Methods and systems for processing gold mineralization data using geologically constrained spatiotemporal attention networks

Abstract

This disclosure relates to the field of data processing technology, specifically to a method and system for processing gold mineralization data using a geologically constrained spatiotemporal attention network. The method includes: preprocessing gold mineralization data; training a network model using geochemical spatial sequence samples. The network model, from input to output, sequentially includes a deep feature extraction module, a data reconstruction module, and an anomaly score calculation module. The deep feature extraction module extracts local spatial features and models global correlation patterns. The data reconstruction module reconstructs the latent vectors output by the deep feature extraction module into the original data space. The anomaly score calculation module calculates the reconstruction error and normalizes it to an anomaly score, which is then linearly weighted and fused with a fault distance factor to generate a comprehensive anomaly score. The trained network model is used to identify anomalies in the gold mineralization data to be processed. This disclosure does not rely on strong distribution assumptions or manual threshold settings, and can efficiently and accurately identify gold mineralization anomalies in complex geological backgrounds.

Description

TECHNICAL FIELD

[0001] The present disclosure relates to the technical field of data processing, in particular to a geological constraint spatio-temporal attention network gold mineralization data processing method and system. BACKGROUND

[0002] With the continuous deepening of geological survey and mineral exploration work, a large amount of high-precision, multi-source geological data has been accumulated, among which gold mineralization data is a potential information source in mineral exploration, and is of decisive significance for accurate prediction of mineral target areas. Through analysis of the data, it can help regional key mineral resource exploration, reduce exploration costs, and provide certain technical support for target area prediction in important metallogenic belts. There are currently three methods for analyzing gold mineralization data: classical statistical methods, machine learning methods, and deep learning methods.

[0003] The classical statistical method, such as the element ratio method, has certain limitations in dealing with non-normal and non-linear data. Machine learning methods, including support vector machines (SVM) and random forests (RF), have the ability to further improve in capturing the spatial continuity and structural characteristics of gold mineralization anomalies, and are difficult to fully meet the demand for fine description of anomaly morphology. SUMMARY

[0004] In view of the problems of insufficient spatial feature extraction, limited long-range dependence modeling capability, and insufficient geological constraint fusion in the prior art, the present disclosure proposes a geological constraint spatio-temporal attention network gold mineralization data processing method and system.

[0005] The geological constraint spatio-temporal attention network gold mineralization data processing method provided by the embodiments of the present disclosure comprises:

[0006] The gold mineralization data is preprocessed, including: constructing a spatial neighborhood sequence centered on each sampling point to form a geochemical spatial sequence sample; calculating the shortest distance from each sampling point to the fault, and generating a fault distance factor according to the fault grade distribution weight;

[0007] The geochemical spatial sequence sample is used to train a network model, which comprises a deep feature extraction module, a data reconstruction module and an anomaly score calculation module from input to output. The deep feature extraction module extracts local spatial features and models global correlation patterns. The data reconstruction module reconstructs the latent vector output by the deep feature extraction module into the original data space. The anomaly score calculation module calculates the reconstruction error and normalizes it into an anomaly score. The fault distance factor is linearly weighted and fused according to the weight to generate a comprehensive anomaly score;

[0008] The trained network model is used for anomaly identification of the gold mineralization data to be processed.

[0009] Furthermore, the construction of the spatial neighborhood sequence includes: calculating the Euclidean distance based on the planar coordinates of the sampling points; using a dynamic radius adjustment strategy to find neighboring points for each sampling point, with the search radius gradually increasing from an initial value until a sufficient number of neighboring points are found or the search radius exceeds a set value; for each sampling point, selecting the n neighboring points with the smallest Euclidean distance from the spatial neighborhood, arranging them in ascending order of distance to form a spatial sequence; traversing all sampling points, stacking the obtained sequences to generate a three-dimensional tensor representing the entire dataset.

[0010] Furthermore, the deep feature extraction module includes:

[0011] One-dimensional convolutional neural networks are used to mine low-level local spatial features.

[0012] The spatial attention layer performs spatial domain recalibration on the convolutional feature map output by the one-dimensional convolutional neural network. It enhances key features by calculating the importance weights of spatial locations.

