A method for efficient detection and positioning of gold deposits based on the characteristics of volcanic cryptoexplosion breccia

Abstract

The application provides a gold mine efficient detection positioning method based on the characteristics of volcanic cryptoexplosion breccia, relates to the technical field of reinforcement learning, and comprises the following steps: acquiring multi-source geological data, pre-processing the multi-source geological data, performing feature extraction on the pre-processed data based on a feature extraction module to obtain a feature vector, performing screening on the feature vector based on a feature vector screening module, performing fusion on the feature vector based on a feature fusion module, and identifying the fused feature vector to realize volcanic cryptoexplosion breccia type gold mine detection. The application can automatically extract weak mineralization information from massive multi-source data, intelligently focus on the optimal target area, reduce artificial dependence and subjective misjudgment, strengthen the learning mechanism to introduce economic cost constraints, promote the exploration decision from experience-driven to benefit-driven, form a dynamic optimization closed loop, adjust the strategy according to real-time results, and greatly improve the exploration success rate and efficiency.

Description

Technical Field

[0001] This invention relates to the field of reinforcement learning technology, and in particular to an efficient method for gold mine detection and location based on the characteristics of cryptovolcanic breccia. Background Technology

[0002] Volcanic cryptovolcanic breccia-type gold deposits are an important type of gold deposit. Their formation is closely related to deep magmatic-hydrothermal cryptovolcanic activity and often occurs in favorable tectonic locations such as volcanic structures and regional faults. While these deposits have enormous economic value, their detection remains a global challenge, primarily due to the following difficulties: ore bodies are often found underground in cylindrical or vein-like forms, with few surface outcrops and weak mineralization markers, exhibiting little difference in physical properties from the surrounding rocks, making traditional geological prospecting methods ineffective; exploration target areas are typically located in remote areas with complex terrain, resulting in extremely low efficiency and high safety risks for manual exploration; modern geological exploration can acquire massive amounts of multi-source heterogeneous data (geological, geophysical, geochemical, remote sensing), but these data come from diverse sources, have different formats, and are at different scales, resulting in severe "information silos." Effectively integrating these data and extracting weak but crucial mineralization information relies heavily on the experience of individual geologists. Experience and intuition are highly subjective and uncertain. Existing computer-aided mineral prediction methods are mostly based on simple statistical analysis or shallow machine learning models, which can only handle single types of data or perform linear analysis. They lack the ability to characterize complex nonlinear geological processes, cannot deeply integrate multi-source information, and have limited prediction accuracy and reliability. The current exploration model is an "open-loop" system, that is, first interpreting and deploying exploration projects based on existing data, and then adjusting the understanding based on the project results. This model is time-consuming, costly, and cannot dynamically optimize and adjust the overall exploration strategy in real time based on newly acquired data, often leading to blind exploration deployment and huge waste of funds and time. Therefore, the field of mineral exploration urgently needs a new method that can automatically, intelligently, and efficiently integrate multi-source data, can imitate expert thinking for reasoning and decision-making, and ultimately dynamically optimize exploration strategies. In summary, it is essential to design an efficient gold deposit detection and location method based on the characteristics of cryptovolcanic breccia. Summary of the Invention

[0003] To overcome the shortcomings of existing technologies, the purpose of this invention is to provide an efficient method for gold mine detection and location based on the characteristics of cryptovolcanic breccia.

[0004] To achieve the above objectives, the present invention provides the following solution:

[0005] This invention provides a highly efficient method for gold deposit detection and location based on the characteristics of cryptovolcanic breccia, including:

[0006] Step 1: Acquire multi-source geological data;

[0007] Step 2: Preprocess the multi-source geological data;

[0008] Step 3: Extract features from the preprocessed data using the feature extraction module to obtain feature vectors;

[0009] Step 4: Filter the feature vectors based on the feature vector filtering module;

[0010] Step 5: Fuse the feature vectors based on the feature fusion module;

[0011] Step 6: Identify the fused feature vectors to achieve the detection of gold deposits in volcanic rock cryptovolcanic breccia.

