Geological guarantee intelligent planning mining system based on deep reinforcement learning

Abstract

The invention relates to the technical field of intelligent mines, and discloses a geological guarantee intelligent planning mining system based on deep reinforcement learning, which comprises a geological guarantee basic module, an intelligent planning engine module and a closed-loop execution system module, the geological guarantee basic module constructs a dynamic probabilistic digital twinning environment through a geological uncertainty implicit neural field, and realizes data real-time fusion updating in combination with a gradient flow; the intelligent planning engine module realizes long-period risk perception planning based on a geological perception Markov decision process and a causal world model; the closed-loop execution system module deploys a lightweight model through strategy distillation, establishes a twinborn reality bidirectional evolution closed loop, and realizes planning and execution adaptive collaboration; according to the method, the problems of static fragility of geological guarantee, short view non-strategy planning and system splitting open loop of a traditional mining system are solved, full-life-cycle dynamic cognition, deep planning and closed-loop self-evolution of a mine are achieved, and mining benefits and safety are remarkably improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent mine, more particularly to a geological guarantee intelligent planning mining system based on deep reinforcement learning. BACKGROUND

[0002] Modern large-scale mining is a complex system engineering with high risk, high investment and long cycle, and its benefits and safety are highly dependent on the scientificity of decision-making. However, the existing technical system faces three fundamental difficulties in realizing true intelligence:

[0003] First, the static and vulnerability of geological guarantee. Current geological guarantee mainly relies on static three-dimensional geological models based on sparse drilling data. Such models are essentially discrete snapshots of geological bodies, with unquantified uncertainty, and are updated with lag and dependence on manual work, which cannot dynamically reflect new information in the mining process, resulting in geological guarantee becoming a static hypothesis, which poses a hidden danger to planning.

[0004] Second, the short-sightedness and non-strategic nature of intelligent planning. Existing intelligent planning is mostly based on defective geological models for short-term local optimization. Even if deep reinforcement learning is introduced, it does not incorporate geological constraints and learns short-sighted strategies, such as preferentially mining rich ore blocks, leading to stress concentration in ore pillars or overlying low-grade resources, and lacks consideration of geological and mechanical laws, which is not truly intelligent.

[0005] Third, the fragmentation and open-loop nature of the mining system. There is an information gap between planning and execution departments, and on-site geological deviations need to be adjusted manually with lag, lacking automatic real-time closed-loop feedback mechanism, resulting in a disconnection between plan and reality, and low system efficiency and safety.

[0006] Based on this, the present application proposes a geological guarantee intelligent planning mining system based on deep reinforcement learning to solve the above problems. SUMMARY

[0007] In order to overcome the above-mentioned defects of the prior art, the present application provides a geological guarantee intelligent planning mining system based on deep reinforcement learning to solve the problems existing in the background art.

[0008] The present application provides the following technical solution: a geological guarantee intelligent planning mining system based on deep reinforcement learning, a geological guarantee basic module for constructing a high-fidelity, dynamic digital twin environment for deep reinforcement learning;

[0009] An intelligent planning engine module for developing a long-cycle deep reinforcement learning decision-making agent with geological risk perception;

[0010] A closed-loop execution system module for realizing adaptive collaborative execution of planning and reality bidirectional evolution.

[0011] Through the collaborative architecture of the three major modules, the core pain points of traditional mining systems, such as static fragility of geological support, short-sighted and non-strategic planning, and fragmented and open-loop systems, are fundamentally addressed. This enables dynamic cognition, far-reaching planning, and closed-loop self-evolution throughout the entire life cycle of the mine, providing a systematic solution for efficient and safe mining.

[0012] As a further aspect of the present invention: the geological support infrastructure module includes:

[0013] Based on the continuous probability representation unit of the Geological Uncertainty Implicit Neural Field (GUINF), a depth coordinate network F defined by parameter θ is used. θ Receive the spatiotemporal coordinates c = (x, y, z, t) and output the joint probability distribution p(v|c) of the key geological attribute vector V of that point;

[0014] The gradient-flow-based dynamic evolution and data fusion unit realizes the dynamic evolution of the digital twin environment by calculating the negative log-likelihood loss of new data and updating the network parameters through gradient descent.

