-element geochemical data, and generate a standardized geochemical dataset through data cleaning, spatial coordinate verification and data format standardization processing;

[0006] The spatial sensing normalization module is used to calculate the local background statistics of each element based on the standardized geochemical dataset through spatial neighborhood statistics, and to process them by applying a local adaptive normalization method to generate spatially normalized geochemical data.

[0007] The multi-scale feature encoding module is used to extract spatial features at different scales based on the spatially normalized geochemical data through a multi-branch convolutional neural network structure, using conventional convolutional layers and dilated convolutional layers respectively, and then fuse them to generate multi-scale encoded features.

[0008] The cross-correlation attention fusion module is used to generate channel attention weights by calculating the cross-correlation matrix between feature channels based on the multi-scale encoded features, and to perform weighted fusion of features to enhance the correlation information between elements and generate element correlation enhanced features.

[0009] The collaborative anomaly reconstruction module is used to reconstruct the input data through a decoder network based on the element correlation enhancement features, calculate the difference between the reconstructed data and the original data, and generate reconstructed geochemical data and the corresponding reconstruction error.

[0010] The gradient-weighted parsing module is used to calculate the feature map channel weights and generate a heatmap based on the reconstruction error and the element association enhancement features, using the gradient-weighted class activation mapping method, to reveal the spatial locations and features that significantly contribute to anomaly detection and generate a feature contribution heatmap.

[0011] The active learning training module is used to select samples with high information content based on the reconstruction error and its uncertainty estimate, submit them for expert annotation, and update the model weights using the newly annotated samples to generate an expert annotation dataset and updated model weights.

[0012] The anomaly contribution mapping module is used to generate a spatial distribution map showing the relative importance of each element in the anomaly region by analyzing the correlation between high contribution feature channels and original elements based on the feature contribution heatmap, and to generate a spatial distribution map of element contribution.

[0013] The comprehensive results output module is used to integrate the reconstruction error, the spatial distribution map of element contribution, and the significant anomalous areas based on dynamic threshold segmentation. Through multi-dimensional data fusion and visualization technology, it generates a comprehensive report and map containing information on the location, intensity, and dominant elements of the anomalous anomalies, as well as a comprehensive identification report and spatial distribution map of multi-element geochemical anomalies.

[0014] Preferably, the data acquisition and preprocessing module includes a data cleaning submodule, a coordinate verification submodule, and a format standardization submodule;

[0015] The data cleaning submodule, based on the original multi-element geochemical data, uses missing value imputation algorithm and outlier identification and processing algorithm to process the data and obtain cleaned geochemical data.

[0016] The coordinate verification submodule, based on the cleaned geochemical data, uses a spatial topological relationship verification method to obtain coordinate-verified data;

[0017] The format standardization submodule, based on the coordinate-verified data, uses a data format conversion and structuring process to obtain a standardized geochemical dataset.

[0018] Preferably, the spatial perception normalization module includes a neighborhood statistics submodule, a local normalization submodule, and a spatial stability submodule;

[0019] The neighborhood statistics submodule, based on the standardized geochemical dataset, uses a spatial neighborhood statistics algorithm to obtain local background statistics;

[0020] The local normalization submodule, based on the standardized geochemical dataset and local background statistics, uses a local adaptive normalization method to obtain preliminary normalized data;

[0021] The space stationarity submodule, based on the aforementioned preliminary normalized data, uses a space autocorrelation analysis method to obtain space-normalized geochemical data.

[0022] Preferably, the multi-scale feature coding module includes a fine feature extraction submodule, a region feature extraction submodule, and a feature pyramid fusion submodule;

[0023] The fine feature extraction submodule, based on the spatially normalized geochemical data, uses a conventional convolutional layer network structure to obtain fine-scale features.

[0024] The regional feature extraction submodule, based on the spatially normalized geochemical data, uses a hollow convolutional layer network structure to obtain regional-scale features.

[0025] The feature pyramid fusion submodule, based on the fine-scale features and regional-scale features, uses the feature pyramid network fusion method to obtain multi-scale encoded features.

[0026] Preferably, the cross-correlation attention fusion module includes a cross-correlation calculation submodule, an attention weight generation submodule, and a feature fusion submodule;

[0027] The cross-correlation calculation submodule, based on the multi-scale coding features, uses the cross-correlation matrix calculation method between feature channels to obtain the channel cross-correlation matrix;

[0028] The attention weight generation submodule obtains the channel attention weights based on the channel cross-correlation matrix and using an attention weight generation algorithm.

[0029] The feature fusion submodule, based on the multi-scale encoded features and channel attention weights, uses a weighted feature fusion algorithm to obtain element association enhancement features.

[0030] Preferably, the collaborative anomaly reconstruction module includes a decoder submodule, a reconstruction calculation submodule, and an error measurement submodule;

[0031] The decoder submodule, based on the element association enhancement features, uses a decoder network structure to obtain preliminary reconstructed data;

[0032] The reconstruction calculation submodule, based on the spatially normalized geochemical data and the preliminary reconstruction data, uses an element-level difference calculation method to obtain the elemental reconstruction error.

[0033] The error measurement submodule reconstructs the error based on the elements and uses an error aggregation method to obtain the reconstruction error.

[0034] Preferably, the gradient weighted parsing module includes a gradient calculation submodule, a channel weight pooling submodule, and a heatmap generation submodule;

[0035] The gradient calculation submodule, based on the reconstruction error and element correlation enhancement features, uses the gradient backpropagation algorithm to obtain the feature gradient map;

[0036] The channel weight pooling submodule, based on the feature gradient map, uses a global average pooling method to obtain the channel contribution weights;

[0037] The heatmap generation submodule, based on the channel contribution weights and element association enhancement features, uses a weighted class activation mapping method to obtain a feature contribution heatmap.