[0013] A bidirectional long short-term memory network layer captures the contextual dependencies and trends of the spatial sequences output by the spatial attention layer, and learns sequence features through LSTM units in both forward and backward directions.

[0014] A multi-head self-attention layer models the global correlation pattern of spatial sequences and captures the correlation information of different subspaces in parallel through multiple attention heads;

[0015] The bottleneck layer forms a latent space representation of the output features of the multi-head self-attention layer.

[0016] Furthermore, the one-dimensional convolutional neural network adopts a two-layer structure, with the number of output channels corresponding to the first and second layers being 128 and 256, respectively. The parameters of each layer of the one-dimensional convolutional neural network are set as follows: the initial number of input channels is the number of geochemical element features M; the convolution kernel size decreases sequentially; each layer of the one-dimensional convolutional neural network includes convolution operation, spatial attention mechanism, SiLU activation function, batch normalization processing, and max pooling operation in sequence; the spatial attention mechanism dynamically enhances the key spatial location features related to mineralization by calculating the attention weights of the query matrix, key matrix, and value matrix.

[0017] Furthermore, the formula for calculating the fracture distance factor is as follows:

[0018] ,

[0019] in, The distance from the sampling point to its nearest fault zone is the Euclidean distance. For scale parameters, This is the fracture distance factor.

[0020] Furthermore, the comprehensive anomaly score is calculated through linear weighted fusion:

[0021] ,

[0022] in The normalized reconstruction error anomaly score, These are the weighting coefficients. This is a comprehensive score for abnormalities.

[0023] A gold mineralization data processing system based on a geologically constrained spatiotemporal attention network, according to an embodiment of this disclosure, includes: a network model trained using geochemical spatial sequence samples and used for anomaly identification of the gold mineralization data to be processed; the network model, from input to output, includes a deep feature extraction module, a data reconstruction module, and an anomaly score calculation module; the deep feature extraction module extracts local spatial features and models global correlation patterns; the data reconstruction module reconstructs the latent vector output by the deep feature extraction module into the original data space; the anomaly score calculation module calculates the reconstruction error and normalizes it into an anomaly score, which is then linearly weighted and fused with a fault distance factor to generate a comprehensive anomaly score.

[0024] Furthermore, the deep feature extraction module includes:

[0025] One-dimensional convolutional neural networks are used to mine low-level local spatial features.

[0026] The spatial attention layer performs spatial domain recalibration on the convolutional feature map output by the one-dimensional convolutional neural network. It enhances key features by calculating the importance weights of spatial locations.

[0027] A bidirectional long short-term memory network layer captures the contextual dependencies and trends of the spatial sequences output by the spatial attention layer, and learns sequence features through LSTM units in both forward and backward directions.

[0028] A multi-head self-attention layer models the global correlation pattern of spatial sequences and captures the correlation information of different subspaces in parallel through multiple attention heads;

[0029] The bottleneck layer forms a latent space representation of the output features of the multi-head self-attention layer.

[0030] Furthermore, the one-dimensional convolutional neural network adopts a two-layer structure, with the number of output channels corresponding to the first and second layers being 128 and 256, respectively. The parameters of each layer of the one-dimensional convolutional neural network are set as follows: the initial number of input channels is the number of geochemical element features M; the convolution kernel size decreases sequentially; each layer of the one-dimensional convolutional neural network includes convolution operation, spatial attention mechanism, SiLU activation function, batch normalization processing, and max pooling operation in sequence; the spatial attention mechanism dynamically enhances the key spatial location features related to mineralization by calculating the attention weights of the query matrix, key matrix, and value matrix.

[0031] Compared with the prior art, the embodiments disclosed herein have at least the following beneficial effects:

[0032] This disclosure employs differentiated feature extraction strategies for the distribution characteristics of geochemical elements in different spatial locations. A one-dimensional convolutional neural network slides a one-dimensional convolutional kernel across the spatial sequence to explicitly extract local spatial patterns between adjacent sampling points, capturing local symbiotic combinations and variation features of elements. A spatial attention mechanism dynamically enhances the response of key location features related to mineralization and suppresses background noise interference. A bidirectional long short-term memory network comprehensively learns the directionality and trend of geochemical anomalies in spatial distribution through forward and backward processing sequences, simulating the spatial evolution of element migration along tectonic channels. A multi-head self-attention mechanism calculates the global correlation between any two locations in the sequence, macroscopically revealing the complex symbiotic combinations and spatial distribution patterns of geochemical elements. This progressive feature learning architecture ensures both the fine extraction of local anomaly features and the effective capture of long-range spatial dependencies. This disclosure does not rely on strong distribution assumptions or manual threshold settings, and can efficiently and accurately identify gold mineralization anomalies in complex geological backgrounds.

