Intelligent planning method for geological mineral exploration analysis model

Abstract

The invention discloses an intelligent planning method of a geological mineral exploration analysis model, and relates to the field of mineral exploration, and the intelligent planning method comprises the following specific steps: S101, obtaining regional geological mineral data through satellite remote sensing, ground survey and historical data collection, and preprocessing the collected geological mineral data; according to the method, powerful prior knowledge support can be provided for simulation learning, exploration planning is more scientific and reliable, mineral resource distribution can be predicted more comprehensively, model generation efficiency and calculation performance are improved, multi-level reasoning and prediction can be carried out on different geological levels, limitation of traditional exploration can be broken through, and the method is suitable for popularization and application. The overall effect of the intelligent planning method is improved; according to the method, global search can be carried out in a large range, local optimum is avoided, meanwhile, the accuracy and reliability of path planning are improved, the diversity and adaptability of path selection are ensured, and a model can more intelligently select a path with high exploration value when facing an unknown area.

Description

Technical Field

[0001] The present invention relates to the field of mineral exploration, and particularly to an intelligent planning method for a geological and mineral exploration analysis model. Background Art

[0002] Geological and mineral exploration is a key link in resource development and economic construction, and is of great significance for realizing the efficient development and sustainable utilization of mineral resources. With the rapid development of social economy, the demand for mineral resources continues to grow, and the task of geological and mineral exploration becomes increasingly arduous. Under complex geological conditions. However, traditional mineral exploration methods mainly rely on empirical analysis and on-site investigation, and there are problems such as information asymmetry, low analysis efficiency, and blind exploration path planning. In a complex geological environment, these methods usually have difficulty in quickly locating ore-forming areas, which not only increases the exploration time and cost, but also easily misses key mineral resources due to insufficient data or misjudgment;

[0003] Existing intelligent planning methods for geological and mineral exploration analysis models cannot provide strong prior knowledge support for simulation learning, reduce the model generation efficiency and computational performance, and cannot perform multi-level reasoning and prediction at different geological levels; in addition, existing intelligent planning methods for geological and mineral exploration analysis models cannot perform global search over a large area, reduce the accuracy and reliability of path planning, and cannot intelligently select paths with high exploration value when facing unknown areas. For this reason, we propose an intelligent planning method for a geological and mineral exploration analysis model. Summary of the Invention

[0004] The object of the present invention is to solve the above-mentioned problems, and provide an intelligent planning method for a geological and mineral exploration analysis model.

[0005] The present invention proposes an intelligent planning method for a geological and mineral exploration analysis model. The specific steps of the intelligent planning method are as follows:

[0006] S101: Obtain regional geological and mineral data through satellite remote sensing, ground survey and historical data collection, and preprocess the collected geological and mineral data;

[0007] S102: Based on the preprocessed data, construct a geological and mineral atlas, and construct a virtual geological scene through simulation learning;

[0008] S103: Integrate multi-mining area data, extract the characteristic information of different types of geological data, and then construct a geological exploration model to predict the mineral distribution area;

[0009] S104: Search and optimize the mineral distribution area according to the prediction result, evaluate the benefits of different exploration paths, and select the mineral exploration path;

[0010] S105: Dynamically optimize the exploration path, apply the optimized exploration path to the actual exploration work, verify the exploration results, and dynamically adjust the parameters of the geological exploration model based on the measured data.

[0011] As a further solution of the present invention, the specific steps of preprocessing the collected geological and mineral data in S101 are as follows:

[0012] S1.1: Collect various remote sensing data of multi-spectral remote sensing, thermal infrared remote sensing, and microwave remote sensing through satellite remote sensing technology. Then, perform wavelet transform on each group of remote sensing data and retain the main signal components to obtain the smoothed remote sensing data.

[0013] S1.2: Collect various geological measurement data of lithology, geological structure, and fault zone through on-site surveys. Sort the different types of geological measurement data separately from small to large. Use the quartile algorithm to divide the sorted groups of geological measurement data into 4 parts according to a ratio of 25%, and calculate the first quartile Q Bp1 and the third quartile Q Bp3 ;

[0014] S1.3: Based on the first quartile Q Bp1 and the third quartile Q Bp3 , calculate the interquartile range IQR of various geological measurement data = Q Bp3 -Q Bp1 . Then, set Q Bp3 +1.5×IQR as the upper limit of the normal data range, and set Q Bp1 -1.5×IQR as the lower limit of the normal data range. Identify all data in the geological measurement data that is less than the lower limit or greater than the upper limit, regard it as an outlier and eliminate it. After eliminating the outliers, re-check the data distribution.

[0015] S1.4: Standardize each remote sensing data, geological measurement data, and historical data, convert the data to the same scale, then use the inverse distance weighting method to fill in the missing areas in the remote sensing data or the blank points in the geological measurement. After that, fuse the geological and mineral data from different data sources through the principal component analysis method.

[0016] It should be further noted that the specific calculation formula of the wavelet transform described in S1.1 is as follows:

[0017]

[0018] In the formula, f WT (x WT ) represents the denoised remote sensing data; represents the wavelet basis function; represents the wavelet coefficient; j WT and k WTrepresent the scale and location parameters respectively;

[0019] The specific calculation formula of the inverse distance weighting method described in S1.4 is as follows:

[0020]

[0021] In the formula, represents the i T -th interpolation weight; p ide represents the weight exponent; and represent the distances between the i T -th and j T -th points to be filled and the known points respectively; n ide represents the total number of known points; f IDE (x IDE ) represents the value of the point to be filled; represents the value of the known point; represents the coordinates of the i T -th known point.