[0038] Preferably, the active learning training module includes an uncertainty estimation submodule, a sample selection submodule, an expert annotation interface submodule, and a model fine-tuning submodule;

[0039] The uncertainty estimation submodule, based on the reconstruction error, uses the Monte Carlo random deactivation method or the model ensemble method to obtain the prediction uncertainty;

[0040] The sample selection submodule, based on the reconstruction error and prediction uncertainty, adopts an uncertainty-based sample selection strategy to obtain a candidate labeled sample set;

[0041] The expert annotation interface submodule obtains expert annotation information based on the candidate annotation sample set using a human-computer interaction interface;

[0042] The model fine-tuning submodule, based on standardized geochemical datasets, expert annotation information, and existing model weights, employs a transfer learning fine-tuning strategy to obtain updated model weights.

[0043] Preferably, the anomaly contribution mapping module includes a feature channel analysis submodule, an element association submodule, and a contribution mapping submodule;

[0044] The feature channel analysis submodule, based on the feature contribution heatmap, uses a feature channel importance analysis method to obtain high-contribution feature channel identifiers;

[0045] The element association submodule, based on the high-contribution feature channel identifier, uses a feature-to-element mapping method to obtain the dominant element identifier;

[0046] The contribution mapping submodule, based on the dominant element identifier and feature contribution heatmap, uses spatial mapping technology to obtain a spatial distribution map of element contribution.

[0047] Preferably, the comprehensive result output module includes an anomaly threshold segmentation submodule, a multi-dimensional fusion submodule, and a visualization report generation submodule;

[0048] The anomaly threshold segmentation submodule, based on the reconstruction error, uses a dynamic threshold determination method to obtain significant anomaly region identifiers;

[0049] The multidimensional fusion submodule, based on the reconstruction error, the spatial distribution map of element contribution, and the identification of significant abnormal regions, uses data fusion technology to obtain comprehensive abnormal feature data;

[0050] The visualization report generation submodule, based on the comprehensive anomaly feature data, uses geographic information system visualization and report generation technology to obtain a comprehensive identification report of multi-element geochemical anomalies and an anomaly spatial distribution map.

[0051] This invention provides an intelligent identification system for multi-element geochemical anomalies based on a convolutional autoencoder. It has the following advantages:

[0052] This invention effectively eliminates the spatial non-stationarity of geochemical data by calculating local background statistics through spatial neighborhood statistics and applying a local adaptive normalization method, significantly improving the accuracy of local weak anomaly identification. Utilizing a multi-branch convolutional neural network structure, combined with conventional and dilated convolutional layers, it simultaneously captures spatial features at different scales, enhancing the model's ability to perceive regional background trends and local details, and improving the detection sensitivity of weak or hidden mineralization signals. It introduces a cross-correlation matrix calculation and weighted fusion mechanism between feature channels to explicitly model element symbiotic antagonistic relationships, making the anomaly identification results more geochemically indicative. Based on backpropagation of the reconstruction error gradient, it generates heat maps, enabling visual analysis of anomaly causes and providing element-level contribution distributions to reduce the subjectivity of manual interpretation. It employs a Monte Carlo random inactivation and uncertainty-driven sample selection strategy to optimize expert annotation resource allocation, improving model iteration efficiency and generalization ability. Dynamic threshold segmentation and multi-dimensional data fusion technology generate comprehensive reports, integrating anomaly location, intensity, and dominant element information, providing high-precision decision support for mineral target area delineation. Attached Figure Description

[0053] Figure 1 This is a system block diagram of the present invention;

[0054] Figure 2 This is a schematic diagram of the data acquisition and preprocessing module of the present invention;

[0055] Figure 3 This is a schematic diagram of the spatial perception normalization module of the present invention;

[0056] Figure 4 This is a schematic diagram of the multi-scale feature encoding module of the present invention;

[0057] Figure 5 This is a schematic diagram of the cross-correlation attention fusion module of the present invention;

[0058] Figure 6 This is a schematic diagram of the collaborative anomaly reconstruction module of the present invention;

[0059] Figure 7 This is a schematic diagram of the gradient weighted analysis module of the present invention;

[0060] Figure 8 This is a schematic diagram of the active learning training module of the present invention;

[0061] Figure 9 This is a schematic diagram of the abnormal contribution mapping module of the present invention;

[0062] Figure 10 This is a schematic diagram of the comprehensive result output module of the present invention. Detailed Implementation

[0063] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0064] Example:

[0065] like Figure 1-10 As shown, this embodiment of the invention provides a multi-element geochemical anomaly intelligent identification system based on a convolutional autoencoder, including the following modules: a data acquisition and preprocessing module, used to acquire raw multi-element geochemical data, and generate a standardized geochemical dataset through data cleaning, spatial coordinate verification and data format standardization processing;

[0066] The spatial sensing normalization module is used to calculate the local background statistics of each element based on the standardized geochemical dataset through spatial neighborhood statistics, and to process them by applying a local adaptive normalization method to generate spatially normalized geochemical data.

[0067] The multi-scale feature encoding module is used to extract spatial features at different scales based on the spatially normalized geochemical data through a multi-branch convolutional neural network structure, using conventional convolutional layers and dilated convolutional layers respectively, and then fuse them to generate multi-scale encoded features.

[0068] The cross-correlation attention fusion module is used to generate channel attention weights by calculating the cross-correlation matrix between feature channels based on the multi-scale encoded features, and to perform weighted fusion of features to enhance the correlation information between elements and generate element correlation enhanced features.