[0015] By overcoming the limitations of static geological models through the continuous probabilistic representation of GUINF and combining it with the gradient flow dynamic update mechanism, geological assurance is upgraded from static assumptions to dynamic, quantitative, real-time evolutionary models, effectively reducing the planning risks caused by geological uncertainties.

[0016] As a further aspect of the present invention: the GUINF adopts a sinusoidal representation network architecture, with the activation function being φ(x) = sin(x), making the network and its derivatives of any order continuous and analytically calculable with respect to the input coordinates; its output layer is a mixture density network, outputting a Gaussian mixture model parameter set. Geological attribute probability density satisfies

[0017] Sine representation networks ensure the continuity and differentiability of geological attribute characterization and support direct coupling with physical solvers; hybrid density networks accurately quantify the uncertainty of geological attributes, providing reliable probabilistic basis for planning decisions and improving the accuracy and credibility of geological assurance.

[0018] As a further aspect of the present invention: the intelligent planning engine module includes:

[0019] The geologically-aware Markov decision process (MDP) building unit formalizes the planning problem into tuples (S, A, P, R, γ), where the state space S contains the global geological security representation, the geological risk map of key areas and the resource endowment status, the action space A is the composite hierarchical decision instruction, and the reward function R contains the long-term value gain, geological risk penalty and information uncertainty penalty.

[0020] Long-term programming algorithm unit based on causal world model, by learning world model Mφ Predict state transitions and reward dynamics, and imagine and optimize long-term trajectories in the world model.

[0021] Geological perception MDP deeply integrates geological constraints into the decision-making framework, and the causal world model supports long-term trajectory planning, breaking through the short-sightedness of traditional planning, enabling decisions to take into account both short-term benefits and long-term geological safety, and maximizing the value of the entire life cycle of the mine.

[0022] As a further aspect of the present invention: the reward function is specifically in the form of:

[0023] R t =w val ·ΔNPV t -w risk ·f risk (S t+1 )-w info ·g info (S t )

[0024] Among them, w val w risk w info The weighting coefficients for economic value, geological risk, and information uncertainty, respectively, are ΔNPV. t f is the net present value increment. risk As a penalty item for geological risks, g info This is a penalty item for information uncertainty.

[0025] Multi-objective reward functions balance economic value, geological risks, and information acquisition, guiding agents to proactively avoid risks and reduce cognitive uncertainty while pursuing benefits, avoiding sacrificing long-term safety for short-term gains, and improving the scientific and robust nature of planning.

[0026] As a further aspect of the present invention: the geological risk penalty item f risk The calculation formula is: f risk =λ1∫ Ω I[σ>σ yield ]dV+λ2·E seismic Where λ1 and λ2 are the risk item weight coefficients, [σ>σ yield ] is the indicator function, σ is the stress value, σ yield Let dV be the rock yield stress, dV be the volume integral of the stope area, and E be the volume integral of the stope area. seismic This represents the total energy of the microseismic event.

[0027] By quantifying the volume of rock mass stress exceeding limits and microseismic energy, geological risks (such as rock mass instability and ground pressure disasters) can be accurately measured, enabling intelligent agents to proactively avoid high-risk areas in the planning process and significantly improve the safety of the mining process.

[0028] As a further aspect of the present invention: the closed-loop execution system module includes:

[0029] The strategy distillation and secure deployment unit distills the decision-making knowledge of the trained DRL agent into a lightweight student neural network π. student In this context, real-time deployment is achieved through supervised learning;

[0030] The twin-reality bidirectional evolution closed-loop unit enables data flow updates from reality to the twin and proactive exploration from the twin to reality, triggering supplementary exploration tasks when key decisions depend on areas with high uncertainty.

[0031] Strategy distillation solves the problem of real-time deployment of complex models; bidirectional evolutionary closed loop breaks down the information barriers between planning and execution, enabling real-time data feedback and proactive inquiry, allowing the system to continuously learn and optimize from practice, and improve its adaptability to the actual mining environment.