[0022] As a further solution of the present invention, the specific steps of constructing the geological and mineral atlas described in S102 are as follows:

[0023] S2.1: Use NER technology to automatically identify each group of entities such as mining areas, ore types, geological structures, metallogenic conditions, ore deposits and strata, as well as each attribute such as ore body depth, grade, reserves and geological age in the preliminarily processed geological and mineral data;

[0024] S2.2: Extract the logical relationships between entities from the geological and mineral data based on the rule matching method of dependency syntax, calculate the similarity between each group of entities, remove redundant entities with a similarity higher than the preset threshold, and convert the identified entities and their relationships into triple form (h MA ,r MA ,t MA ), that is, "head entity - relationship - tail entity", and store the constructed triples in the graph database for storage to generate the corresponding geological and mineral atlas, where entities are used as nodes and relationships are used as edges;

[0025] S2.3: Based on the path search algorithm and logical rule reasoning, calculate the inference probability between each entity, and judge the credibility of the implicit relationship according to the inference probability between each entity. If the inference probability is higher than the preset threshold, it is judged that there is an implicit relationship between the two groups of entities and record it;

[0026] S2.4: Collect the addition of new geological and mineral data and exploration results in real time, continuously update the entities and relationships in the geological and mineral atlas, and display the geological and mineral atlas through a visualization tool, and provide interactive query and analysis.

[0027] It should be further noted that the specific calculation formula for the inference probability described in S2.3 is as follows:

[0028]

[0029] In the formula, P Ip (h MA →t MA ) represents the inference probability from the head entity h MA to the tail entity t MA ; Ρ MA represents the set of all paths from the head entity h MA to the tail entity t MA ; p MA represents a path from the head entity h MA to the tail entity t MA ; |p MA | represents the length of the path p MA ; represents the i MA -th relationship on the path p Ip .

[0030] As a further solution of the present invention, the specific steps for constructing a virtual geological scene through simulation learning described in S102 are as follows:

[0031] S3.1: Extract various prior knowledge of geological and mineral distributions from the constructed geological and mineral atlas, including the association rules between geological structures, mineral types, and metallogenic conditions, and represent the extracted prior knowledge in the form of triples. Then, generate an initial virtual geological scene model according to the extracted prior knowledge;

[0032] S3.2: Generate multiple groups of different combinations of geological condition parameters through random sampling, input the sampled geological condition parameters into the virtual geological scene model. Based on real-time geological and mineral data and historical mineral data, the model calculates the mineral distribution under each group of geological conditions through physical equations and geological laws to simulate the mineral distribution results, and compares the simulation results with the geological and mineral atlas or actual exploration data to calculate the simulation error of the virtual geological scene model;

[0033] S3.3: According to the calculated simulation error, optimize the geological parameters in the simulation model through numerical simulation methods to adjust the prediction ability of the model, and take minimizing the simulation error as the goal to iteratively update the geological parameters until the simulation error of the virtual geological scene model converges to a preset range. Then, collect the available computing resources of each mineral exploration device, and construct a lightweight model based on the virtual geological scene model that has been trained and completed according to the available computing resources of each mineral exploration device;

[0034] S3.4: Input the historical mineral data into the trained virtual geological scene model and the lightweight model respectively. Use the true value of the historical mineral data and the output of the trained virtual geological scene model as the hard target and the soft target respectively. Then construct a total loss function to calculate the total loss between the output of the lightweight model and the hard target and the soft target.

[0035] S3.5: Perform backpropagation on the calculated total loss, and optimize the parameters of the lightweight model based on the backpropagation results. After the parameter optimization is completed, evaluate the prediction accuracy of the lightweight model for mineral distribution and the computational efficiency of the model on the unused historical mineral data. According to the verification results, adjust the hyperparameters in the model structure or the distillation process, and repeat the training and verification of the lightweight model until the total loss value of each lightweight model converges to the preset threshold.

[0036] S3.6: According to the available computing resources of different mineral exploration devices, deploy the virtual geological scene model and the lightweight model to the corresponding devices respectively, input the real-time geological mineral data into each model, generate a virtual geological scene containing the mineral distribution under different geological conditions through forward propagation, and display the generated geological scene through a graphical interface. At the same time, update the parameters of each model in real time according to the actual exploration data.

[0037] It should be further noted that the specific calculation formula for the simulation error described in S3.2 is as follows:

[0038]

[0039] In the formula, VG Error represents the average error of the simulation results; n Gp represents the number of groups of geological condition parameters sampled; represents the mineral distribution output by the simulation model under the i Gp th group of geological condition parameters; represents the actual mineral distribution under the i Gp th group of geological condition parameters;

[0040] The specific calculation formula for the total loss function described in S3.4 is as follows:

[0041]

[0042] L hard =||z Sd -y Sd || 2

[0043] L 总 =αL soft +(1-α)L hard

[0044] In the formula, Represents the soft target of the i-th mineral data generated by the virtual geological scene model Hg for the mineral data; Represents the score of the i-th mineral data output by the virtual geological scene model Hg ; T tp Represents the temperature parameter used to smooth the probability distribution; Represents the score of the j-th mineral data output by the virtual geological scene model Hg ; L soft Represents the soft target loss; Represents the predicted probability of the i-th mineral data of the lightweight model Hg ; L hard Represents the hard target loss; z Sd Represents the predicted value of the lightweight model; y Sd Represents the true value of the actual mineral distribution; L 总 Represents the total loss function; α represents the trade-off parameter.