[0069] The collaborative anomaly reconstruction module is used to reconstruct the input data through a decoder network based on the element correlation enhancement features, calculate the difference between the reconstructed data and the original data, and generate reconstructed geochemical data and the corresponding reconstruction error.

[0070] The gradient-weighted parsing module is used to calculate the feature map channel weights and generate a heatmap based on the reconstruction error and the element association enhancement features, using the gradient-weighted class activation mapping method, to reveal the spatial locations and features that significantly contribute to anomaly detection and generate a feature contribution heatmap.

[0071] The active learning training module is used to select samples with high information content based on the reconstruction error and its uncertainty estimate, submit them for expert annotation, and update the model weights using the newly annotated samples to generate an expert annotation dataset and updated model weights.

[0072] The anomaly contribution mapping module is used to generate a spatial distribution map showing the relative importance of each element in the anomaly region by analyzing the correlation between high contribution feature channels and original elements based on the feature contribution heatmap, and to generate a spatial distribution map of element contribution.

[0073] The comprehensive results output module is used to integrate the reconstruction error, the spatial distribution map of element contribution, and the significant anomalous areas based on dynamic threshold segmentation. Through multi-dimensional data fusion and visualization technology, it generates a comprehensive report and map containing information on the location, intensity, and dominant elements of the anomalous anomalies, as well as a comprehensive identification report and spatial distribution map of multi-element geochemical anomalies.

[0074] The data acquisition and preprocessing module includes a data cleaning submodule, a coordinate verification submodule, and a format standardization submodule;

[0075] The data cleaning submodule, based on the original multi-element geochemical data, uses missing value imputation algorithm and outlier identification and processing algorithm to process the data and obtain cleaned geochemical data.

[0076] Based on raw multi-element geochemical data, field soil samples containing elements such as lead, zinc, and copper were collected. Some sampling points had missing or abnormally high lead values. A mean imputation algorithm was used to handle the missing values: the background mean lead value within the sample area was calculated. The missing points are assigned the mean; outliers are identified using box plots; the first quartile of lead content is calculated. Third and quartiles Interquartile range Set the outlier threshold upper limit to Sampling points with lead content exceeding this threshold (such as point P01) will be sampled. Values ​​marked as outliers are replaced with the median of their neighborhood (replaced with the median lead content of the five nearest neighboring sampling points). After data cleaning is completed, cleaned geochemical data is generated.

[0077] The coordinate verification submodule, based on the cleaned geochemical data, uses a spatial topological relationship verification method to obtain coordinate-verified data;

[0078] Based on the cleaned geochemical data, a set of spatial coordinates of the sampling points was extracted. Using the spatial topology analysis tool of GIS software, with a tolerance distance of 50 meters, the distance between adjacent sampling points was checked: the Euclidean distance between point A (100,200) and point B (98,195) was calculated. If the distance between point C (300, 400) and its nearest neighbor D is 82.5 meters, exceeding the tolerance, a verification alarm will be triggered. The coordinates of point C will be corrected to (305, 398) according to the field logbook. The distance between point C and its nearest neighbor D after correction will be... Meters, pass the verification; after traversing all points, generate the verified coordinate data.

[0079] The format standardization submodule, based on the coordinate-verified data, uses a data format conversion and structuring process to obtain a standardized geochemical dataset.

[0080] Based on the coordinate-verified data, longitude, latitude, elevation, and element content fields are extracted to establish a unified data dictionary: the field names are defined as follows. The Python Pandas library was used to convert the Excel spreadsheet into a three-dimensional tensor. Longitude values ​​(e.g., 116.25°E-116.30°E) were used as the X-axis index, latitude values ​​(e.g., 39.85°N-39.90°N) were used as the Y-axis index, and element types (Pb, Zn, Cu, etc.) were used as the channel dimensions. Missing grid points were filled with 0s by interpolation to generate a standard tensor structure with a size of 100×100×8. The format was standardized to generate a standardized geochemical dataset.

[0081] The spatial perception normalization module includes a neighborhood statistics submodule, a local normalization submodule, and a spatial stability submodule.

[0082] The neighborhood statistics submodule, based on the standardized geochemical dataset, uses a spatial neighborhood statistics algorithm to obtain local background statistics;

[0083] Based on a standardized geochemical dataset, the spatial coordinates and lead content values ​​of sampling points were extracted. A circular neighborhood definition method was adopted, with a radius of 300 meters. Centered on point P (longitude 116.25°, latitude 39.86°), all sampling points within the radius (e.g., points A, B, C, etc., a total of 50 points) were selected using a KD-tree spatial index. The arithmetic mean of the lead content was calculated: summing the lead content values ​​of all points (e.g., 15.2 + 16.1 + ... + 14.8 = 780 μg / g), and dividing by the number of points (50) yielded a mean of 15. 0.6μg / g, calculate the standard deviation: calculate the sum of squares of the differences between the content of each point and the mean, such as (15.2-15.6)²+(16.1-15.6)²+…=24.8), divide by the number of points minus 1 (49) and take the square root to get 0.71μg / g, and calculate the median at the same time: sort the 50 values ​​and take the average of the 25th and 26th values ​​(15.5μg / g), and finally generate the local background statistics of point P (mean 15.6μg / g, standard deviation 0.71μg / g, median 15.5μg / g).

[0084] The local normalization submodule, based on the standardized geochemical dataset and local background statistics, uses a local adaptive normalization method to obtain preliminary normalized data;

[0085] Based on the standardized geochemical dataset and the local background statistics of point P, the original lead content value of point P was extracted as 18.3 μg / g. The local Z-score normalization method was used: (original value - local mean) / local standard deviation = (18.3 - 15.6) / 0.71 ≈ 3.80. If the original lead content of point Q is 25.6 μg / g, its normalized value is (25.6 - 15.6) / 0.71 ≈ 14.08. The same operation was performed on all sampling points to generate a normalized value sequence of lead (such as point P: 3.80, point Q: 14.08, etc.), thus obtaining preliminary normalized data.