[0033] To better coordinate spatial feature extraction and sequence dependency modeling, this disclosure designs a four-tiered structure in the deep feature extraction module. Through the cascading cooperation of a one-dimensional convolutional neural network, a spatial attention layer, a bidirectional long short-term memory network layer, and a multi-head self-attention layer, it achieves multi-scale feature learning of geochemical anomalies from local to global perspectives. Specifically, this disclosure introduces the fault distance factor as a geological constraint into the anomaly evaluation system, establishing a dual constraint mechanism of "data-driven + geological prior." In the anomaly score calculation module, the geochemical anomaly score calculated from the model reconstruction error and the tectonic influence score calculated based on the fault distance are input respectively. Geochemical anomalies and geological structural constraints are fused through linear weighting, and the weights are adjusted according to the spatial correlation between the intensity of geochemical anomalies and fault structures to achieve accurate identification of gold mineralization data anomalies. This structure allows the model to not only learn the characteristics of the geochemical data itself but also incorporate the geological laws of fault-controlled mineralization, effectively avoiding false anomalies with unclear geological significance in low background areas and significantly improving the accuracy and reliability of mineral target area prediction. Attached Figure Description

[0034] Figure 1 A flowchart of the method provided in the embodiments of this disclosure;

[0035] Figure 2 This is a schematic diagram of the network structure of the network model provided in the embodiments of this disclosure. Detailed Implementation

[0036] To make the objectives, technical solutions, and advantages of this disclosure clearer, the following detailed description is provided in conjunction with embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of this disclosure.

[0037] See Figure 1 Combination Figure 2 As shown, a method for processing gold mineralization data using a geologically constrained spatiotemporal attention network includes:

[0038] S1 preprocessing of gold mineralization data includes: constructing a spatial neighborhood sequence centered on each sampling point to form a geochemical spatial sequence sample; calculating the shortest distance from each sampling point to the fault, and generating a fault distance factor by assigning weights according to the fault level;

[0039] Gold mineralization data, namely multi-element geochemical sampling datasets and regional fault data, were used for data preprocessing. This included outlier removal using the 3σ criterion and interquartile range method, and generating regular grid data for each element using Kriging interpolation. A spatial neighborhood sequence was constructed centered on each sampling point using KDTree (K-Dimensional Tree) spatial indexing and a dynamic radius search strategy, forming a geochemical spatial sequence sample. Simultaneously, the shortest distance from each sampling point to the fault was calculated, and a fault distance factor was generated by assigning weights according to the fault level. The processed data was then divided into training and testing sets. This embodiment uses a gold mining area as an example, employing a 1:50,000 scale geochemical exploration dataset containing key ore-forming elements such as Au, Ag, As, and Sb, totaling 4,662 sampling points. The fault data was divided into three categories according to level: Level 1 faults had a weight of 0.6, Level 2 faults had a weight of 0.3, and Level 3 faults had a weight of 0.1.

[0040] The construction of the spatial neighborhood sequence is specifically as follows: Euclidean distance is calculated based on the planar coordinates of the sampling points; a dynamic radius adjustment strategy is used to find at least 5 neighboring points for each center point; the search radius starts from an initial value. Start by following the formula Expand gradually until enough neighboring points or a radius exceeding [a certain value] is found. For each center point Select the five nearest neighbors with the smallest Euclidean distance from its neighborhood, and arrange them in ascending order of distance to form its spatial sequence. ; Traverse all central sampling points, stack the resulting sequences, and generate a three-dimensional tensor representing the entire dataset. , where N is the number of samples, 5 is the size of the spatial neighborhood window, and M is the number of geochemical element features.