[0012] Preferably, the multi-source geological data includes geological and geographical data, geophysical data, geochemical data, and remote sensing data;

[0013] The geological and geographical data includes digital geological maps, structural maps, and digital elevation models;

[0014] The geophysical data includes gravity data, magnetic data, magnetotelluric data, and seismic data;

[0015] The geochemical data includes rock geochemical measurement data, soil geochemical measurement data, and elemental content and distribution data;

[0016] The remote sensing data includes multispectral data, hyperspectral remote sensing data, and InSAR data.

[0017] Preferably, in step 2, the multi-source geological data is preprocessed, specifically as follows:

[0018] Acquire multi-source geological data;

[0019] Noise removal, distortion correction, and normalization were performed on it.

[0020] Spatial registration and gridding are performed on the processed multi-source geological data.

[0021] Preferably, in step 3, feature extraction is performed on the preprocessed data based on the feature extraction module to obtain a feature vector, specifically as follows:

[0022] A feature extraction module is constructed based on dilated convolution;

[0023] The processed multi-source geological data is input into the feature extraction module to obtain feature vectors.

[0024] Preferably, in step 4, the feature vectors are filtered based on the feature vector filtering module, specifically as follows:

[0025] A feature vector filtering module is constructed based on the MaxMin-DQN network;

[0026] The feature vector filtering module sorts the feature vectors to determine the foreground feature vector set and the background feature vector set.

[0027] Preferably, a feature vector selection module is constructed based on the MaxMin-DQN network, specifically as follows:

[0028] Define an agent to select a set of foreground feature vectors that contribute highly to the detection from the feature vectors. Define the agent's action as selecting a set of foreground feature vectors from the candidate regions. Define the set of all current feature vectors as the state of the environment. Establish a MaxMin-DQN network algorithm mechanism. By building multiple DQN networks with the same structure, each DQN network participates in scoring the feature vectors in the iteration. Calculate the scores of multiple feature vectors selected by each DQN network through reinforcement learning and accumulate them. Select the DQN network with the smallest accumulated score as the benchmark and update all networks.

[0029] Preferably, in step 5, the feature vectors are fused based on the feature fusion module, specifically as follows:

[0030] The feature vector fusion module processes the foreground feature vector set and the background feature vector set separately, and then fuses them to obtain a fused feature vector set.

[0031] Preferably, in step 6, the fused feature vector is identified to achieve the detection of gold deposits in volcanic cryptovolcanic breccia type, specifically as follows:

[0032] Based on the Transformer module, the fused feature vector set is identified to realize the detection of gold deposits in volcanic rock cryptovolcanic breccia.

[0033] According to specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0034] This invention provides an efficient method for gold deposit detection and location based on the characteristics of cryptovolcanic breccia. The method includes acquiring multi-source geological data, preprocessing the multi-source geological data, extracting features from the preprocessed data using a feature extraction module to obtain feature vectors, filtering the feature vectors using a feature vector filtering module, fusing the feature vectors using a feature fusion module, and identifying the fused feature vectors to achieve the detection of gold deposits in cryptovolcanic breccia. This invention first achieves intelligent and automated exploration processes. By using deep convolutional networks and Transformer models, it automatically extracts and fuses deep features from massive, multi-source geological data, significantly reducing reliance on the personal experience of geological experts and overcoming the subjectivity and uncertainty of manual interpretation. This transforms mineral prediction from an "art" into a quantifiable and reproducible "science." Secondly, the core innovation is the MaxMin-DQN reinforcement learning intelligent screening mechanism introduced in this invention. It simulates the thought process of experienced exploration experts—"separating the wheat from the chaff, focusing on the most promising areas"—adaptively focusing on a few target areas most likely to form mineral deposits from thousands of grid cells. Furthermore, the cleverly designed "cost penalty" term in the reward function forces the agent to learn to weigh information value against economic costs, thereby recommending a "few but precise" optimal target areas. This avoids blind exploration from the outset, significantly reducing exploration costs and time, and improving exploration efficiency. Thirdly, at the technical level, the use of dilated convolutional neural networks (DetNet) can extract a wide range of upper and lower features while maintaining high resolution. The method is highly suitable for capturing annular and linear structural anomalies associated with cryptovolcanic breccia pipes, effectively avoiding the loss of small target features in the traditional convolutional downsampling process. Furthermore, the Transformer's self-attention mechanism can globally analyze the relationships between all features, accurately capturing key mineralization information combinations such as "high Au-low magnetic field-strong silicification-ring structure," resulting in predictions with significantly higher accuracy and robustness than traditional methods. Fourth, this invention constructs a dynamic closed-loop system of "perception-decision-verification," allowing the agent to update the model and optimize subsequent exploration strategies in real time based on newly acquired data such as drilling verification. This achieves dynamic adjustment through simultaneous exploration, learning, and optimization, completely changing the traditional static, open-loop exploration model and making the entire exploration process as precise and efficient as "surgery." Finally, this method has strong universality and scalability. Its technical framework is not only applicable to cryptovolcanic breccia-type gold deposits but, with adaptive adjustments, can also be widely applied to the exploration of other types of concealed deposits, providing a powerful tool for promoting technological transformation and intelligent upgrading in the mineral exploration industry. Attached Figure Description