[0086] The space stationarity submodule, based on the aforementioned preliminary normalized data, uses a space autocorrelation analysis method to obtain space-normalized geochemical data.

[0087] Based on the preliminary normalized data, the normalized value sequence of lead and its corresponding coordinates were extracted. The Moran's I index was used to verify spatial autocorrelation: a spatial weight matrix was constructed, with a distance threshold of 500 meters. If the distance between two points is ≤500 meters, the weight is 1 (e.g., the distance between point P and point A is 320 meters, weight 1), otherwise it is 0. The global Moran's I value was calculated: the numerator is the weighted sum of the products of all points with respect to the normalized values ​​(e.g., point P value 3.80 × point A value 2.15 × weight 1), and the denominator is the total number of weights. After standardization, I = 0.15 (range -1 to 1). If I > 0.1, it is considered a weak positive correlation, indicating that spatial nonstationarity has been significantly reduced (original data I = 0.43), generating spatially normalized geochemical data that meets the requirements of spatial stationarity.

[0088] The multi-scale feature coding module includes a fine feature extraction submodule, a region feature extraction submodule, and a feature pyramid fusion submodule;

[0089] The fine feature extraction submodule, based on the spatially normalized geochemical data, uses a conventional convolutional layer network structure to obtain fine-scale features.

[0090] Based on spatially normalized geochemical data, gridded data (100×100 matrix) of lead element were extracted in the range of longitude 116.25°-116.30° and latitude 39.85°-39.90°. Feature extraction was performed using a 3×3 convolution kernel: the convolution kernel weight matrix was initialized (e.g., center weight 0.2, adjacent weight 0.1). Taking the point (39.86N, 116.25E) as the center, the normalized lead element values ​​of the surrounding 8 grid points were taken (e.g., top left point 2.1, top point 2.3, top right point 2.0, left point 3). (8, center point 3.80, right side point 3.5, lower left point 1.9, lower point 1.8, lower right point 2.2), the convolution calculation is the sum of the values ​​of each point multiplied by their corresponding weights: 2.1×0.1+2.3×0.1+2.0×0.1+3.8×0.1+3.80×0.2+3.5×0.1+1.9×0.1+1.8×0.1+2.2×0.1=1.18. After adding a bias term of 0.01, the output feature value is 1.19. A fine-scale feature matrix of 100×100 is generated by traversing all grid points.

[0091] The regional feature extraction submodule, based on the spatially normalized geochemical data, uses a hollow convolutional layer network structure to obtain regional-scale features.

[0092] Based on the same lead element grid matrix of spatially normalized geochemical data, a hollow convolutional layer with an expansion rate of 2 (kernel size 3×3) was used: the initial weights were 0.3 at the center and 0.05 at each of the four sides. Taking the point (39.86N, 116.25E) as the center, adjacent points were skipped and the values ​​of the interval grid points were directly taken, such as the value of 1.2 for the top left (39.84N, 116.23E), 2.5 for the top (39.88N, 116.25E), 1.5 for the top right (39.84N, 116.27E), 3.0 for the left (39.86N, 116.21E), 3.80 for the center point, and 3.0 for the right (39.86N, 116.29E). The values ​​are 3.2 for the lower left (39.82N, 116.23E), 1.0 for the lower left (39.90N, 116.25E), 1.7 for the lower right (39.82N, 116.27E), and 1.3 for the lower right (39.82N, 116.27E). The convolution calculation is 1.2×0.05+2.5×0.05+1.5×0.05+3.0×0.05+3.80×0.3+3.2×0.05+1.0×0.05+1.7×0.05+1.3×0.05=1.58. With a bias of 0.02, the output feature value is 1.60. After traversal, a 50×50 region-scale feature matrix is ​​generated (the receptive field is expanded to 600 meters due to dilated convolution).

[0093] The feature pyramid fusion submodule, based on the fine-scale features and regional-scale features, uses the feature pyramid network fusion method to obtain multi-scale encoded features.

[0094] Based on the fine-scale feature matrix (size 100×100) and the regional-scale feature matrix (size 50×50), bilinear interpolation is used to upsample the fine features to 50×50: the fine matrix point (25,25) has a value of 1.19, and its corresponding upsampled position (12.5,12.5) needs to be interpolated. The weighted average of the neighboring points (12,12) values ​​1.10, (12,13) ​​values ​​1.15, (13,12) values ​​1.08, and (13,13) values ​​1.20 is calculated to be 1.13. The regional-scale matrix point (25,25) value 1.60 is added to the upsampled value 1.13 with a weight of 1:1 to obtain a fusion value of 2.73. All points are traversed to generate a 50×50 multi-scale encoded feature matrix.

[0095] The cross-correlation attention fusion module includes a cross-correlation calculation submodule, an attention weight generation submodule, and a feature fusion submodule;

[0096] The cross-correlation calculation submodule, based on the multi-scale coding features, uses the cross-correlation matrix calculation method between feature channels to obtain the channel cross-correlation matrix;

[0097] Based on multi-scale coding features, lead element feature maps (50×50 matrix) and zinc element feature maps (50×50 matrix) are extracted. Taking point (25,25) as an example, the lead feature value is 1.13 and the zinc feature value is 0.92. The cosine similarity between the two channels at this point is calculated as follows: the numerator is the sum of the products of all spatial position feature values ​​of the two channels (lead map point (1,1) value 1.10 × zinc map point (1,1) value 0.95 + … + lead map point (50,50) value 1.20 × zinc map point (50,50) value 0.88), and the denominator is the magnitude of the lead map feature vector. With the eigenvector magnitude of the zinc map The product of the two is calculated to have a similarity of 0.75. An 8×8 channel cross-correlation matrix is ​​generated by traversing all channel pairs (such as lead-copper, zinc-copper, etc.) (where the lead-zinc similarity of 0.75 is located in the 1st row and 2nd column).