[0032] As a further aspect of the present invention: a geologically sound intelligent planning and mining method based on deep reinforcement learning, comprising the following steps:

[0033] Step S1: Construct a high-fidelity, dynamic digital twin environment for deep reinforcement learning, specifically including continuous probability representation based on implicit neural fields of geological uncertainty and dynamic evolution and data fusion based on gradient flow;

[0034] Step S2: Develop a long-term deep reinforcement learning decision-making agent for geological risk perception, specifically including constructing a Markov decision process for geological perception and adopting a long-term planning algorithm based on a causal world model.

[0035] Step S3: Achieve adaptive and collaborative execution of the two-way evolution of planning and reality, specifically including strategy distillation and secure deployment, and establishing a closed loop for the two-way evolution of twin reality.

[0036] It provides a complete end-to-end technology path, covering everything from environment construction and intelligent agent development to closed-loop execution, ensuring that the system can be implemented and systematically solving the problems of poor dynamism, short-sighted planning, and insufficient coordination in traditional mining methods.

[0037] As a further aspect of the present invention: step S1 includes:

[0038] S1.1: The mine is implicitly represented by the GUINF network, and the joint probability distribution of geological attributes at any spatiotemporal coordinate is output.

[0039] S1.2: The sinusoidal representation of the network architecture enables the network to have continuous differentiability.

[0040] S1.3: Output Gaussian mixture model parameters through a mixture density network;

[0041] S1.4: Negative log-likelihood loss L based on new observation data new (θ)=-logp(v new |c new ;θ), through gradient descent Update network parameters.

[0042] Step S1 ensures that the digital twin environment has high fidelity and dynamic evolution characteristics, can absorb new data in real time and quantify uncertainty, and provide an accurate and reliable geological understanding basis for subsequent planning and decision-making.

[0043] As a further aspect of the present invention: step S2 includes:

[0044] S2.1: Define the MDP state space as including global geological characterization, local risk map and resource status, the action space as a hierarchical discrete structure, and the reward function as a multi-objective weight;

[0045] S2.2: Train a Transformer-based world model, and optimize long-cycle trajectories in the world model using the cross-entropy method, with the planning vision set to 240 monthly decision steps.

[0046] Step S2 enables the agent to deeply perceive geological constraints and conduct long-term planning, avoiding short-term optimization traps and ensuring that planning decisions achieve the optimal balance between efficiency and safety throughout the entire life cycle of the mine.

[0047] The technical effects and advantages of this invention are as follows:

[0048] (1) Fundamentally improve the level of geological security: Through differential probability digital twins, this invention elevates geological security from a static and qualitative concept to a dynamic, quantitative, and reliable commitment that can evolve in real time with mining activities, providing an unprecedented solid foundation for all planning decisions.

[0049] (2) Achieving intelligent planning: Through the DRL framework for geological risk perception, this invention enables planning decisions to transcend short-term profit orientation, and to proactively manage and avoid long-term geological risks, achieving a deep unity of economic benefits and geological safety, and making it possible to maximize the value of the entire life cycle of the mine.

[0050] (3) A closed-loop self-evolving mining system was constructed: Through the bidirectional evolution and active inquiry mechanism of twin reality, this invention breaks down the barriers between planning and execution, and creates an intelligent system that can learn from practice, actively improve its own cognition, and achieve the co-evolution of planning and reality. Attached Figure Description

[0051] The invention will now be further described with reference to the accompanying drawings.

[0052] Figure 1A schematic diagram of the overall architecture of the intelligent planning and mining system for geological protection based on deep reinforcement learning provided in an embodiment of the present invention;

[0053] Figure 2 This is a flowchart illustrating the bidirectional evolution closed loop of twin reality in an embodiment of the present invention.

[0054] Figure 3 This is a schematic diagram of the framework of the geological perception Markov decision-making process in an embodiment of the present invention:

[0055] Figure 4 This is a system block diagram of the closed-loop execution system module in an embodiment of the present invention. Detailed Implementation

[0056] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. These embodiments are only used to explain the present invention and are not intended to limit the scope of protection of the present invention.

[0057] Example 1

[0058] like Figure 1 As shown, this embodiment provides a geological support intelligent planning and mining system based on deep reinforcement learning. The system includes three core modules: a geological support basic module, an intelligent planning engine module, and a closed-loop execution system module.