[0045] As a further solution of the present invention, the specific steps of constructing the geological exploration model in S103 to predict the mineral distribution area are as follows:

[0046] S4.1: Extract exploration areas with different geological characteristics from the geological map mineral atlas, and store the exploration data of each exploration area in an independent regional server k RS , and the original exploration data is not shared among the regional servers k RS . Then, initialize a set of global optimization models in the central server and randomly generate the model parameters of this global optimization model ;

[0047] S4.2: Each regional server k RS receives the model parameters of the current iteration of the global model from the central server . Each regional server k RS replaces the original local model's model parameters according to the received model parameters. Then, each regional server uses the local exploration data set to train the local model with the goal of minimizing the loss value of this local model, and optimizes the local model parameters through the gradient descent method; ;

[0048] S4.3: After local training is completed on each regional server k RS , the updated model parameters are sent to the central server. The central server receives the local model parameters uploaded by each region and aggregates them through the weighted average algorithm to generate new global model parameters;

[0049] S4.4: Evaluate the performance metrics of the global model using historical mineral data. If the performance metrics of the global model do not reach the preset threshold, distribute the aggregated global model parameters back to each regional server k RS , and continue with the next round of local training and aggregation. Repeat multiple rounds of iteration until the performance of the global model converges to the preset threshold. Apply the trained global model to the virtual geological scenario model and the lightweight model, and optimize the mineral distribution under different geological conditions simulated by each model.

[0050] As a further solution of the present invention, the specific steps for searching and optimizing the mineral distribution area according to the prediction results in S104 are as follows:

[0051] S5.1: Collect the optimized mineral distribution under different geological conditions and use it as the search space. Based on this search space, construct a search tree, and use this search space as the root node of the tree. Each child node in the tree represents the mineral distribution under a certain geological condition in the search space. At the same time, initialize the evaluation value and access count of each node in the search tree;

[0052] S5.2: Starting from the root node, calculate the UCB1 value of each child node, and according to the upper confidence bound selection strategy, sequentially select the child node with the highest UCB1 value until reaching an unexpanded child node, that is, the currently unexplored sub-region. Then, expand new child nodes by dividing the sub-region of the current child node into smaller sub-regions or adjusting the value range of geological parameters;

[0053] S5.3: In the corresponding search space of the newly added child node, randomly generate a set of geological parameters and calculate its corresponding metallogenic potential. Propagate the simulation results backward from the current node upwards to update the evaluation value and access count of each node;

[0054] S5.4: Repeat the steps of selection, expansion, simulation, and backtracking until the metallogenic potential converges within the preset threshold. After the search stops, traverse each node in the search tree and select the child node with the highest evaluation value as the optimal mineral distribution area, and plan the exploration path based on the selected groups of mineral distribution areas.

[0055] As a further solution of the present invention, the specific steps for dynamically optimizing the exploration path are as follows:

[0056] S6.1: With the goal of minimizing the total path cost and maximizing the path value, model the optimization of the exploration path as a combinatorial optimization problem, and generate a set of initial paths based on the selected groups of mineral distribution areas. Each path p Tpc is a set of individuals in the population P path , and through the objective function F Tpc (p Tpc ) = V Tpc(p Tpc ) - λC Tpc (p Tpc ) Evaluate the fitness of each path, where F Tpc (p Tpc ) represents the fitness value of path p Tpc , V Tpc (p Tpc ) represents the sum of the metallogenic potential and information gain of path p Tpc , and C Tpc (p Tpc ) represents the total cost of path p Tpc ; λ represents the weight of the cost;

[0057] S6.2: Select the path p Tpc (p Tpc ) whose fitness F Tpc meets the preset demand value for cloning to generate a clone pool. Mutate each group of paths in the clone pool by adjusting the order of geological points on path p Tpc or inserting new geological points, and record the mutated path as p' Tpc ;

[0058] S6.3: Select the paths that meet the preset demand value from the mutated clone pool and the initial population to form a new generation population Select the new population and the corresponding fitness F Tpc of each path p Tpc (p Tpc ) in it, and then construct a GP model as a surrogate model;

[0059] S6.4: Use the GP model to calculate the mean prediction and variance prediction of each group of paths p Tpc , and then calculate the standardized improvement amount of each path p Tpc according to the mean prediction and variance prediction. Then, through the cumulative distribution function and probability density function of the standard normal distribution, calculate the expected improvement of the corresponding path p Tpc . Select the path p with the maximum expected improvement in the population Tpc by the grid search method;

[0060] S6.5: Calculate the fitness F Tpc of the selected path p Tpc (p Tpc ), and add it to the historical data set. Then, based on the updated historical data set, retrain the GP model, update the mean prediction and variance prediction, and repeat the expected improvement calculation and historical data set update until the improvement amplitude of the fitness F Tpc (p Tpc ) is lower than the preset threshold;

[0061] S6.6: Recalculate the fitness F Tpc of each group of paths p Tpc (p Tpc ) in the population, and re - select and mutate each group of paths p Tpc . Then, optimize the expected improvement of each path p Tpc in the updated population through the GP model, and perform repeated update iterations until the fitness value F Tpc (p Tpc ) converges to the preset index, and select the path with the highest fitness as the optimal exploration path.

[0062] Advantages of the present invention:

[0063] 1. The present invention obtains regional geological and mineral data through satellite remote sensing, ground surveys, and historical data collection, constructs a corresponding geological and mineral atlas, extracts various prior knowledge of geological and mineral distribution from the constructed geological and mineral atlas, including the association rules between geological structures, mineral types, and metallogenic conditions, and represents the extracted prior knowledge in the form of triples. Then, an initial virtual geological scene model is generated based on the extracted prior knowledge. Multiple different combinations of geological condition parameters are generated through random sampling, and the sampled geological condition parameters are input into the virtual geological scene model. Based on real - time geological and mineral data and historical mineral data, the model calculates the mineral distribution under each group of geological conditions through physical equations and geological laws to simulate the mineral distribution results, and optimizes the virtual geological scene model based on the simulation results. Then, a lightweight model based on the trained virtual geological scene model is constructed according to the available computing resources of each mineral exploration device, and the real values of historical mineral data and the outputs of the trained virtual geological scene model are used as hard targets and soft targets respectively to optimize each group of lightweight models. According to the available computing resources of different mineral exploration devices, the virtual geological scene model and the lightweight model are respectively deployed to the corresponding devices, and real - time geological and mineral data are input into each model. Through forward propagation, a virtual geological scene containing mineral distributions under different geological conditions is generated, and the generated geological scene is displayed through a graphical interface. At the same time, the parameters of each model are updated in real time according to the actual exploration data, which can provide strong prior knowledge support for simulation learning, make the exploration plan more scientific and reliable, help to more comprehensively predict the distribution of mineral resources, improve the model generation efficiency and computing performance, enable multi - level reasoning and prediction at different geological levels, is conducive to breaking through the limitations of traditional exploration, and improves the overall effect of intelligent planning methods.