[0098] The attention weight generation submodule obtains the channel attention weights based on the channel cross-correlation matrix and using an attention weight generation algorithm.

[0099] Based on the channel cross-correlation matrix, the similarity vector between the lead channel and other channels is extracted ([lead-lead: 1.0, lead-zinc: 0.75, lead-copper: 0.62]). The attention weight is calculated using the Softmax function: an exponent is taken for each similarity value. The summation yields 6.693. The weights are the exponents divided by the sum (lead-lead weight 2.718 / 6.693≈0.42, lead-zinc weight 2.117 / 6.693≈0.32, lead-copper weight 1.858 / 6.693≈0.26). This generates the attention weight vector [0.42, 0.32, 0.26] for the lead channel. By traversing all channels, eight sets of weight vectors are generated to form the channel attention weight matrix.

[0100] The feature fusion submodule, based on the multi-scale encoded features and channel attention weights, uses a weighted feature fusion algorithm to obtain element association enhancement features.

[0101] Based on multi-scale coding features and channel attention weights, taking the spatial point (25,25) as an example, the lead feature value is extracted as 1.13, the zinc feature value as 0.92, and the copper feature value as 0.85. According to the lead channel attention weight vector [0.42,0.32,0.26], the weighted fusion value is calculated as follows: lead value × lead weight + zinc value × zinc weight + copper value × copper weight = 1.13 × 0.42 + 0.92 × 0.32 + 0.85 × 0.26 ≈ 0.475 + 0.294 + 0.221 = 0.990. A 50 × 50 element association enhancement feature matrix is ​​generated by traversing all spatial locations.

[0102] The collaborative anomaly reconstruction module includes a decoder submodule, a reconstruction calculation submodule, and an error measurement submodule.

[0103] The decoder submodule, based on the element association enhancement features, uses a decoder network structure to obtain preliminary reconstructed data;

[0104] Based on element association enhancement features, the lead element feature value of 0.990 was extracted from the spatial point (39.86N, 116.25E). A deconvolution operation was performed: the weights of the 3×3 deconvolution kernel were initialized (0.2 at the center, 0.1 on each of the surrounding edges). The neighboring points of the feature matrix were then selected with this point as the center (top left 0.95, top 0.98, top right 0.92, left 0.96, center 0.990, right 0.94, bottom left 0.93, bottom 0). 0.97 (bottom right 0.91), calculate the weighted sum 0.95×0.1+0.98×0.1+0.92×0.1+0.96×0.1+0.990×0.2+0.94×0.1+0.93×0.1+0.97×0.1+0.91×0.1=0.978, add an offset of 0.01 to output the reconstructed lead value of 0.988, and traverse all points to generate a 50×50 preliminary reconstructed lead element data matrix.

[0105] The reconstruction calculation submodule, based on the spatially normalized geochemical data and the preliminary reconstruction data, uses an element-level difference calculation method to obtain the elemental reconstruction error.

[0106] Based on spatially normalized geochemical data and preliminary reconstructed data, the original normalized lead value of point (39.86N, 116.25E) is 3.80, and the reconstructed lead value is 0.988. The elemental absolute difference is calculated as: |3.80 - 0.988| = 2.812. If the original value of point Q is 14.08 and the reconstructed value is 13.26, the difference is |14.08-13.26|=0.82. After traversal, a lead element reconstruction error matrix is ​​generated (e.g., error of point P is 2.812, error of point Q is 0.82).

[0107] The error measurement submodule reconstructs the error based on the elements and uses an error aggregation method to obtain the reconstruction error.

[0108] Based on the lead element reconstruction error matrix, the errors at point P (2.812), Q (0.82), and R (0.70) were extracted. The arithmetic mean aggregation method was used: summation (2.812 + 0.82 + 0.70 = 4.332), divided by the number of points (3) to obtain the average error of 1.444. The lead element reconstruction error sequence containing the overall anomaly degree of each point was generated by traversing all sampling points.

[0109] The gradient weighted parsing module includes a gradient calculation submodule, a channel weight pooling submodule, and a heatmap generation submodule;

[0110] The gradient calculation submodule, based on the reconstruction error and element correlation enhancement features, uses the gradient backpropagation algorithm to obtain the feature gradient map;

[0111] Based on the reconstruction error of 2.812 at point P and the lead element correlation enhancement feature value of 0.990, the gradient is approximated using the numerical difference method: keeping other features unchanged, the lead feature value is increased by a small increment of 0.01 (becoming 0.990+0.01=1.000), and the reconstruction error at point P is recalculated (e.g., increased to 2.85). The gradient value is (new error - original error) / increment = (2.85 - 2.812) / 0.01 = 3.8. The gradient value of the lead channel at point P is generated by traversing each position of the feature map, forming a 50×50 feature gradient map.

[0112] The channel weight pooling submodule, based on the feature gradient map, uses a global average pooling method to obtain the channel contribution weights;

[0113] Based on the feature gradient map of the lead channel, the gradient values ​​of all spatial locations are extracted, such as point (1,1): 0.2, point (1,2): 0.3, ..., point (50,50): 0.1. Global average pooling is used: summing all gradient values ​​(0.2+0.3+...+0.1=180.0), dividing by the total number of positions 2500, yields the average gradient value of 0.072. This value is used as the contribution weight of the lead channel to the reconstruction error. A weight vector is generated by traversing all element channels (lead weight 0.072, zinc weight 0.058, etc.).