[0059] The Geological Support Foundation Module is responsible for building a high-fidelity, dynamic digital twin environment for deep reinforcement learning. This module includes two key units:

[0060] Continuous probabilistic representation unit based on Geological Uncertainty Implicit Neural Field (GUINF):

[0061] like Figure 2 As shown, this cell uses a single depth coordinate network F defined by parameter θ. θ This implicitly represents the entire mine. The network receives arbitrary spatiotemporal coordinates c = (x, y, z, t) as input and outputs the joint probability distribution p(v|c) of all key geological attribute vectors v at that point. Vector v includes two types of key information: resource attributes (such as ore grade and density) and engineering attributes (such as rock quality indicators RMR, elastic modulus, Poisson's ratio, cohesion C, and internal friction angle φ).

[0062] The GUINF network employs a sinusoidal representation network (SIREN) architecture, with its activation function being φ(x) = sin(x). This makes the network F θ Its derivatives of any order are continuous and analytically solvable with respect to the input coordinates. This property allows for direct coupling of the digital twin with a physical solver based on partial differential equations.

[0063] To provide true geological assurance, namely a reliable quantification of uncertainty, the output layer of GUINF is a Mixture Density Network (MDN), whose output is the parameter set defining a Gaussian Mixture Model (GMM). Therefore, at any point c, the probability density of any geological attribute v is:

[0064] Gradient-flow-based dynamic evolution and data fusion unit:

[0065] This unit makes the digital twin a "living" entity, capable of dynamically absorbing new data generated during the mining process. When on-site mining reveals new geological information, these new observational data (c...) with precise spatiotemporal tags... new v new This will be used to fine-tune the GUINF network online. The negative log-likelihood loss will be calculated using the new data in the current model:

[0066] L new (θ)=-logp(v new |c new ;θ)

[0067] And update the network parameters using gradient descent:

[0068] The intelligent planning engine module is responsible for developing a long-term deep reinforcement learning decision-making agent for geological risk perception. For example... Figure 3 As shown, this module includes:

[0069] Geologically-aware Markov Decision Process (MDP) building blocks:

[0070] The intelligent planning problem is formalized as a complex geological sensing MDP, whose tuples are defined as (S, A, P, R, γ). The state space S is a high-dimensional, structured vector that comprehensively describes the mine's situation at time t, including: global geological assurance representation and digital twin network. The parameter θ t The features of the entity itself, or its feature embedding on a predefined grid; Geological risk map of key areas: spatial distribution of key physical fields obtained by querying through digital twins and calculating through stress field solvers; Resource endowment status: spatial distribution of remaining recoverable reserves, grade distribution and its uncertainty.

[0071] Reward function R(S) t A t Designed as a multi-objective, risk-penalty model:

[0072] R t =w val ·ΔNPV t -wrisk ·f risk (S t+1 )-w info ·g info (S t )

[0073] Among them, geological risk penalty item f risk The calculation formula is:

[0074] f risk =λ1∫ Ω I[σ>σ yield ]dV+λ2·E seismic

[0075] Long-term programming algorithm unit based on causal world model:

[0076] The agent first learns a world model M φ The model is able to predict the approximate dynamics of state transitions and reward functions. The agent then performs massively parallel long-period trajectory visualization within its world model, using gradient-based trajectory optimization algorithms or Monte Carlo tree search to select the action sequence with the optimal long-term cumulative reward.

[0077] The closed-loop execution system module is responsible for achieving adaptive and collaborative execution of the bidirectional evolution of planning and reality. For example... Figure 4 As shown, this module includes:

[0078] Strategy Distillation and Secure Deployment Unit: Employing strategy distillation technology, the wisdom of the teacher model is "distilled" into a lightweight student neural network π. student In this approach, the teacher model makes decisions on a massive number of states within a digital twin, forming a huge dataset of (states, optimal actions), which is then used to train the student model using supervised learning.