[0064] 2. The present invention generates a set of initial paths based on the selected mineral distribution areas of each group. With the goal of minimizing the total path cost and maximizing the path value, it calculates the fitness of each path, clones, mutates, and selects each group of paths based on the fitness. Then, it uses the GP model to calculate the mean prediction and variance prediction of each group of paths. After that, it calculates the standardized improvement amount of each path according to the mean prediction and variance prediction, and then calculates the expected improvement of the corresponding path through the cumulative distribution function and probability density function of the standard normal distribution. It selects the path with the maximum expected improvement through the grid search method, calculates the fitness of the selected path, and adds it to the historical data set. Then, based on the updated historical data set, it retrains the GP model, updates the mean prediction and variance prediction, repeats the expected improvement calculation and historical data set update until the fitness improvement amplitude is lower than the preset threshold, recalculates the fitness of each group of paths in the population, and re-selects and mutates each group of paths. Then, it optimizes the expected improvement of each path in the updated population through the GP model and performs repeated update iterations until the fitness value converges to the preset index. Finally, it selects the path with the highest fitness as the optimal exploration path, which can perform global search in a large range, avoid falling into local optima, improve the accuracy and reliability of path planning, and ensure the diversity and adaptability of path selection, enabling the model to more intelligently select paths with high exploration value when facing unknown areas. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] The present invention will be further described below with reference to the accompanying drawings.

[0066] Figure 1 It is a framework diagram of an intelligent planning method for a geological and mineral exploration analysis model. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0067] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0068] All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0069] Embodiment 1

[0070] The embodiment of the present invention provides an intelligent planning method for a geological and mineral exploration analysis model. Refer to Figure 1 , Figure 1A framework diagram of an intelligent planning method for a geological and mineral exploration analysis model provided by an embodiment of the present invention. The intelligent planning method comprises the following steps:

[0071] Through satellite remote sensing, ground surveys and historical data collection, regional geological and mineral data are obtained, and the collected geological and mineral data are pre-processed.

[0072] Specifically, satellite remote sensing technology is used to collect multispectral remote sensing, thermal infrared remote sensing and microwave remote sensing data, and then wavelet transform is performed on each group of remote sensing data, and the main signal components are retained to obtain smoothed remote sensing data. Through ground surveys, various geological measurement data on lithology, geological structure and fault zones are collected, and different types of geological measurement data are sorted from small to large. The sorted groups of geological measurement data are divided into 4 parts at a ratio of 25% using the quartile algorithm, and the first quartile Q of the data is calculated. Bp1 and the third quartile Q Bp3 , according to the first quartile Q Bp1 and the third quartile Q Bp3 , calculate the interquartile range IQR = Q of various geological measurement data Bp3 -Q Bp1 , then Q Bp3 +1.5×IQR is set as the upper limit of the normal data range, and Q Bp1 -1.5×IQR is set as the lower limit of the normal data range. All data in the geological survey data that are less than the lower limit or greater than the upper limit are identified, regarded as outliers and eliminated. After eliminating the outliers, the data distribution is rechecked, and the remote sensing data, geological survey data and historical data are standardized. The data are converted to the same scale, and the inverse distance weighted method is used to fill in the missing areas in the remote sensing data or the blank points in the geological measurement. Then, the geological and mineral data from different data sources are fused through principal component analysis.

[0073] In this embodiment, the specific calculation formula of wavelet transform is as follows:

[0074]

[0075] In the formula, f WT (x WT ) represents the denoised remote sensing data; represents the wavelet basis function; represents the wavelet coefficient; j WT With k WT represent scale and location parameters respectively;

[0076] The specific calculation formula of the inverse distance weighted method is as follows:

[0077]

[0078] Wherein, represents the i T -th interpolation weight; p ide represents the weight exponent; and respectively represent the distances between the i T -th and the j T -th points to be completed and the known points; n ide represents the total number of known points; f IDE (x IDE ) represents the value of the point to be completed; represents the value of the known point; represents the coordinates of the i T -th known point.

[0079] Based on the preprocessed data, a geological and mineral atlas is constructed, and a virtual geological scene is constructed through simulation learning.

[0080] Specifically, the NER technology is used to automatically identify each group of entities such as mining areas, ore types, geological structures, metallogenic conditions, ore deposits and strata, as well as attributes such as ore body depth, grade, reserves and geological age in the preliminarily processed geological and mineral data. Based on the rule matching method of dependency syntax, the logical relationships between entities are extracted from the geological and mineral data, the similarity between each group of entities is calculated, and redundant entities with similarity higher than the preset threshold are removed. The identified entities and their relationships are transformed into the triple form (h MA , r MA , t MA ), that is, "head entity - relationship - tail entity", and the constructed triples are stored in the graph database to generate the corresponding geological and mineral atlas. Among them, the entities are used as nodes and the relationships are used as edges. Based on the path search algorithm and logical rule reasoning, the inference probability between each entity is calculated, and the credibility of the implicit relationship is judged according to the inference probability between each entity. If the inference probability is higher than the preset threshold, it is judged that there is an implicit relationship between the two groups of entities and recorded. New geological and mineral data and exploration results are collected in real time and the entities and relationships in the geological and mineral atlas are continuously updated. The geological and mineral atlas is displayed through a visualization tool, and interactive query and analysis are provided.