[0114] The heatmap generation submodule, based on the channel contribution weights and element association enhancement features, uses a weighted class activation mapping method to obtain a feature contribution heatmap.

[0115] Based on the lead channel contribution weight of 0.072 and the lead element correlation enhancement eigenvalue of point P of 0.990, the thermal value is calculated as: weight × eigenvalue = 0.072 × 0.990 ≈ 0.071. The thermal value of the zinc channel is calculated simultaneously (zinc weight 0.058 × zinc eigenvalue 0.92 ≈ 0.053). The total thermal value of point P is obtained by summing the thermal values ​​of each channel: 0.071 + 0.053 + ... = 0.25. A 50×50 feature contribution heatmap is generated by traversing the spatial locations.

[0116] The active learning training module includes an uncertainty estimation submodule, a sample selection submodule, an expert annotation interface submodule, and a model fine-tuning submodule.

[0117] The uncertainty estimation submodule, based on the reconstruction error, uses the Monte Carlo random deactivation method or the model ensemble method to obtain the prediction uncertainty;

[0118] Based on the reconstruction error of 2.812 for point P, the Monte Carlo random deactivation method was used: 20% of neurons in the model's decoder layer were randomly deactivated, and the reconstruction error of predicted point P was repeated 5 times (e.g., 2.80, 2.85, 2.78, 2.82, 2.83). The standard deviation was calculated, and the mean was taken. The standard deviation √0.00068 ≈ 0.026 is used as the prediction uncertainty for point P. An uncertainty sequence is generated by traversing all sampling points.

[0119] The sample selection submodule, based on the reconstruction error and prediction uncertainty, adopts an uncertainty-based sample selection strategy to obtain a candidate labeled sample set;

[0120] Based on the reconstruction error of point P of 2.812 and its uncertainty of 0.026, an uncertainty threshold of 0.02 is set: if the uncertainty is greater than 0.02 (e.g., 0.026 for point P > 0.02), and the reconstruction error is greater than the mean error of 1.444 (2.812 > 1.444), then it is marked as a candidate sample; the top 10% of samples are selected and sorted (e.g., sorted from highest to lowest error: point P first, point Q second, etc.), and the top 5 samples are used to form a candidate labeled sample set.

[0121] The expert annotation interface submodule obtains expert annotation information based on the candidate annotation sample set using a human-computer interaction interface;

[0122] Based on the candidate annotation sample set, the coordinates (39.86N, 116.25E) and error value 2.812 of point P are extracted. Its spatial location is displayed through the GIS interface. Geological experts determine that the point is located on the periphery of the mineralization zone based on the regional geological map and mark it as "non-anomaly". The annotation results are received and stored as expert annotation information for point P.

[0123] The model fine-tuning submodule, based on standardized geochemical datasets, expert annotation information, and existing model weights, employs a transfer learning fine-tuning strategy to obtain updated model weights.

[0124] Based on the standardized geochemical dataset, the "non-anomaly" labeling information of point P, and the existing model weights, the encoder weights are frozen, and only the decoder layer is updated: with the original lead value of point P (3.80) as the supervision target, the decoder outputs a reconstructed value of 0.988, the loss is calculated as (3.80-0.988)²≈7.92, the decoder kernel weights are adjusted by backpropagation (e.g., the center weight is adjusted from 0.2 to 0.19), and the updated model weights are generated after 50 iterations of training.

[0125] The anomaly contribution mapping module includes a feature channel analysis submodule, an element association submodule, and a contribution mapping submodule;

[0126] The feature channel analysis submodule, based on the feature contribution heatmap, uses a feature channel importance analysis method to obtain high-contribution feature channel identifiers;

[0127] Based on the feature contribution heatmap, the total heat value of point P is extracted to be 0.25. The proportion of heat value of each channel is calculated: the heat value of the lead channel is 0.071, accounting for 28.4% of the total heat value of 0.25; the zinc channel is 0.053, accounting for 21.2%; and the copper channel is 0.042, accounting for 16.8%. A contribution threshold of 20% is set, and channels with a proportion ≥20% are selected (lead 28.4%>20%, zinc 21.2%>20%) and marked as high contribution feature channels (lead channel number 1, zinc channel number 2). A high contribution channel identifier sequence is generated by traversing all sampling points.

[0128] The element association submodule, based on the high-contribution feature channel identifier, uses a feature-to-element mapping method to obtain the dominant element identifier;

[0129] Based on the high contribution feature channel identifiers, extract the high contribution channel numbers 1 (lead) and 2 (zinc) for point P. Referring to the feature channel definition table (channel 1 corresponds to lead and channel 2 corresponds to zinc), map to obtain the dominant element identifiers (lead symbol Pb and zinc symbol Zn), and generate a dominant element identifier list [Pb, Zn] for point P.

[0130] The contribution mapping submodule, based on the dominant element identifier and feature contribution heatmap, uses spatial mapping technology to obtain a spatial distribution map of element contribution.

[0131] Based on the dominant element identifiers [Pb, Zn] and the thermal values ​​of the lead channel (0.071) and zinc channel (0.053) at point P, the contribution of each element is calculated as follows: lead contribution 0.071 / (0.071+0.053)≈57.3%, zinc contribution 0.053 / (0.071+0.053)≈42.7%. Centered on the coordinates of point P (39.86N, 116.25E), isosurface maps of lead contribution (57.3%) and zinc contribution (42.7%) are drawn on the GIS platform and overlaid to generate a spatial distribution map of element contribution.