[0079] Twin Reality Two-Way Evolution Closed-Loop Unit: Establishing a two-way information flow: Data flow from reality to the twin: All new geological data from actual mining activities will be used to update the digital twin in real time and smoothly; Exploration flow from the twin to reality: When the DRL agent discovers that the quality of a key decision is highly dependent on a high-uncertainty area in the digital twin, it will actively trigger the "active exploration" task to calculate the information value (VoI) of supplementary exploration in the uncertain area.

[0080] Example 2

[0081] This embodiment provides a geological protection intelligent planning and mining method based on deep reinforcement learning. The specific implementation steps are as follows:

[0082] Step 1: Construct a high-fidelity, dynamic digital twin environment for deep reinforcement learning.

[0083] First, initial geological data for model training is collected and processed. Data sources include: Historical borehole data: Core logging data from all boreholes during the mine exploration and production exploration phases, totaling 2,500 boreholes, yielding approximately 150,000 samples with spatial coordinates (x, y, z) and corresponding attributes such as ore grade, lithology, and RMR values. Geological profile maps: 25 existing key geological profile maps are digitized, extracting spatial location information of major lithological contact zones, faults, and alteration zones, generating approximately 30,000 control point data. Existing stope exposure data: Geological sketches, ore grade acceptance records, and ground pressure monitoring data from mined areas over the past 5 years are compiled and aligned with the 3D model of the stopes, generating approximately 50,000 data points with precise spatiotemporal labels (x, y, z, t) and corresponding attributes. Geophysical data: If available, 3D seismic exploration or resistivity tomography data can be introduced as soft constraints to provide regional geological structure trend information.

[0084] Secondly, the collected multi-source heterogeneous data underwent preprocessing: Coordinate System 1: All data were transformed to a unified mine coordinate system. Data Cleaning: Outlier data points with incomplete records or obvious errors were removed. Normalization: All geological attribute values ​​(such as grade, RMR, elastic modulus, etc.) were subjected to min-max normalization, scaling them to the [0,1] interval to facilitate stable training of the neural network. Training Set Construction: The processed data points were integrated into a training set with a unified format. Where N≈230,000, c i For 4-dimensional spatiotemporal coordinates (for static data, time t is set to 0), v i This is the normalized geological attribute vector for that point.

[0085] Then, the GUINF network is initialized and trained. Network architecture: The network adopts an 8-layer fully connected structure, with 256 neurons per layer and the activation function being sin(x). The input layer receives position-encoded 4-dimensional coordinates (x, y, z, t), and the output layer is a mixture density network, outputting parameters (mixing coefficients π) of K = 5 Gaussian components for each geological attribute. k Mean μ k ,variance Pre-training process: The network is pre-trained using the training set D. The loss function is the sum of the negative log-likelihoods of all training data points.

[0086]

[0087] Formula explanation: L(θ): Total loss function over the entire training set D. Summing is performed on all data points from 1 to N in the training set. logp(v) i |c i;θ): The model for the i-th data point (c i ,v i The goal of training is to find a set of parameters θ that maximizes the total probability (i.e., the total likelihood) of all known data points, which is equivalent to minimizing the negative log-likelihood loss.

[0088] The Adam optimizer is used with an initial learning rate of 1×10⁻⁶. -4 The batch size is 64, and the training is conducted for 10,000 epochs until the loss function converges, thus obtaining the initial digital twin environment.

[0089] Step 2: Develop a long-term deep reinforcement learning decision-making agent for geological risk perception.

[0090] First, the specific parameters of the geological perception MDP are defined. State space S: The dimension is set to 1024, specifically composed of three parts: Global geological representation (512-dimensional): The current GUINF network is queried on a fixed 32×32×32 grid to obtain the mean and variance of geological attributes for each grid point, which are then compressed into a 512-dimensional feature vector using a small convolutional autoencoder. Local risk map (256-dimensional): For several key areas requiring decision-making, rapid stress simulation is performed using software such as FLAC3D based on the rock mechanics parameters queried from GUINF, extracting key indicators such as maximum principal stress and stress concentration factor, and encoding them into a 256-dimensional vector. Resource status (256-dimensional): Information such as the spatial distribution, average grade, and total reserves of remaining oreable blocks is statistically analyzed and encoded into a 256-dimensional vector. Action Space A: Designed as a layered discrete structure: the upper layer is the selection of 10 candidate stopes (discrete action 110), the middle layer is the selection of 3 mining methods (such as stratified caving, segmented open stope, and backfilling) (discrete action 13), and the lower layer is the monthly mining intensity (discrete action, divided into low, medium, and high levels). Reward Function Weights: Weights are set according to the mine's long-term strategic goals: w val =1.0, w risk =0.3, w info =0.1. In the geological risk penalty, λ1 = 100 (penalty for volumetric instability), λ2 = 50 (penalty for microseismic energy). The discount factor γ = 0.99 to encourage long-term planning.