[0081] It should be further noted that the specific calculation formula of the inference probability is as follows:

[0082]

[0083] Wherein, P Ip (h MA →t MA ) represents the inference probability from the head entity h MA to the tail entity t MA ; Ρ MA represents from the head entity hMA All path sets to the tail entity t MA ; p MA Represents a path from the head entity h MA to the tail entity t MA ; |p MA | represents the length of the path p MA ; Represents the i-th relationship on the path p MA i Ip th relationship

[0084] Specifically, extract various prior knowledge of geological and mineral distribution from the constructed geological and mineral atlas, including the association rules between geological structures, mineral types, and metallogenic conditions, and represent the extracted prior knowledge in the form of triples. Then, generate an initial virtual geological scenario model based on the extracted prior knowledge, generate multiple groups of different geological condition parameter combinations through random sampling, input the sampled geological condition parameters into the virtual geological scenario model, and based on real-time geological and mineral data and historical mineral data, the model calculates the mineral distribution under each group of geological conditions through physical equations and geological laws to simulate the mineral distribution results. Compare the simulation results with the geological and mineral atlas or actual exploration data, calculate the simulation error of the virtual geological scenario model, and optimize the geological parameters in the simulation model through numerical simulation methods to adjust the prediction ability of the model. With the goal of minimizing the simulation error, iteratively update the geological parameters until the simulation error of the virtual geological scenario model converges to a preset range. Then, collect the available computing resources of each mineral exploration device, and construct a lightweight model based on the trained virtual geological scenario model according to the available computing resources of each mineral exploration device. Input the historical mineral data into the trained virtual geological scenario model and the lightweight model respectively, use the true value of the historical mineral data and the output of the trained virtual geological scenario model as the hard target and the soft target respectively, then construct a total loss function to calculate the total loss of the output of the lightweight model and the hard target and the soft target, perform backpropagation on the calculated total loss, and optimize the parameters of the lightweight model based on the backpropagation results. After the parameter optimization is completed, evaluate the prediction accuracy of the mineral distribution of the lightweight model and the computing efficiency of the model on the unused historical mineral data. According to the verification results, adjust the hyperparameters in the model structure or the distillation process, repeat the training and verification of the lightweight model until the total loss value of each lightweight model converges to a preset threshold. According to the available computing resources of different mineral exploration devices, deploy the virtual geological scenario model and the lightweight model to the corresponding devices respectively, input the real-time geological and mineral data into each model, generate a virtual geological scenario containing the mineral distribution under different geological conditions through forward propagation, and display the generated geological scenario through a graphical interface, and at the same time, update the model parameters in real time according to the actual exploration data.

[0085] In addition, it should be noted that the specific calculation formula for the simulation error is as follows:

[0086]

[0087] In the formula, VG Error represents the average error of the simulation results; n Gp represents the number of groups of sampled geological condition parameters; represents the mineral distribution output by the simulation model under the i Gp -th group of geological condition parameters; represents the actual mineral distribution under the i Gp -th group of geological condition parameters;

[0088] The specific calculation formula for the total loss function is as follows:

[0089]

[0090] L hard = ||z Sd - y Sd || 2

[0091] L 总 = αL soft + (1 - α)L hard

[0092] In the formula, represents the soft target of the i Hg -th mineral data generated by the virtual geological scenario model; represents the score of the i Hg -th mineral data output by the virtual geological scenario model; T tp represents the temperature parameter used to smooth the probability distribution; represents the score of the j Hg -th mineral data output by the virtual geological scenario model; L soft represents the soft target loss; represents the predicted probability of the i Hg -th mineral data of the lightweight model; L hard represents the hard target loss; z Sd represents the predicted value of the lightweight model; y Sd represents the true value of the actual mineral distribution; L 总 represents the total loss function; α represents the trade-off parameter.

[0093] Integrate the data of multiple mining areas, extract the characteristic information of different types of geological data, and then construct a geological exploration model to predict the mineral distribution area.

[0094] Specifically, exploration areas with different geological characteristics are extracted from the geological map and mineral atlas, and the exploration data of each exploration area are stored in an independent regional server k RS and there is no sharing of the original exploration data among the regional servers k RS After that, a set of global optimization models is initialized in the central server and the model parameters of the global optimization model are randomly generated Each regional server k RS receives the model parameters of the global model for the current iteration from the central server Each regional server k RS replaces the original local model's model parameters according to the received model parameters After that, each regional server uses the local exploration data set to train the local model with the goal of minimizing the loss value of the local model, and optimizes the local model parameters by the gradient descent method. Each regional server k RS After the local training is completed, the updated model parameters are sent to the central server. The central server receives the local model parameters uploaded by each region and aggregates them through the weighted average algorithm to generate new global model parameters, and evaluates the performance indicators of the global model through historical mineral data. If the performance indicators of the global model do not reach the preset threshold, the aggregated global model parameters are distributed back to each regional server k RS to continue the next round of local training and aggregation, and iterate multiple rounds repeatedly until the performance of the global model converges to the preset threshold, and the trained global model is applied to the virtual geological scenario model and the lightweight model, and the mineral distribution under different geological conditions simulated by each model is optimized.