[0132] The comprehensive results output module includes an anomaly threshold segmentation submodule, a multi-dimensional fusion submodule, and a visualization report generation submodule;

[0133] The anomaly threshold segmentation submodule, based on the reconstruction error, uses a dynamic threshold determination method to obtain significant anomaly region identifiers;

[0134] Based on the reconstructed error sequence (e.g., error at point P = 2.812, error at point Q = 0.82, error at point R = 0.70), calculate the global error mean: sum all error values ​​(2.812 + 0.82 + 0.70 = 4.332), divide by the total number of points 3 to get the mean 1.444, and calculate the standard deviation: calculate the sum of the squares of the differences between each error and the mean (…). Divide by the number minus 1 (2) and take the square root to get The dynamic threshold formula is set as the mean plus twice the standard deviation (1.444 + 2 × 1.185 = 3.814). All points are iterated. If the error is ≥ 3.814 (e.g., if the error of point P is 2.812 < 3.814, it is not abnormal, and if the error of point S is 4.0 > 3.814, it is marked as abnormal), a list of significant abnormal areas is generated.

[0135] The multidimensional fusion submodule, based on the reconstruction error, the spatial distribution map of element contribution, and the identification of significant abnormal regions, uses data fusion technology to obtain comprehensive abnormal feature data;

[0136] Based on the reconstruction error (value of point P: 2.812), the spatial distribution map of element contributions (point P: lead contribution 57.3%, zinc 42.7%), and the identification of significant abnormal areas (point P is marked "abnormal"), the fusion weights are set as follows: error weight 0.6, contribution weight 0.4. The comprehensive outlier value is calculated as: error value × error weight + lead contribution × contribution weight = 2.812 × 0.6 + 57.3 × 0.4 ≈ 1.687 + 22.92 = 24.607. If point Q has an error of 0.82 and a lead contribution of 30%, the comprehensive value is 0.82 × 0.6 + 30 × 0.4 = 0.492 + 12 = 12.492. A grid data matrix containing the comprehensive outlier value is generated by traversing all points.

[0137] The visualization report generation submodule, based on the comprehensive anomaly feature data, uses geographic information system visualization and report generation technology to obtain a comprehensive identification report of multi-element geochemical anomalies and an anomaly spatial distribution map.

[0138] Based on the comprehensive anomaly feature data (e.g., the comprehensive value of point P is 24.607), import it into ArcGIS software and create a point layer: define the coordinate field as longitude 116.25° and latitude 39.86°, and the value field as the comprehensive value 24.607. Set the rendering rules: comprehensive value ≥20 is red, 10-20 is yellow, and <10 is green. Generate a heat map, synchronously call the report template, insert the heat map screenshot and data table (including point coordinates, comprehensive value, and dominant elements), and export a multi-element geochemical anomaly comprehensive identification report and anomaly spatial distribution map file in PDF format.

[0139] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A multi-element geochemical anomaly intelligent identification system based on a convolutional autoencoder, characterized in that, It includes the following modules: a data acquisition and preprocessing module, which is used to acquire raw multi-element geochemical data and generate a standardized geochemical dataset through data cleaning, spatial coordinate verification and data format standardization; The spatial sensing normalization module is used to calculate the local background statistics of each element based on the standardized geochemical dataset through spatial neighborhood statistics, and to process them by applying a local adaptive normalization method to generate spatially normalized geochemical data. The multi-scale feature encoding module is used to extract spatial features at different scales based on the spatially normalized geochemical data through a multi-branch convolutional neural network structure, using conventional convolutional layers and dilated convolutional layers respectively, and then fuse them to generate multi-scale encoded features. The cross-correlation attention fusion module is used to generate channel attention weights by calculating the cross-correlation matrix between feature channels based on the multi-scale encoded features, and to perform weighted fusion of features to enhance the correlation information between elements and generate element correlation enhanced features. The collaborative anomaly reconstruction module is used to reconstruct the input data through a decoder network based on the element correlation enhancement features, calculate the difference between the reconstructed data and the original data, and generate reconstructed geochemical data and the corresponding reconstruction error. The gradient-weighted parsing module is used to calculate the feature map channel weights and generate a heatmap based on the reconstruction error and the element association enhancement features, using the gradient-weighted class activation mapping method, to reveal the spatial locations and features that significantly contribute to anomaly detection and generate a feature contribution heatmap. The active learning training module is used to select samples with high information content based on the reconstruction error and its uncertainty estimate, submit them for expert annotation, and update the model weights using the newly annotated samples to generate an expert annotation dataset and updated model weights. The anomaly contribution mapping module is used to generate a spatial distribution map showing the relative importance of each element in the anomaly region by analyzing the correlation between high contribution feature channels and original elements based on the feature contribution heatmap, and to generate a spatial distribution map of element contribution. The comprehensive results output module is used to integrate the reconstruction error, the spatial distribution map of element contribution, and the significant anomalous areas based on dynamic threshold segmentation. Through multi-dimensional data fusion and visualization technology, it generates a comprehensive report and map containing information on the location, intensity, and dominant elements of the anomalous anomalies, as well as a comprehensive identification report and spatial distribution map of multi-element geochemical anomalies.

2. The intelligent identification system for multi-element geochemical anomalies based on a convolutional autoencoder according to claim 1, characterized in that: The data acquisition and preprocessing module includes a data cleaning submodule, a coordinate verification submodule, and a format standardization submodule; The data cleaning submodule, based on the original multi-element geochemical data, uses missing value imputation algorithm and outlier identification and processing algorithm to process the data and obtain cleaned geochemical data. The coordinate verification submodule, based on the cleaned geochemical data, uses a spatial topological relationship verification method to obtain coordinate-verified data; The format standardization submodule, based on the coordinate-verified data, uses a data format conversion and structuring process to obtain a standardized geochemical dataset.