[0035] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0036] Figure 1 This is a schematic diagram of the method flow of the present invention;

[0037] Figure 2 This is a schematic diagram of a holed convolutional block structure.

[0038] Figure 3 This is a schematic diagram of DetNet stage 5 and stage 6 structures. Detailed Implementation

[0039] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0040] The purpose of this invention is to provide a highly efficient gold deposit detection and location method based on the characteristics of cryptovolcanic breccia. This method can automatically and accurately extract weak mineralization information from massive multi-source data, intelligently focus on the optimal target area, and reduce reliance on manual labor and subjective misjudgment. The reinforcement learning mechanism introduces economic cost constraints, promoting exploration decision-making from experience-driven to benefit-driven. A dynamic optimization closed loop is formed, and the strategy is adjusted according to real-time results, which greatly improves the success rate and efficiency of exploration.

[0041] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0042] like Figure 1 As shown, this invention provides a highly efficient method for gold deposit detection and location based on the characteristics of cryptovolcanic breccia, comprising:

[0043] Step 1: Acquire multi-source geological data;

[0044] Step 2: Preprocess the multi-source geological data;

[0045] Step 3: Extract features from the preprocessed data using the feature extraction module to obtain feature vectors;

[0046] Step 4: Filter the feature vectors based on the feature vector filtering module;

[0047] Step 5: Fuse the feature vectors based on the feature fusion module;

[0048] Step 6: Identify the fused feature vectors to achieve the detection of gold deposits in volcanic rock cryptovolcanic breccia.

[0049] The multi-source geological data includes geological and geographical data, geophysical data, geochemical data, and remote sensing data;

[0050] The geological and geographical data includes digital geological maps, structural maps (especially fault and ring structures), and digital elevation models (DEMs).

[0051] The geophysical data includes gravity data (capturing density differences and identifying concealed rock masses), magnetic data (identifying magnetic anomalies and delineating alteration zones), magnetotelluric / audio-magnetic (AMT / CSAMT) data (characterizing resistivity structures and identifying ore-bearing fluid channels), and seismic data.

[0052] The geochemical data includes rock geochemical measurement data, soil geochemical measurement data, and element content and distribution data (e.g., obtaining the content and distribution of elements such as Au, Ag, As, Sb, Hg, Cu, Pb, and Zn).

[0053] The remote sensing data includes multispectral data (such as Landsat-8, ASTER), hyperspectral remote sensing data (for extracting alteration minerals such as sericite, kaolinite, limonite, etc.), and InSAR data (for identifying tectonic activity).

[0054] In step 2, the multi-source geological data is preprocessed, specifically as follows:

[0055] Acquire multi-source geological data;

[0056] Noise removal, distortion correction, and normalization were performed on it.

[0057] Spatial registration and gridding are performed on the processed multi-source geological data to unify all data into the same coordinate system (such as WGS84 UTM) and the same grid scale (such as 100m x 100m or 50m x 50m grid cells). Each grid cell will contain multidimensional features from different data sources, forming a geoscience data cube.

[0058] In step 3, feature extraction is performed on the preprocessed data using the feature extraction module to obtain feature vectors, specifically as follows:

[0059] A feature extraction module is constructed based on dilated convolution;

[0060] The processed multi-source geological data is input into the feature extraction module to obtain feature vectors.