[0041] S2 uses geochemical spatial sequence samples to train the network model. The network model includes a deep feature extraction module, a data reconstruction module, and an anomaly score calculation module from input to output. The deep feature extraction module extracts local spatial features and models global correlation patterns. The data reconstruction module reconstructs the latent vectors output by the deep feature extraction module into the original data space. The anomaly score calculation module calculates the reconstruction error and normalizes it into an anomaly score, which is then linearly weighted and fused with the fault distance factor to generate a comprehensive anomaly score.

[0042] The network model trained by S3 is used to identify anomalies in the gold mineralization data that needs to be processed.

[0043] In one embodiment, the deep feature extraction module includes:

[0044] One-dimensional convolutional neural networks are used to mine low-level local spatial features.

[0045] The spatial attention layer performs spatial domain recalibration on the convolutional feature map output by the one-dimensional convolutional neural network. It enhances key features by calculating the importance weights of spatial locations.

[0046] A bidirectional long short-term memory network layer captures the contextual dependencies and trends of the spatial sequences output by the spatial attention layer, and learns sequence features through LSTM units in both forward and backward directions.

[0047] A multi-head self-attention layer models the global correlation pattern of spatial sequences and captures the correlation information of different subspaces in parallel through multiple attention heads;

[0048] The bottleneck layer transforms the output features of the multi-head self-attention layer into a latent spatial representation. The one-dimensional convolutional neural network employs a two-layer one-dimensional convolutional structure with 128 and 256 convolutional kernels respectively, extracting local spatial features through convolution operations. The spatial attention layer recalibrates the convolutional output in the spatial domain, enhancing the response to key location features. The bidirectional long short-term memory network layer has 512 hidden units, capturing long-term dependencies in spatial sequences. The multi-head self-attention layer has four attention heads, modeling global correlation patterns. The data reconstruction module uses a two-level fully connected layer structure to reconstruct the latent vectors of the bottleneck layer into the original data space. The anomaly score calculation module first calculates the reconstruction error, normalizes it, and uses it as the anomaly score. Then, it linearly weights and fuses this error with the break distance factor at a weight of 0.7:0.3 to generate a comprehensive anomaly score.

[0049] The specific computational process of the deep feature extraction module is as follows:

[0050] One-dimensional convolutional neural network:

[0051] ,

[0052] in, This represents a one-dimensional convolution operation. and The first The kernel weights and bias terms of the layer This is the output of the previous layer. For a one-dimensional convolutional neural network Layer output;

[0053] After each convolutional layer, the process sequentially involves spatial attention, SiLU activation function, batch normalization, and max pooling. The formula for max pooling is as follows:

[0054] ,

[0055] in, Indicates the area covered by the pooling window. Indicates the first Feature values ​​at each position;

[0056] Spatial attention layer: This layer performs spatial domain recalibration on the convolutional feature map output by the one-dimensional convolutional neural network. It enhances key features by calculating the importance weights of spatial locations. The calculation formula is as follows:

[0057] ,

[0058] in, , , These are the query matrix, key matrix, and value matrix generated by the linear transformation of the input features, respectively. Let be the dimension of the key vector. For the output of the spatial attention layer, , and These are the weight matrices for the query matrix, the key matrix, and the value matrix, respectively.

[0059] Bidirectional Long Short-Term Memory Network Layer:

[0060] To capture the contextual dependencies and trends of spatial sequences, sequence features are learned through LSTM units in both forward and backward directions. The calculation formula is as follows:

[0061] ,

[0062] ,

[0063] ,

[0064] in, and These represent the forward and backward hidden states, respectively. The final output for each time step, For time step Input features, This indicates forward LSTM unit processing. This indicates backward processing by the LSTM unit;

[0065] Multi-head self-attention layer:

[0066] The global correlation pattern of the modeling sequence is obtained by capturing correlation information in different subspaces in parallel using multiple attention heads. The calculation formula is as follows:

[0067] ;

[0068] ;

[0069] in, and For a learnable parameter matrix, The number of attention heads is set to 4 in this disclosure. For the first One point of attention, For the first One point of attention, , and The first The weight matrix of the query matrix, the first query matrix The weight matrix of the i-th key matrix and the i-th key matrix The weight matrix of the value matrix, For transpose;

[0070] The spatial attention layer and the multi-head self-attention layer complement each other functionally. The spatial attention layer acts after the convolutional layer, focusing on the local area. By dynamically calculating the weights of each sampling point, it adaptively enhances the response at key locations related to mineralization and suppresses background noise interference. Finally, the multi-head self-attention layer acts as a global relation modeler, directly capturing the long-range dependencies between any two locations in the sequence, and macroscopically revealing the complex symbiotic combinations and spatial distribution patterns of gold mineralization elements.