[0095] Embodiment 2

[0096] The embodiment of the present invention provides an intelligent planning method for a geological and mineral exploration analysis model. Refer to Figure 1 , Figure 1 which is a framework diagram of an intelligent planning method for a geological and mineral exploration analysis model provided by the embodiment of the present invention. The intelligent planning method includes the following steps:

[0097] Search and optimize the mineral distribution area according to the prediction result, evaluate the benefits of different exploration paths, and select the mineral exploration path.

[0098] Specifically, collect the optimized mineral distributions under different geological conditions and use them as the search space. Based on this search space, construct a search tree, and use this search space as the root node of the tree. Each child node in the tree represents the mineral distribution under a certain geological condition in the search space. At the same time, initialize the evaluation values and access times of each node in the search tree. Starting from the root node, calculate the UCB1 values of each child node, and according to the upper confidence bound selection strategy, sequentially select the child node with the highest UCB1 value until reaching an incompletely expanded child node, that is, the currently unexplored sub-region. Then, expand new child nodes by dividing the sub-region of the current child node into smaller sub-regions or adjusting the value range of geological parameters. In the corresponding search space of the newly added child nodes, randomly generate a set of geological parameters and calculate their corresponding mineralization potential. Propagate the simulation results backward from the current node upward to update the evaluation values and access times of each node. Repeat the steps of selection, expansion, simulation, and backtracking until the mineralization potential converges within the preset threshold. After the search stops, traverse each node in the search tree and select the child node with the highest evaluation value as the optimal mineral distribution area, and plan the exploration path based on the selected groups of mineral distribution areas.

[0099] Dynamically optimize the exploration path, apply the optimized exploration path to the actual exploration work, verify the exploration results, and dynamically adjust the geological exploration model parameters according to the measured data.

[0100] Specifically, aiming to minimize the total path cost and maximize the path value, model the optimized exploration path as a combinatorial optimization problem, and generate a set of initial paths based on the selected groups of mineral distribution areas. Each path p Tpc is a set of individuals in the population P path . Evaluate the fitness of each path through the objective function F Tpc (p Tpc ) = V Tpc (p Tpc ) - λC Tpc (p Tpc ), where F Tpc (p Tpc ) represents the fitness value of path p Tpc , V Tpc (p Tpc ) represents the sum of the mineralization potential and information gain of path p Tpc , C Tpc (p Tpc ) represents the total cost of path p Tpc , λ represents the weight of the cost. Select the path p Tpc whose fitness F Tpc (p Tpc ) meets the preset requirement value for cloning to generate a cloning pool, and adjust the path p TpcMutate each group of paths in the clone pool by the geological point order on [object] or inserting new geological points, and record the mutated paths as p'. Tpc Select paths that meet the preset demand value from the mutated clone pool and the initial population to form a new generation of population Select the new population For each path p in Tpc and its corresponding fitness F Tpc (p Tpc ), then construct a GP model as a surrogate model, use the GP model to calculate the mean prediction and variance prediction of each group of paths p Tpc , then calculate the standardized improvement amount of each path p Tpc according to the mean prediction and variance prediction, and then calculate the expected improvement of the corresponding path p Tpc through the cumulative distribution function and probability density function of the standard normal distribution. Use the grid search method to select the path p with the maximum expected improvement in the population Tpc , calculate the fitness F Tpc of the selected path p Tpc (p Tpc ), and add it to the historical data set. Then, based on the updated historical data set, retrain the GP model, update the mean prediction and variance prediction, repeat the expected improvement calculation and historical data set update until the improvement amplitude of the fitness F Tpc (p Tpc ) is lower than the preset threshold. Recalculate the fitness F Tpc of each group of paths p Tpc (p Tpc ), and re-select and mutate each group of paths p Tpc . Then, optimize the expected improvement of each path p Tpc in the updated population through the GP model, and perform repeated update iterations until the fitness value F Tpc (p Tpc ) converges to the preset index, and select the path with the highest fitness as the optimal exploration path.

[0101] The above has described an embodiment of the present invention in detail, but the described content is only a preferred embodiment of the present invention and cannot be considered as limiting the scope of implementation of the present invention. All equivalent changes and improvements made according to the scope of the present invention application should still fall within the scope covered by the patent of the present invention.

Claims

1. An intelligent planning method for a geological and mineral exploration analysis model, characterized in that: The specific steps of the intelligent planning method are as follows: S101: Obtain regional geological and mineral data through satellite remote sensing, ground surveys and historical data collection, and pre-process the collected geological and mineral data; S102: constructing a geological mineral map based on the preprocessed data, and constructing a virtual geological scene through simulation learning; S103: Integrate data from multiple mining areas, extract characteristic information from different types of geological data, and then construct a geological exploration model to predict the distribution area of ​​mineral resources; S104: Search and optimize the mineral distribution area based on the prediction results, evaluate the benefits of different exploration paths, and select the mineral exploration path; S105: Dynamically optimize the exploration path, apply the optimized exploration path to actual exploration work, verify the exploration results, and dynamically adjust the geological exploration model parameters based on the measured data.