3. The intelligent identification system for multi-element geochemical anomalies based on a convolutional autoencoder according to claim 1, characterized in that: The spatial perception normalization module includes a neighborhood statistics submodule, a local normalization submodule, and a spatial stability submodule. The neighborhood statistics submodule, based on the standardized geochemical dataset, uses a spatial neighborhood statistics algorithm to obtain local background statistics; The local normalization submodule, based on the standardized geochemical dataset and local background statistics, uses a local adaptive normalization method to obtain preliminary normalized data; The space stationarity submodule, based on the aforementioned preliminary normalized data, uses a space autocorrelation analysis method to obtain space-normalized geochemical data.

4. The intelligent identification system for multi-element geochemical anomalies based on a convolutional autoencoder according to claim 1, characterized in that: The multi-scale feature coding module includes a fine feature extraction submodule, a region feature extraction submodule, and a feature pyramid fusion submodule; The fine feature extraction submodule, based on the spatially normalized geochemical data, uses a conventional convolutional layer network structure to obtain fine-scale features. The regional feature extraction submodule, based on the spatially normalized geochemical data, uses a hollow convolutional layer network structure to obtain regional-scale features. The feature pyramid fusion submodule, based on the aforementioned fine-scale features and regional-scale features, A feature pyramid network fusion method is used to obtain multi-scale encoded features.

5. The intelligent identification system for multi-element geochemical anomalies based on a convolutional autoencoder according to claim 1, characterized in that: The cross-correlation attention fusion module includes a cross-correlation calculation submodule, an attention weight generation submodule, and a feature fusion submodule; The cross-correlation calculation submodule, based on the multi-scale coding features, uses the cross-correlation matrix calculation method between feature channels to obtain the channel cross-correlation matrix; The attention weight generation submodule obtains the channel attention weights based on the channel cross-correlation matrix and using an attention weight generation algorithm. The feature fusion submodule, based on the multi-scale encoded features and channel attention weights, uses a weighted feature fusion algorithm to obtain element association enhancement features.

6. The intelligent identification system for multi-element geochemical anomalies based on a convolutional autoencoder according to claim 1, characterized in that: The collaborative anomaly reconstruction module includes a decoder submodule, a reconstruction calculation submodule, and an error measurement submodule. The decoder submodule, based on the element association enhancement features, uses a decoder network structure to obtain preliminary reconstructed data; The reconstruction calculation submodule, based on the spatially normalized geochemical data and the preliminary reconstruction data, uses an element-level difference calculation method to obtain the elemental reconstruction error. The error measurement submodule reconstructs the error based on the elements and uses an error aggregation method to obtain the reconstruction error.

7. The intelligent identification system for multi-element geochemical anomalies based on a convolutional autoencoder according to claim 1, characterized in that: The gradient weighted parsing module includes a gradient calculation submodule, a channel weight pooling submodule, and a heatmap generation submodule; The gradient calculation submodule, based on the reconstruction error and element correlation enhancement features, uses the gradient backpropagation algorithm to obtain the feature gradient map; The channel weight pooling submodule, based on the feature gradient map, uses a global average pooling method to obtain the channel contribution weights; The heatmap generation submodule, based on the channel contribution weights and element association enhancement features, uses a weighted class activation mapping method to obtain a feature contribution heatmap.

8. The intelligent identification system for multi-element geochemical anomalies based on a convolutional autoencoder according to claim 1, characterized in that: The active learning training module includes an uncertainty estimation submodule, a sample selection submodule, an expert annotation interface submodule, and a model fine-tuning submodule. The uncertainty estimation submodule, based on the reconstruction error, uses the Monte Carlo random deactivation method or the model ensemble method to obtain the prediction uncertainty; The sample selection submodule, based on the reconstruction error and prediction uncertainty, adopts an uncertainty-based sample selection strategy to obtain a candidate labeled sample set; The expert annotation interface submodule obtains expert annotation information based on the candidate annotation sample set using a human-computer interaction interface; The model fine-tuning submodule, based on standardized geochemical datasets, expert annotation information, and existing model weights, employs a transfer learning fine-tuning strategy to obtain updated model weights.

9. The intelligent identification system for multi-element geochemical anomalies based on a convolutional autoencoder according to claim 1, characterized in that: The anomaly contribution mapping module includes a feature channel analysis submodule, an element association submodule, and a contribution mapping submodule; The feature channel analysis submodule, based on the feature contribution heatmap, uses a feature channel importance analysis method to obtain high-contribution feature channel identifiers; The element association submodule, based on the high-contribution feature channel identifier, uses a feature-to-element mapping method to obtain the dominant element identifier; The contribution mapping submodule, based on the dominant element identifier and feature contribution heatmap, uses spatial mapping technology to obtain a spatial distribution map of element contribution.

10. The intelligent identification system for multi-element geochemical anomalies based on a convolutional autoencoder according to claim 1, characterized in that: The comprehensive results output module includes an anomaly threshold segmentation submodule, a multi-dimensional fusion submodule, and a visualization report generation submodule; The anomaly threshold segmentation submodule, based on the reconstruction error, uses a dynamic threshold determination method to obtain significant anomaly region identifiers; The multidimensional fusion submodule, based on the reconstruction error, the spatial distribution map of element contribution, and the identification of significant abnormal regions, uses data fusion technology to obtain comprehensive abnormal feature data; The visualization report generation submodule, based on the comprehensive anomaly feature data, uses geographic information system visualization and report generation technology to obtain a comprehensive identification report of multi-element geochemical anomalies and an anomaly spatial distribution map.

Images