[0061] The invention is described in detail as follows: It uses dilated convolution instead of traditional convolution. Dilated convolution, also known as dilated convolution or dilated convolution, is a convolution method improved based on traditional convolution. Dilated convolution changes the receptive field by setting different dilation ratios. The dilation ratio minus one represents the number of spaces added to the convolution kernel. By setting different dilation ratios, different sizes of receptive fields can be achieved, which can better preserve spatial hierarchical information, process small object target information, and avoid information loss.

[0062] Dilated Residual Networks (DetNet) are feature extraction networks specifically designed for detection. Based on the basic structure of Residual Networks (ResNet), they replace ordinary convolutions with dilated convolutions, increasing network depth while maintaining feature map resolution. This allows for better preservation of features and location information of small targets. The network incorporates two novel convolutional block structures. First, as... Figure 2 As shown, the traditional 3×3 convolution is replaced with a 3×3 dilated convolution with a dilation ratio of 2 on the basis of the ResNet convolution block, which increases the range of the receptive field and is called dilated convolution block A. Then, the residual structure is modified on this basis, and 1×1 convolution is added to the residual branch to further enhance the nonlinear ability of the network, so that the network can better express complex features. This is called dilated convolution block A.

[0063] The overall structure of DetNet is similar to that of ResNet. The first four stages are exactly the same as ResNet. Starting from the fifth stage, it uses innovative dilated convolutional blocks to continue extracting features, and a sixth stage is added to continue extracting features. The structure of the fifth and sixth stages is as follows: Figure 3 As shown, the number of channels in stages 5 and 6 remains unchanged at 256, saving memory and computing resources. Stages 5 and 6 have the same structure, consisting of a dilated convolutional block B followed by two dilated convolutional blocks A. The DetNet network can increase the receptive field without increasing the amount of computation. In order to further improve the ability to represent complex features, the feature maps output from stages 4 to 6 are fused to output a feature map with a resolution of 14×14. The structure diagram is shown in Table 1.

[0064] Table 1 DetNet Network Structure

[0065]

[0066]

[0067] After feature extraction, a set of feature vectors F is obtained.

[0068] In step 4, the feature vectors are filtered based on the feature vector filtering module, specifically as follows:

[0069] A feature vector filtering module is constructed based on the MaxMin-DQN network;

[0070] The feature vector filtering module sorts the feature vectors to determine the foreground feature vector set and the background feature vector set;

[0071] First, let's introduce reinforcement learning. Reinforcement learning is an important method in machine learning that solves specific tasks through the interaction between an agent and its environment. In reinforcement learning, the agent observes the environment and takes corresponding actions based on a policy. After each action, the environment changes accordingly, and the agent receives feedback from the environment to adjust its policy, continuously learning to obtain the optimal solution. Compared to supervised and unsupervised learning, reinforcement learning is more flexible, as it doesn't require pre-provided data but learns its policy through interaction with the environment. Reinforcement learning can be divided into two types: probability-based and value-based methods. Probability-based methods are one of the most common methods in reinforcement learning. The agent observes and analyzes data in the environment, outputting the probabilities of various actions for the next step. Then, based on the probability of each action, the agent selects the most likely action to execute. In contrast, value-based methods analyze the value of actions to provide suggestions for the agent's actions. In this method, each action is assigned a value, and selection is based on its value. This method is typically used in discontinuous, discrete environments. It struggles to achieve good learning results for tasks with large sets of actions or continuous actions; in these cases, probability-based methods perform better.

[0072] In reinforcement learning, the reward function determines the reward signal received by the agent during its interaction with the environment. By maximizing the reward signal, the agent can achieve better performance. Therefore, designing a reward function that accurately reflects environmental feedback is crucial in reinforcement learning algorithms. However, designing a reward function is not easy. Because the agent's behavior is based on the reward signal, an inappropriate reward function can lead to the agent learning incorrect behaviors. For example, in a robot-controlled task, if the reward function only rewards the robot's position, the robot might discover that falling down maximizes the reward instead of maintaining balance. Therefore, designing a suitable reward function requires careful consideration of the task objective, as well as the agent's learning speed and stability. Furthermore, the environment and policy also influence the success of reinforcement learning. The environment determines the agent's situation, including the state of the environment and available actions. The policy is the rule governing the agent's decision-making, determining what action the agent should take in a given state. Therefore, to ensure the success of reinforcement learning, it is necessary to understand the characteristics of the environment and the strengths and weaknesses of the policy. Based on this understanding, a better reward function can be designed, enabling the agent to perform the task more effectively. In conclusion, designing a suitable reward function, understanding the environment, and designing excellent policies are all crucial factors in reinforcement learning. Although these factors are complex, the effectiveness of reinforcement learning can only be maximized when their influence is taken into account.