2. The intelligent planning method of a geological and mineral exploration analysis model according to claim 1 is characterized in that: The specific steps of pre-processing the collected geological and mineral data in S101 are as follows: S1.1: Collect multispectral remote sensing, thermal infrared remote sensing and microwave remote sensing data through satellite remote sensing technology, then perform wavelet transform on each group of remote sensing data and retain the main signal components to obtain smoothed remote sensing data. The specific calculation formula of wavelet transform is as follows: In the formula, f WT (x WT ) represents the denoised remote sensing data; represents the wavelet basis function; represents the wavelet coefficient; j WT With k WT represent scale and location parameters respectively; S1.2: Collect various geological survey data on lithology, geological structure and fault zones through ground survey, sort different types of geological survey data from small to large, divide each group of sorted geological survey data into 4 parts at a ratio of 25% using the quartile algorithm, and calculate the first quartile Q of the data Bp1 and the third quartile Q Bp3 ; S1.3: Based on the first quartile Q Bp1 and the third quartile Q Bp3 , calculate the interquartile range IQR = Q of various geological measurement data Bp3 -Q Bp1 , then Q Bp3 +1.5×IQR is set as the upper limit of the normal data range, and Q Bp1 -1.5×IQR is set as the lower limit of the normal data range, and all data in the geological survey data that are less than the lower limit or greater than the upper limit are identified, regarded as outliers and eliminated. After eliminating the outliers, the distribution of the data is re-examined; S1.4: Standardize the remote sensing data, geological survey data and historical data, convert the data to the same scale, and then use the inverse distance weighted method to fill in the missing areas in the remote sensing data or the blank points in the geological survey. Then, use the principal component analysis method to fuse the geological and mineral data from different data sources. The specific calculation formula of the inverse distance weighted method is as follows: In the formula, Represents the i T interpolation weights; p ide represents the weight index; and Represents the i T and j T The distance between the points to be completed and the known points; n ide represents the total number of known points; f IDE (x IDE ) represents the value of the point to be completed; represents the value of a known point; Represents the i T The coordinates of a known point.

3. The intelligent planning method of a geological and mineral exploration analysis model according to claim 1 is characterized in that: The specific steps of constructing the geological mineral map described in S102 are as follows: S2.1: Use NER technology to automatically identify the mining areas, mineral types, geological structures, mineralization conditions, mineral deposits and strata in the preliminary processed geological and mineral data, as well as the attributes of ore body depth, grade, reserves and geological age; S2.2: The rule matching method based on dependency syntax extracts the logical relationship between entities from geological and mineral data, calculates the similarity between each group of entities, removes redundant entities with similarity higher than the preset threshold, and converts the identified entities and their relationships into triples (h MA ,r MA ,t MA ), namely "head entity-relationship-tail entity", and store the constructed triples in the graph database to generate the corresponding geological mineral map, where entities are nodes and relationships are edges; S2.3: Based on the path search algorithm and logical rule reasoning, the inference probability between each entity is calculated, and the credibility of the implicit relationship is judged according to the inference probability between each entity. If the inference probability is higher than the preset threshold, it is judged that there is an implicit relationship between the two groups of entities and recorded. The specific calculation formula of the inference probability is as follows: Where P Ip (h MA →t MA ) represents the de novo entity h MA To the end entity t MA The inference probability of MA Represents the de novo entity h MA To the end entity t MA The set of all paths of MA Represents the de novo entity h MA To the end entity t MA A path between |p MA | represents the path p MA Length; Represents path p MA Previous Ip a relationship; S2.4: Collect new geological and mineral data and exploration results in real time, and continuously update the entities and relationships in the geological and mineral maps, display the geological and mineral maps through visualization tools, and provide interactive query and analysis.

4. The intelligent planning method of a geological and mineral exploration analysis model according to claim 3 is characterized in that: The specific steps of constructing a virtual geological scene through simulation learning in S102 are as follows: S3.1: Extract various prior knowledge of geological mineral distribution from the constructed geological mineral map, including the association rules between geological structure, mineral type and mineralization conditions, and express the extracted prior knowledge in the form of triples, and then generate an initial virtual geological scene model based on the extracted prior knowledge; S3.2: Generate multiple groups of different geological condition parameter combinations through random sampling, input the sampled geological condition parameters into the virtual geological scene model, and calculate the mineral distribution under each group of geological conditions through physical equations and geological laws based on real-time geological mineral data and historical mineral data to simulate the mineral distribution results. Compare the simulation results with the geological mineral map or actual exploration data, and calculate the simulation error of the virtual geological scene model. The specific calculation formula of the simulation error is as follows: Where VG Error represents the average error of the simulation results; n Gp The number of geological condition parameter groups representing the sampling; Represents the i Gp The mineral distribution output by the simulation model under the set of geological condition parameters; Represents the i Gp The actual mineral distribution under the geological conditions and parameters of the group; S3.3: According to the calculated simulation error, the geological parameters in the simulation model are optimized by numerical simulation methods to adjust the prediction ability of the model, and the geological parameters are iteratively updated with the goal of minimizing the simulation error until the simulation error of the virtual geological scene model converges to a preset range. After that, the available computing resources of each mineral exploration equipment are collected, and a lightweight model based on the trained virtual geological scene model is constructed according to the available computing resources of each mineral exploration equipment; S3.4: Input the historical mineral data into the trained virtual geological scene model and the lightweight model respectively, and use the true value of the historical mineral data and the output of the trained virtual geological scene model as hard targets and soft targets respectively. Then, construct a total loss function to calculate the total loss of the lightweight model output and the hard targets and soft targets. The specific calculation formula of the total loss function is as follows: L hard =||z Sd -y Sd || 2 L 总 =αL soft +(1-α)L hard In the formula, Represents the virtual geological scene model generated by the i Hg Soft target of mineral data; Represents the i-th output of the virtual geological scene model Hg The score of each mineral data; T tp represents the temperature parameter, which is used to smooth the probability distribution; Represents the jth output of the virtual geological scene model Hg The score of each mineral data; L soft represents the soft target loss; Represents lightweight model i Hg The predicted probability of mineral data; L hard represents the hard target loss; z Sd Represents the predicted value of the lightweight model; y Sd represents the true value of the actual mineral distribution; L 总 represents the total loss function; α represents the trade-off parameter; S3.5: The calculated total loss is back-propagated, and the parameters of the lightweight model are optimized based on the back-propagation results. After the parameter optimization is completed, the prediction accuracy of the mineral distribution of the lightweight model and the calculation efficiency of the model are evaluated on the unused historical mineral data. According to the verification results, the model structure or the hyperparameters in the distillation process are adjusted, and the lightweight model is repeatedly trained and verified until the total loss value of each lightweight model converges to the preset threshold; S3.6: According to the available computing resources of different mineral exploration equipment, the virtual geological scene model and the lightweight model are deployed on the corresponding equipment respectively, and the real-time geological and mineral data are input into each model. Through forward propagation, a virtual geological scene containing the distribution of minerals under different geological conditions is generated, and the generated geological scene is displayed through a graphical interface. At the same time, the parameters of each model are updated in real time according to the actual exploration data.