[0091] Then, the causal world model is trained. The agent interacts with the digital twin environment constructed in step 1, collecting 10... 6 The model uses empirical data (state, action, next state, reward) for each step. This data is then used to train a Transformer-based sequence model as the world model, with a hidden dimension of 512, containing 8 attention heads and a 6-layer encoder-decoder structure.

[0092] Finally, long-term planning training is performed. The cross-entropy method (CEM) is used for trajectory optimization within the learned world model. The planning horizon is set to 240 monthly decision steps. At each decision point, the agent imagines and samples 1000 action sequences for the next 240 steps in parallel, evaluates the cumulative reward of each sequence, selects the top 10% (i.e., 100) of the best-performing sequences, and updates the sampling distribution for the next round using their mean and variance. After 5 iterations, the first action of the top-ranked action sequence in the updated distribution is executed.

[0093] Step 3: Achieve adaptive and collaborative execution of the two-way evolution of planning and reality.

[0094] First, policy distillation is performed. A trained, computationally expensive CEM-based teacher model is used to infer from 1000 randomly sampled states in a digital twin environment, collecting (state, optimal action) data pairs. These data pairs are then used to train a lightweight student network in a supervised learning manner. This student network is a 3-layer fully connected neural network (128 neurons per layer, ReLU activation), using cross-entropy loss for behavior cloning.

[0095] Then, a closed-loop system is deployed. The distilled student strategy network is deployed to the mine's central scheduling system, automatically generating and issuing the monthly mining plan at the beginning of each month. Simultaneously, an automated data acquisition channel is established to collect new data generated by underground production, measurement, geological survey, and monitoring systems (such as actual ore grade, lithological changes revealed in roadways, microseismic monitoring events, etc.) into the database in real time.

[0096] Finally, proactive exploration is implemented. A monitoring module continuously runs in the system background. This module periodically (e.g., weekly) assesses the geological uncertainties of key areas to be mined. It uses the Monte Carlo method to sample multiple sets (e.g., 100 sets) of possible geological realizations from the probability distribution of GUINF, and runs a planning agent for each realization to evaluate the value of the decision. When the system detects that the geological uncertainty in a certain area causes the expected NPV fluctuation range of subsequent planning to exceed a preset threshold, it automatically generates a "proactive exploration" task, sending supplementary exploration suggestions to the geological survey department. The suggestions include optimized borehole locations, depths, and target sampling densities.

[0097] The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A geological support intelligent planning and mining system based on deep reinforcement learning, characterized in that: include: The geological support module is used to build a high-fidelity, dynamic digital twin environment for deep reinforcement learning. The intelligent planning engine module is used to develop a long-term deep reinforcement learning decision-making intelligent agent for geological risk perception. The closed-loop execution system module is used to achieve adaptive and collaborative execution of the bidirectional evolution of planning and reality.

2. The intelligent planning and mining system for geological support based on deep reinforcement learning according to claim 1, characterized in that: The geological support infrastructure module includes: Based on the continuous probability representation unit of the Geological Uncertainty Implicit Neural Field (GUINF), a depth coordinate network F defined by parameter θ is used. θ Receive the spatiotemporal coordinates c = (x, y, z, t) and output the joint probability distribution p(v|c) of the key geological attribute vector V of that point; The gradient-flow-based dynamic evolution and data fusion unit realizes the dynamic evolution of the digital twin environment by calculating the negative log-likelihood loss of new data and updating the network parameters through gradient descent.