[0071] After the features are extracted and fused through the above-mentioned layers, they are compressed to the bottleneck layer using global average pooling. This forms a highly abstract representation of the potential space.

[0072] The data reconstruction module gradually restores the data dimensionality and details through two fully connected layers. The specific operation process is as follows: The data reconstruction module receives the bottleneck layer latent vector output by the deep feature extraction module. The data is reconstructed into a spatial sequence of the original dimensions using the following computational steps:

[0073] First-level fully connected layer reconstruction calculation:

[0074] ,

[0075] Second-level fully connected layer reconstruction computation:

[0076] ,

[0077] in, and These are the learnable parameters for the fully connected layer. This represents the batch normalization operation, where SiLU is the activation function. This is the output of the first fully connected layer. This is the output of the second fully connected layer;

[0078] The output data is reconstructed by reconstructing the output layer. for:

[0079] ,

[0080] in, These are the weight parameters of the output layer. For the bias term of the output layer, This represents a linear transformation operation. The sigmoid function constrains each feature value of the reconstructed data to the interval [0,1], maintaining the same scale as the normalized input data; the reconstructed output data... With the original input Same dimensions.

[0081] fracture distance factor The calculation formula is:

[0082] ,

[0083] in, Sampling points The Euclidean distance to its nearest fault zone, The scale parameter controls the decay rate of the influence range and is set to 0.5 times the average fracture spacing in the study area.

[0084] Comprehensive Abnormal Score Calculated through linear weighted fusion:

[0085] ,

[0086] in The normalized reconstruction error anomaly score, The weighting factor is set to 0.7.

[0087] During training, cosine annealing learning rate scheduling was used, with an initial learning rate of 0.001 and weight decay of 1e-5. An early stopping mechanism was also added to prevent overfitting. After training, the network model was saved and tested using a test set. Then, the gold mineralization data to be processed was input into the trained network model for anomaly identification, and the output was gold mineralization anomaly detection results and prospecting target area prediction. The discrete point anomaly scores were converted into a continuous spatial distribution map using the Kriging interpolation method.

[0088] During the training of the network model, the mean absolute error is used as the reconstruction loss function:

[0089] ,

[0090] in, For the first One reconstructed output data, For the first One original input data, This represents the number of samples.

[0091] Finally, the training set is used to train the network model. After training, the network model is saved. After testing with the test set, the gold mineralization data to be processed is input into the trained network model for anomaly identification. The output is the gold mineralization anomaly detection result, and the anomaly target area is delineated accordingly.

[0092] On the other hand, see Figure 2 As shown, this disclosure provides a gold mineralization data processing system based on a geologically constrained spatiotemporal attention network, comprising: a network model, which is trained using geochemical spatial sequence samples and used to identify anomalies in the gold mineralization data to be processed. The network model includes, from input to output, a deep feature extraction module, a data reconstruction module, and an anomaly score calculation module. The deep feature extraction module extracts local spatial features and models global correlation patterns. The data reconstruction module reconstructs the latent vectors output by the deep feature extraction module into the original data space. The anomaly score calculation module calculates the reconstruction error and normalizes it into an anomaly score, which is then linearly weighted and fused with a fault distance factor to generate a comprehensive anomaly score.

[0093] The network model also includes an input data module for preprocessing gold mineralization data. The preprocessing of gold mineralization data includes: constructing a spatial neighborhood sequence centered on each sampling point to form a geochemical spatial sequence sample; calculating the shortest distance from each sampling point to the fault; and generating a fault distance factor by assigning weights according to the fault level.

[0094] In one embodiment, the deep feature extraction module includes:

[0095] One-dimensional convolutional neural networks are used to mine low-level local spatial features.

[0096] The spatial attention layer performs spatial domain recalibration on the convolutional feature map output by the one-dimensional convolutional neural network. It enhances key features by calculating the importance weights of spatial locations.