[0073] Reinforcement learning algorithms have a wide range of applications. In the field of autonomous driving, reinforcement learning is used to design autonomous decision-making systems for vehicles, enabling them to react to complex traffic environments. In industrial automation, reinforcement learning is also used to design intelligent control systems to automate equipment operations. In game development, reinforcement learning has been used to train robots to defeat human players in video games. With the continuous development of artificial intelligence technology, reinforcement learning will play an increasingly important role in future applications.

[0074] Reinforcement learning is an interactive, cyclical process in which an agent's actions change the state of the environment and earn rewards. This process can be described by a Markov decision process (MFD), where the agent makes the optimal action based on the current information state, and the future state of the environment depends only on the current state and is independent of historical states. MFDs are widely used in reinforcement learning problems involving both discrete and continuous states, and can be represented as a quadruple. <S,A,R,P s，a >

[0075] S represents a finite set of states, where each state s tLet S represent the environmental state of the agent at a specific time t. These states are abstractly represented and contain all the information needed to make a decision. Once the current state is determined, all historical states can be discarded because reinforcement learning has the Markov property, which states that the future state of the environment depends only on the current state and is independent of historical states. Therefore, state st is a very important concept and is the fundamental information for the agent to make decisions.

[0076] A represents the set of actions, where each action a∈A refers to the action that the agent can take in the current environmental state. The actions the agent can take may differ in each state, therefore the action set A may vary across different states. By selecting appropriate actions in each state st, the agent can continuously change the environmental state and obtain reward feedback from the environment, thereby optimizing its strategy. Action a is a crucial component of the interaction between the agent and the environment and is one of the key elements in realizing the reinforcement learning process.

[0077] R represents the set of rewards that the agent receives through continuous interaction with the environment, used to encourage the agent to take beneficial actions. In each state s t After the agent takes action 'a', it will enter the next state St+1, at which point the agent will receive a reward. The reward function is usually represented as r(s). t+1 |s, a), is determined by the result of environmental feedback, in the transition state s t+1 When the reward is unique, it can be simplified as r(s, a). When calculating the reward, the rewards obtained from past actions are usually weighted by a discount factor γ∈[0,1] and then summed. The smaller the value of the discount factor γ, the more emphasis is placed on the reward of the current action. Therefore, when designing the reward function, it is necessary to consider the magnitude of the reward and the value of the discount factor γ to maximize the reward obtained by the agent in the long run. The design of the reward function is a very critical issue in reinforcement learning.

[0078] P represents the probability function of state transition, which describes the state transition from state s at time t. t Take action t Then it enters the next state S. t+1 The probability is expressed by P(S) t+1 |S t a tThe probability P is represented by a variable P. This probability value is an important parameter, crucial for determining the efficiency and accuracy of reinforcement learning algorithms. In most practical reinforcement learning tasks, the probability is unknown and must be learned from the data. To effectively estimate the probability P, it is usually necessary to collect a large amount of empirical data and use machine learning algorithms to fit the model. Through accurate modeling, the future state of the agent can be better predicted, and optimal actions can be taken in different states, thus achieving better reinforcement learning results.

[0079] A Markov Decision Process (MDP) is a model for an agent to make optimal behavioral decisions in an environment. For each state *s* in which the agent is located, an optimal action, or policy, can be determined through a Markov Decision Process, denoted by π(s). In the agent's decision-making process, long-term reward is an important objective; that is, the agent's behavior should aim to maximize long-term reward.