5. The intelligent planning method of a geological and mineral exploration analysis model according to claim 4 is characterized in that: The specific steps of constructing a geological exploration model and predicting the mineral distribution area in S103 are as follows: S4.1: Extract exploration areas with different geological characteristics based on the geological map and mineral map, and store the exploration data of each exploration area on an independent regional server k RS In each regional server k RS The original survey data is not shared between them, and a set of global optimization models are initialized in the central server. And randomly generate the global optimization model Model parameters of S4.2: Each regional server k RS Receive the global model of the current iteration from the central server The model parameters of each regional server k RS Replace the original local model according to the received model parameters Then, each regional server trains the local model through the local survey data set with the goal of minimizing the loss value of the local model, and optimizes the local model parameters through the gradient descent method; S4.3: Each region server k RS After local training is completed, the updated model parameters The central server receives the local model parameters uploaded by each region and aggregates them through a weighted average algorithm to generate new global model parameters. S4.4: Evaluate the performance indicators of the global model through historical mineral data. If the performance indicators of the global model do not reach the preset threshold, the aggregated global model parameters are distributed back to each regional server k RS , continue the next round of local training and aggregation, and repeat multiple rounds of iterations until the performance of the global model converges to the preset threshold. The trained global model is applied to the virtual geological scene model and the lightweight model, and the mineral distribution under different geological conditions simulated by each model is optimized.

6. The intelligent planning method of a geological and mineral exploration analysis model according to claim 5 is characterized in that: The specific steps of searching and optimizing the mineral distribution area according to the prediction results described in S104 are as follows: S5.1: Collect the optimized mineral distribution under different geological conditions and use it as the search space, build a search tree based on the search space, and use the search space as the root node of the tree. Each child node in the tree represents the mineral distribution under a geological condition in the search space, and initialize the evaluation value and access count of each node in the search tree; S5.2: Starting from the root node, calculate the UCB1 value of each child node, and select the child node with the highest UCB1 value in turn according to the upper confidence interval selection strategy until reaching the incompletely expanded child node, that is, the sub-area that has not been explored yet. Then, expand the new child node by dividing the sub-area of ​​the current child node into smaller sub-areas or adjusting the value range of the geological parameters; S5.3: In the corresponding search space of the newly added child node, a set of geological parameters is randomly generated, and the corresponding mineralization potential is calculated, the simulation results are back-propagated upward from the current node, and the evaluation value and the number of visits of each node are updated; S5.4: Repeat the selection, expansion, simulation and backtracking steps until the mineralization potential converges to the preset threshold. After the search stops, traverse each node in the search tree and select the child node with the highest evaluation value as the optimal mineral distribution area, and plan the exploration path based on the selected groups of mineral distribution areas.

7. The intelligent planning method of a geological and mineral exploration and analysis model according to claim 6 is characterized in that: The specific steps of dynamically optimizing the exploration path are as follows: S6.1: With the goal of minimizing the total cost of the path and maximizing the path value, the optimization of the exploration path is modeled as a combinatorial optimization problem, and a set of initial paths is generated based on the selected groups of mineral distribution areas. Each path p Tpc is the population P path A group of individuals in Tpc (p Tpc )=V Tpc (p Tpc )-λC Tpc (p Tpc ) evaluates the fitness of each path, where F Tpc (p Tpc ) represents the path p Tpc The fitness value, V Tpc (p Tpc ) represents the path p Tpc The sum of the mineralization potential and information gain, C Tpc (p Tpc ) represents the path p Tpc Total cost, λ represents the weight of cost; S6.2: Select fitness F Tpc (p Tpc ) The path p that meets the preset demand value Tpc Perform cloning to generate a clone pool, and adjust the path p in the clone pool Tpc The sequence of geological points on the cloning pool or the insertion of new geological points is used to mutate each group of paths in the clone pool, and the mutated path is recorded as p′ Tpc ; S6.3: Select the path that meets the preset demand value from the mutated clone pool and the initial population to form a new generation of population Select new population Each path p Tpc And its corresponding fitness F Tpc (p Tpc ), and then construct the GP model as the proxy model; S6.4: Calculate each group of paths p using the GP model Tpc The mean prediction and variance prediction of each path p are then calculated based on the mean prediction and variance prediction. Tpc The standardized improvement is then calculated through the cumulative distribution function and probability density function of the standard normal distribution to obtain the corresponding path p Tpc The expected improvement of In the above example, we select the path p with the largest expected improvement. Tpc ; S6.5: Calculate the selected path p Tpc The fitness F Tpc (p Tpc ), and add it to the historical data set, then retrain the GP model based on the updated historical data set, update the mean prediction and variance prediction, repeat the expected improvement calculation and historical data set update until the fitness F Tpc (p Tpc ) The improvement is lower than the preset threshold; S6.6: Recalculate each group of paths p in the population Tpc The fitness F Tpc (p Tpc ), and recalculate each group of paths p Tpc Select and mutate, and then optimize the updated paths p in the population through the GP model Tpc The expected improvement is repeated until the fitness value F Tpc (p Tpc ) converges to the preset index, and selects the path with the highest fitness as the optimal exploration path.

Images