[0097] A bidirectional long short-term memory network layer captures the contextual dependencies and trends of the spatial sequences output by the spatial attention layer, and learns sequence features through LSTM units in both forward and backward directions.

[0098] A multi-head self-attention layer models the global correlation pattern of spatial sequences and captures the correlation information of different subspaces in parallel through multiple attention heads;

[0099] The bottleneck layer forms a latent space representation of the output features of the multi-head self-attention layer.

[0100] In one embodiment, the one-dimensional convolutional neural network adopts a two-layer structure, with the number of output channels corresponding to the first layer and the second layer being 128 and 256, respectively. The parameters of each layer of the one-dimensional convolutional neural network are set as follows: the initial number of input channels is the number of geochemical element features M; the convolution kernel size decreases sequentially; each layer of the one-dimensional convolutional neural network includes convolution operation, spatial attention mechanism, SiLU activation function, batch normalization processing and max pooling operation in sequence; the spatial attention mechanism dynamically enhances the key spatial location features related to mineralization by calculating the attention weights of the query matrix, key matrix and value matrix.

[0101] The data reconstruction module reconstructs the latent vectors output by the deep feature extraction module into the original data space, including the first-level fully connected layer and the second-level fully connected layer, and calculates the reconstructed output data through the reconstruction output layer;

[0102] The anomaly score calculation module calculates the reconstruction error and normalizes it into anomaly score. It then weights and fuses the anomaly score with the fracture distance factor and outputs a comprehensive anomaly score.

[0103] After the network model is trained, the gold mineralization data to be processed is input into the trained network model for anomaly identification, and the output is the gold mineralization anomaly detection result, which is used to delineate the anomaly target area.

[0104] The above description is merely a preferred embodiment of this disclosure and is not intended to limit this disclosure. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this disclosure should be included within the scope of protection of this disclosure.

Claims

1. A method for processing gold mineralization data of a geologically constrained spatiotemporal attention network, characterized in that, The application relates to a gold mineralization data processing method based on a deep learning network model. The method comprises the following steps: ， wherein, is the Euclidean distance of a sample point to its nearest fault zone, is a scale parameter, controlling the rate of decay of the influence range, is a fault distance factor; Preprocessing gold mineralization data, including: constructing a spatial neighborhood sequence centered on each sampling point to form a geochemical spatial sequence sample; calculating the shortest distance from each sampling point to a fault, generating a fault distance factor according to the fault grade distribution weight, and the calculation formula of the fault distance factor is: Training the network model by using the geochemical spatial sequence sample, wherein the network model comprises a deep feature extraction module, a data reconstruction module and an anomaly score calculation module from input to output, the deep feature extraction module extracts local spatial features and models global correlation patterns, the data reconstruction module reconstructs the latent vector output by the deep feature extraction module into the original data space, the anomaly score calculation module calculates the reconstruction error and normalizes it into an anomaly score, and the anomaly score is linearly weighted and fused with the fault distance factor according to the weight to generate a comprehensive anomaly score; The trained network model is used for anomaly recognition of gold mineralization data to be processed. The method comprises the following steps: The method comprises the following steps: A one-dimensional convolutional neural network is used for bottom local spatial feature mining. A spatial attention layer is used for recalibrating the convolution feature map output by the one-dimensional convolutional neural network in the spatial domain, and the importance weight of the spatial position is calculated to enhance the key features. A bidirectional long short-term memory network layer is used for capturing the context dependence and trend of the spatial sequence output by the spatial attention layer, and the sequence features are learned by the forward and backward LSTM units.

2. The method of claim 1, wherein the method is applied to gold mineralization data. A multi-head self-attention layer is used for modeling the global correlation pattern of the spatial sequence, and the correlation information of different subspaces is captured by multiple attention heads in parallel.

3. The method of claim 1, wherein the method is applied to gold mineralization data. A bottleneck layer is used for forming the latent space representation of the output features of the multi-head self-attention layer. ， wherein is the normalized reconstruction error anomaly score, is a weight coefficient, is the composite anomaly score.