[0080] Therefore, maxE[∑tγ] t R t (S t a t [)] represents the agent's optimization objective. To better describe the agent's behavior in a Markov decision process, we need to define the agent's value function as V(s). With policy π, the expected reward in state s is as follows:

[0081]

[0082] Recursively expanding this value function, we get:

[0083] V π (s)=R(s,π(s))+γ∑ s′∈S P(s|s′,a)V π (s′) (2)

[0084] Where R(s, π(s)) represents the immediate reward obtained after choosing an action in state s using policy π(s), γ represents the discount factor, P(s|s', a) represents the probability of entering the next state s after taking action a from state s' at time t, and the value function can evaluate the value of different states to make the optimal decision. To obtain the maximum cumulative reward, the optimal policy π*(s) needs to be adopted, which is:

[0085] π * (s)=arg max r(s,a)+γ∑ s′∈S P(s|s′,a)V π (s′) (3)

[0086] This optimal policy can be obtained through a value function iterative algorithm. Specifically, starting from an initial policy, iterative steps are taken until the policy converges, thus obtaining the optimal policy. This allows the agent to take the optimal action in any situation, maximizing long-term rewards. In Markov decision-making, the value function V(s) and the policy π(s) are closely related. By evaluating the state value and adopting the optimal policy, the agent can make optimal behavioral decisions in the environment, thereby maximizing long-term rewards.

[0087] The goal of the agent is to select the set of foreground feature vectors that contribute significantly to the final object detection from the feature vector set. The agent's action 'a' is defined as selecting a set of foreground feature vectors of size N from among the candidate regions. This includes:

[0088] Action 1 (Retain): Determine that the feature vector belongs to the foreground (related to mineralization);

[0089] Action 0 (Discard): Determine that the feature vector belongs to the background (unrelated to mineralization);

[0090] The set of all current feature vectors Defined as the state of the environment;

[0091] Reward: Designing a reward function is crucial for the success of reinforcement learning. This invention designs a reward based on the feedback of the final exploration results:

[0092] R pos (Positive Reward): If a grid cell containing a feature vector that is retained (Action 1) is subsequently drilled and verified to have gold mineralization of economic grade, a large positive reward is given.

[0093] R neg (Negative Reward): If a grid cell containing a feature vector that was retained (Action 1) is found to be free of minerals after drilling, a negative reward (penalty) is given.

[0094] R cost (Cost Penalty): For each feature vector selected to be retained, a small fixed negative reward is deducted. This simulates the economic cost of exploration, forcing the agent to learn to "be frugal" and only select the target area with the most confidence, avoiding blindly retaining too many areas.

[0095] This invention uses a reinforcement learning DQN network to score and filter feature vectors, then uses the resulting scores as modulation factors for each feature vector, updating them along with the backbone network. A scoring network (DQN) scores each feature vector, and the Q-value for each region is δ. n，wThe DQN network consists of two fully connected layers:

[0096] δ h，w =DQN(f h，w (4)

[0097] Q value δ n，w The larger the value, the better the eigenvector f. i，j The higher the value of the selected score, the more all the scores {f} will be allocated. n，w After sorting, we obtain [δ] l [l = 1, ..., 196], select the N vectors with the highest scores to construct the foreground feature vector. for:

[0098]

[0099] After establishing the foreground feature vector set, the scores obtained from the foreground feature vectors are used as weighting factors for the foreground feature set. A layer normalization operation is then performed on the feature vectors to eliminate gradient vanishing during training, as follows:

[0100]

[0101] When a single neural network estimates the reward value of the Q-function, the randomness during each training process can lead to biases in the network loss. To reduce the impact of estimation bias on the training convergence speed, the MaxMin-DQN network algorithm mechanism is established.

[0102] By establishing ρ DQN networks with identical structures ρ (1, ..., ρ), in the iteration, each DQN network participates in scoring the feature vectors. Each feature vector is calculated to obtain a different score value. The scores of the N feature vectors selected by each DQN network through reinforcement learning are accumulated, and the DQN network with the smallest cumulative score value is selected as the benchmark. All networks are updated. The calculation formula is as follows:

[0103] DQN ρ ←DQN ρ +α[Y MQ -DQN ρ (7)

[0104] α represents the predetermined step size, Y MQ This represents the vector network of the DQN network with the smallest cumulative score.