3. The intelligent planning and mining system for geological support based on deep reinforcement learning according to claim 2, characterized in that: The GUINF uses a sinusoidal network architecture with the activation function φ(x) = sin(x), ensuring that the network and its derivatives of any order are continuous and analytically calculable with respect to the input coordinates. Its output layer is a mixture density network, outputting a Gaussian mixture model parameter set. Geological attribute probability density satisfies 4. The intelligent planning and mining system for geological protection based on deep reinforcement learning according to claim 1, characterized in that: The intelligent planning engine module includes: The geologically-aware Markov decision process (MDP) building unit formalizes the planning problem into tuples (S, A, P, R, γ), where the state space S contains the global geological security representation, the geological risk map of key areas and the resource endowment status, the action space A is the composite hierarchical decision instruction, and the reward function R contains the long-term value gain, geological risk penalty and information uncertainty penalty. Long-term programming algorithm unit based on causal world model, by learning world model M φ Predict state transitions and reward dynamics, and imagine and optimize long-term trajectories in the world model.

5. The intelligent planning and mining system for geological support based on deep reinforcement learning according to claim 4, characterized in that: The reward function takes the following specific form: R t =w val ·ΔNPV t -w risk ·f risk (S t+1 )-w info ·g info (S t ) Among them, w val w risk w info The weighting coefficients for economic value, geological risk, and information uncertainty, respectively, are ΔNPV. t f is the net present value increment. risk As a penalty item for geological risks, g info This is a penalty item for information uncertainty.

6. The intelligent planning and mining system for geological support based on deep reinforcement learning according to claim 5, characterized in that: The geological risk penalty item f risk The calculation formula is: f risk =λ1∫ Ω I[σ>σ yield ]dV+λ2·E seismic Where λ1 and λ2 are the risk item weight coefficients, [σ>σ yield ] is the indicator function, σ is the stress value, σ yield Let dV be the rock yield stress, dV be the volume integral of the stope area, and E be the volume integral of the stope area. seismic This represents the total energy of the microseismic event.

7. The intelligent planning and mining system for geological support based on deep reinforcement learning according to claim 1, characterized in that: The closed-loop execution system module includes: The strategy distillation and secure deployment unit distills the decision-making knowledge of the trained DRL agent into a lightweight student neural network π. student In this context, real-time deployment is achieved through supervised learning; The twin-reality bidirectional evolution closed-loop unit enables data flow updates from reality to the twin and proactive exploration from the twin to reality, triggering supplementary exploration tasks when key decisions depend on areas with high uncertainty.

8. A geological protection intelligent planning and mining method based on deep reinforcement learning, characterized in that, Includes the following steps: Step S1: Construct a high-fidelity, dynamic digital twin environment for deep reinforcement learning, specifically including continuous probability representation based on implicit neural fields of geological uncertainty and dynamic evolution and data fusion based on gradient flow; Step S2: Develop a long-term deep reinforcement learning decision-making agent for geological risk perception, specifically including constructing a Markov decision process for geological perception and adopting a long-term planning algorithm based on a causal world model. Step S3: Achieve adaptive and collaborative execution of the two-way evolution of planning and reality, specifically including strategy distillation and secure deployment, and establishing a closed loop for the two-way evolution of twin reality.

9. The intelligent planning and mining method for geological protection based on deep reinforcement learning according to claim 8, characterized in that: Step S1 includes: S1.1: The mine is implicitly represented by the GUINF network, and the joint probability distribution of geological attributes at any spatiotemporal coordinate is output. S1.2: The sinusoidal representation of the network architecture enables the network to have continuous differentiability. S1.3: Output Gaussian mixture model parameters through a mixture density network; S1.4: Negative log-likelihood loss L based on new observation data new (θ)=-logp(v new |c new ;θ), through gradient descent Update network parameters.

10. The intelligent planning and mining method for geological protection based on deep reinforcement learning according to claim 8, characterized in that: Step S2 includes: S2.1: Define the MDP state space as including global geological characterization, local risk map and resource status, the action space as a hierarchical discrete structure, and the reward function as a multi-objective weight; S2.2: Train a Transformer-based world model, and optimize long-cycle trajectories in the world model using the cross-entropy method, with the planning vision set to 240 monthly decision steps.

Images