4. A geological-constrained spatio-temporal attention network-based gold mineralization data processing system for implementing the gold mineralization data processing method of any one of claims 1-3, characterized in that, The one-dimensional convolutional neural network adopts a two-layer structure, and the output channel numbers of the first layer and the second layer are 128 and 256 respectively. The comprehensive anomaly score is calculated by linear weighted fusion: The application relates to a gold mineralization data processing method based on a deep learning network model. The method comprises the following steps: Preprocessing gold mineralization data, including: constructing a spatial neighborhood sequence centered on each sampling point to form a geochemical spatial sequence sample; calculating the shortest distance from each sampling point to a fault, generating a fault distance factor according to the fault grade distribution weight, and the calculation formula of the fault distance factor is: Training the network model by using the geochemical spatial sequence sample, wherein the network model comprises a deep feature extraction module, a data reconstruction module and an anomaly score calculation module from input to output, the deep feature extraction module extracts local spatial features and models global correlation patterns, the data reconstruction module reconstructs the latent vector output by the deep feature extraction module into the original data space, the anomaly score calculation module calculates the reconstruction error and normalizes it into an anomaly score, and the anomaly score is linearly weighted and fused with the fault distance factor according to the weight to generate a comprehensive anomaly score; The trained network model is used for anomaly recognition of gold mineralization data to be processed. The method comprises the following steps: The method comprises the following steps: A one-dimensional convolutional neural network is used for bottom local spatial feature mining. A spatial attention layer is used for recalibrating the convolution feature map output by the one-dimensional convolutional neural network in the spatial domain, and the importance weight of the spatial position is calculated to enhance the key features. A bidirectional long short-term memory network layer is used for capturing the context dependence and trend of the spatial sequence output by the spatial attention layer, and the sequence features are learned by the forward and backward LSTM units. A multi-head self-attention layer is used for modeling the global correlation pattern of the spatial sequence, and the correlation information of different subspaces is captured by multiple attention heads in parallel. A bottleneck layer is used for forming the latent space representation of the output features of the multi-head self-attention layer. The one-dimensional convolutional neural network adopts a two-layer structure, and the output channel numbers of the first layer and the second layer are 128 and 256 respectively. The comprehensive anomaly score is calculated by linear weighted fusion: The application relates to a gold mineralization data processing method based on a deep learning network model. The method comprises the following steps: Preprocessing gold mineralization data, including: constructing a spatial neighborhood sequence centered on each sampling point to form a geochemical spatial sequence sample; calculating the shortest distance from each sampling point to a fault, generating a fault distance factor according to the fault grade distribution weight, and the calculation formula of the fault distance factor is: Training the network model by using the geochemical spatial sequence sample, wherein the network model comprises a deep feature extraction module, a data reconstruction module and an anomaly score calculation module from input to output, the deep feature extraction module extracts local spatial features and models global correlation patterns, the data reconstruction module reconstructs the latent vector output by the deep feature extraction module into the original data space, the anomaly score calculation module calculates the reconstruction error and normalizes it into an anomaly score, and the anomaly score is linearly weighted and fused with the fault distance factor according to the weight to generate a comprehensive anomaly score; The trained network model is used for anomaly recognition of gold mineralization data to be processed. The method comprises the following steps: The method comprises the following steps: The network model is trained via geochemical spatial sequence samples and used for anomaly identification of gold mineralization data to be processed, and sequentially comprises a deep feature extraction module, a data reconstruction module and an anomaly score calculation module from input to output, the deep feature extraction module extracts local spatial features and models global correlation patterns, the data reconstruction module reconstructs the latent vector output by the deep feature extraction module into the original data space, the anomaly score calculation module calculates the reconstruction error and normalizes it into an anomaly score, linearly weights and fuses the anomaly score with a fracture distance factor according to weights to generate a comprehensive anomaly score.

5. The gold mineralization data processing system of claim 4, wherein, The one-dimensional convolutional neural network adopts a two-layer structure, and the output channel numbers corresponding from the first layer to the second layer are 128 and 256 respectively; the parameter setting of each one-dimensional convolutional neural network is that: the initial input channel number is the number M of geochemical element features; the convolution kernel size decreases successively; each one-dimensional convolutional neural network successively comprises convolution operation, spatial attention mechanism, SiLU activation function, batch normalization processing and maximum pooling operation; the spatial attention mechanism dynamically enhances the key spatial position features related to mineralization by calculating the attention weights of the query matrix, the key matrix and the value matrix.

Images