[0105] In step 5, the feature vectors are fused based on the feature fusion module, specifically as follows:

[0106] The feature vector fusion module processes the foreground and background feature vector sets separately, and then fuses them to obtain a fused feature vector set, specifically:

[0107] After reinforcement learning, a foreground feature set was selected. The remaining feature vectors mainly correspond to background feature vectors, which are numerous but contain relatively little effective information. During detection, it is necessary to retain useful feature information from the background context to the maximum extent possible without reducing detection accuracy. This invention designs a pooling sampler based on a bilinear attention transformation mechanism, which can compress the background features to obtain a fixed number of M background context feature vectors. This reduces the proportion of redundant background information in the feature set, decreasing computational load while improving detection performance. The remaining feature vectors after reinforcement learning are... It is expressed as follows:

[0108]

[0109] Bilinear attention mechanism transformation can reduce the number of vectors while fully preserving the information of interest in the feature vectors. Its goal is to compress the number of background region features and retain key information. First, a weight matrix W is used. O Get f γ Aggregate weight vector O γ :

[0110] O r =f r W O (9)

[0111] Among them, W O The dimension is C×M, O γ The dimension is 1×M, and f is obtained using a weight matrix. γ The projected eigenvector f' γ :

[0112] f′ r =f r Wv (10)

[0113] Among them, W v The dimension is C×C, and the aggregate weight vector O γ Softmax normalization is performed on each corresponding element in the data:

[0114]

[0115] Using the normalized aggregate weight vector, for f γ By aggregating the projected feature vectors, the background region feature vector f can be obtained. m :

[0116]

[0117] The compressed set of background region feature vectors is

[0118]

[0119] The pooling sampler dynamically generates aggregation weights to freely acquire background context of varying sizes. A portion of the feature vector captures local context, while the other portion encodes global context information. Combined with the foreground feature set F selected through reinforcement learning, the resulting concatenation yields the abstract feature vector set to be fed into the Transformer encoding layer.

[0120] In step 6, the fused feature vectors are identified to achieve the detection of gold deposits in volcanic cryptovolcanic breccia type, specifically as follows:

[0121] Based on the Transformer module, the fused feature vector set is identified to realize the detection of gold deposits in volcanic cryptovolcanic breccia.

[0122] An introduction to the Transformer module:

[0123] The Transformer module uses the classic Transformer architecture, consisting of several encoding layers and several decoding layers. Each encoding layer has the same structure, comprising a multi-head self-attention mechanism and a feedforward network. Feature vectors from the abstract feature vector set F* are input into the encoding layer. Since spatial structure is lost after inputting into the Transformer, positional encoding is obtained through fixed encoding after feature extraction by the backbone feature network. After positional encoding is calculated, it is processed by the feature fusion module, just like the feature vector set, to obtain abstract positional encoding information, which is then added to each encoding layer. The decoding part is the same as the classic Transformer decoding part, also consisting of several decoding layers with the same structure. Multiple target queries are decoded in parallel at each decoder layer, and these target queries are also input into each decoding layer. Finally, the output of the decoding part outputs the final prediction result through a fully connected linear layer.

[0124] The Transformer module uses existing technology, so it will not be described in detail here.

[0125] After obtaining high-probability areas or recommended drilling targets based on the Transformer module, experts are organized to verify their rationality. If the verification is successful, the project can be implemented.

[0126] The present invention also provides an embodiment to demonstrate the effects of the present invention:

[0127] Taking a key gold mining area in my country as an example, the exploration practice using the method described in this invention was carried out. The geological background of this area is complex, and the main ore-bearing rocks are Mesozoic volcanic rocks. Although several mineral deposits were discovered in the early exploration work, the breakthroughs in deep and peripheral exploration have not met expectations.

[0128] Step 1: Data Acquisition and Preprocessing. This involved collecting 1:50,000 geological maps, ASTER multispectral remote sensing data, 1:50,000 high-precision aeromagnetic data, 1:100,000 gravity data, 1:50,000 soil geochemical measurement data (analyzing 12 elements including Au, Ag, As, and Cu), and 12.5-meter resolution ALOSDEM data for the area. All data underwent coordinate unification (WGS84 UTM Zone 50N), gridding (50m × 50m grid cells), and normalization to form a "geological data cube" containing over 30 feature channels, covering an area of ​​approximately 400 square kilometers.

[0129] The second step is feature extraction. The processed data is input into a dilated convolutional feature extraction network based on the DetNet architecture. The network outputs a 14×14×256 feature map, which is divided into 196 grid cells, and each cell consists of a 256-dimensional feature vector f. i，j This indicates that the vector deeply integrates ground, physical, chemical, and remote sensing information at this unit.

[0130] Step 3: Reinforcement learning intelligent selection. Each of the 256-dimensional feature vectors is input into a pre-trained MaxMin-DQN network (K = 8 Q-networks). The network outputs a Q-value for each vector, representing its mineralization potential score. Finally, the system automatically selects the top 60 grid cells (i.e., 60 feature vectors) with the highest Q-values ​​as foreground target areas. Visualization shows that these cells clearly delineate the boundary of a known mineralized breccia pipe and indicate a previously overlooked northwest-trending mineralization fault extension zone.

[0131] Step 4: Feature fusion. The remaining 136 background feature vectors are compressed into 60 condensed background context vectors using bilinear attention pooling. Simultaneously, the 60 foreground vectors are weighted and modulated using Q-values ​​to enhance their importance. Finally, the two sets are concatenated to form a fused feature set of 120 vectors.

[0132] Step 5: Mineralization Prediction and Location. Input F into the Transformer prediction module (encoder layer number 6). The model outputs a "mineralization probability map" and directly recommends the coordinates of 3 optimal drilling target points and their expected depths.

[0133] Verification results: Engineering verification showed that drilling at the recommended target point yielded mineralization near the designed depth, revealing a thick, cryptovolcanic breccia-type gold deposit.

[0134] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0135] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A highly efficient method for gold deposit detection and location based on the characteristics of cryptovolcanic breccia, characterized in that, The process includes: Step 1: Acquiring multi-source geological data; Step 2: Preprocessing the multi-source geological data; Step 3: Extracting features from the preprocessed data using a feature extraction module to obtain feature vectors.

4. Feature vector filtering module: The feature vectors are filtered based on the feature vector filtering module. Specifically, a feature vector filtering module is constructed based on the MaxMin-DQN network. Specifically, an agent is defined to select the set of foreground feature vectors that contribute highly to the detection from the feature vectors. The agent's action is defined as selecting the set of foreground feature vectors from the candidate regions. The set of all current feature vectors is defined as the state of the environment. The MaxMin-DQN network algorithm mechanism is established. Multiple DQN networks with identical structures are built. In the iteration, each DQN network participates in scoring the feature vectors. The scores of multiple feature vectors selected by each DQN network through reinforcement learning are accumulated. The DQN network with the smallest accumulated score is selected as the benchmark, and all networks are updated. The feature vectors are sorted based on the feature vector filtering module to determine the foreground feature vector set and the background feature vector set. Step 5. Feature vector fusion module: The feature vectors are fused based on the feature fusion module. Step 6. The fused feature vectors are identified to realize the detection of gold deposits in volcanic rock cryptovolcanic breccia.

2. The method according to claim 1, characterized in that, The multi-source geological data includes geological and geographical data, geophysical data, geochemical data, and remote sensing data; the geological and geographical data includes digital geological maps, structural maps, and digital elevation models; the geophysical data includes gravity data, magnetic data, magnetotelluric data, and seismic data; the geochemical data includes rock geochemical measurement data, soil geochemical measurement data, and elemental content and distribution data; and the remote sensing data includes multispectral data, hyperspectral remote sensing data, and InSAR data.

3. The method according to claim 2, characterized in that, In step 2, the multi-source geological data is preprocessed, specifically: acquiring multi-source geological data; performing noise removal, distortion correction and normalization on the data; and performing spatial registration and gridding on the processed multi-source geological data.

4. The method according to claim 3, characterized in that, In step 3, feature extraction is performed on the preprocessed data based on the feature extraction module to obtain feature vectors. Specifically, the feature extraction module is constructed based on dilated convolution; the processed multi-source geological data is input into the feature extraction module to obtain feature vectors.

5. The method according to claim 4, characterized in that, In step 5, the feature vectors are fused based on the feature fusion module. Specifically, the feature vector fusion module processes the foreground feature vector set and the background feature vector set respectively, and then fuses them to obtain a fused feature vector set.

6. The method according to claim 5, characterized in that, In step 6, the fused feature vectors are identified to realize the detection of gold deposits in volcanic rock cryptovolcanic breccia. Specifically, the fused feature vector set is identified based on the Transformer module to realize the detection of gold deposits in volcanic rock cryptovolcanic breccia.

Images