A method and system for residual ore filling and recovery based on multi-layer intelligent decision-making

Abstract

The application provides a residual ore filling recovery method and system based on multi-layer intelligent decision, and relates to the technical field of mine exploitation. The system comprises several modules, such as data acquisition and preprocessing, hierarchical evaluation, data enhancement, neural network prediction, rule correction, economic benefit evaluation, monitoring and early warning and feedback closed loop, and man-machine interaction. The method is as follows: import mine basic data and output standard feature matrix; quantify residual ore recovery potential and engineering geological conditions through a three-level evaluation system; expand the data set when the sample is insufficient; use graph neural network multi-task cascade to predict the filling scheme and perform strength checking; correct the scheme through hard constraint checking and soft constraint scoring; carry out full life cycle economic benefit analysis; output visual decision report and realize model continuous optimization through post-construction monitoring. The method solves the problem that the prior art relies on experience and lacks a systematic decision mechanism, and improves the scientificity, safety and economic rationality of the residual ore filling scheme.

Description

Technical Field

[0001] This invention relates to the field of mine backfilling and recycling technology, specifically to a method and system for backfilling and recycling residual ore based on multi-level intelligent decision-making. Background Technology

[0002] As metal mining in my country continues to extend deeper, the ore body occurrence conditions are becoming increasingly complex, ground pressure activity is significantly increasing, and the number of goaf areas is continuously growing, posing severe challenges to safety management. After mines enter the mid-to-deep mining stage or the remining stage of old mining areas, a large number of marginal ore, bottom ore, and residual ore bodies have not yet been effectively recovered due to their complex occurrence conditions. These resources typically have high grades and significant economic value, but their mining decisions heavily rely on the experience and judgment of engineering and technical personnel, leading to differences in the understanding of geological conditions and safety risks among different personnel, resulting in highly subjective and uncertain decision-making outcomes.

[0003] Currently, mine backfilling technology mainly focuses on equipment automation control, construction parameter monitoring, and slurry transportation process optimization. For example, the invention patent with publication number CN113389592A improves the level of automation by deploying multiple sensors in the backfilling equipment to achieve real-time sensing of process parameters and automatic adjustment of backfilling actions. However, this type of technology is mostly concentrated at the equipment execution and construction control levels, mainly addressing how to complete backfilling construction safely and efficiently, without systematically responding to the core decision-making issue of what backfilling and recovery schemes should be adopted under different mine, intermediate section, and stope conditions. On the other hand, the invention patent with publication number CN119575827A calculates and controls the stability of the backfilling process in real time by establishing ore body stress models and slurry physical property models to prevent safety risks caused by excessive pore water pressure or insufficient strength. Although this method improves construction safety to some extent, its core technology is still limited to the dynamic control of a single backfilling process, lacking a systematic consideration of the overall feasibility of mine residual ore recovery, multi-level decision-making logic, and multi-scheme comparison mechanisms.

[0004] Residual ore backfilling and recovery is essentially a complex decision-making problem involving multiple factors, including mine resource conditions, intermediate level structural characteristics, stope spatial relationships, backfill material performance, economic efficiency, and safety. Existing technologies generally suffer from the following shortcomings: First, they lack a three-tiered decision-making system at the mine-intermediate level-stope level, making it difficult to achieve step-by-step screening and optimization of residual ore recovery schemes. Second, they rely heavily on single models or empirical rules, lacking consistency verification and fusion mechanisms between different evaluation methods. Third, there is a disconnect between backfill scheme design and engineering safety verification, as well as construction feedback, failing to form a closed-loop decision-making process for sustainable optimization.

[0005] In summary, existing technologies have significant limitations in decision-making regarding the backfilling and recovery of residual ore in metal mines: they lack both quantitative evaluation standards for the recoverability of residual ore and scientific basis for comparing and optimizing backfilling schemes, making it difficult to achieve a balance between safety and efficiency. Therefore, there is an urgent need to construct a multi-layered intelligent decision-making system that integrates on-site mine data with artificial intelligence algorithms to provide a more scientific, safe, and economically reasonable solution for the formulation of residual ore backfilling schemes. Summary of the Invention

[0006] In view of the technical problems existing in the background art, this invention provides a method and system for residual ore backfilling and recovery based on multi-level intelligent decision-making to guide the backfilling operations of residual ore bodies. It objectively assesses the recoverability of residual ore and intelligently optimizes backfilling mining schemes under complex and variable mining conditions, ensuring safe mine production. It mainly constructs a three-level evaluation system at the mine level, intermediate level, and stope level; uses state-variable weighting functions and fuzzy comprehensive evaluation to achieve dynamic weight adjustment of indicators; establishes an indicator relationship network and a cloud correlation matrix; solves the problem of insufficient samples through data augmentation technology; and introduces a dual verification mechanism of neural network prediction and expert rule correction; ultimately forming a complete closed-loop decision-making process from data collection, intelligent evaluation, scheme optimization to monitoring feedback, significantly improving the scientificity, safety, and economy of residual ore recovery schemes.

[0007] In a first aspect, embodiments of the present invention provide a residual ore backfilling and recovery system based on multi-layer intelligent decision-making, comprising: The data acquisition and preprocessing module is used to import geological, production, and environmental data from the mine, perform data preprocessing, and output a standard feature matrix. The hierarchical evaluation module, connected to the data acquisition and preprocessing module, includes mine-level macro evaluation units, intermediate-level evaluation units, and stope-level evaluation units. It is used to calculate mine-level recoverability indicators, intermediate-level evaluation indicators, and stope-level evaluation indicators in sequence, so as to realize the quantitative evaluation of residual ore recovery potential and engineering geological conditions from macro to micro. The neural network prediction module is connected to the data acquisition and preprocessing module and the hierarchical evaluation module, respectively. It is used to model the spatial relationship of the mining area and predict suitable filling schemes, material types and proportion parameters in a multi-task cascade based on the standard feature matrix and evaluation index. The rule correction module, connected to the neural network prediction module, has a built-in expert rule base, including a hard constraint verification unit, a soft constraint scoring unit, and a comprehensive weighted fusion unit. It is used to correct the safety and engineering feasibility of the predicted scheme through hard constraint verification, soft constraint scoring, and comprehensive weighted fusion, and output a corrected scheme ranking list. The economic benefit assessment module, connected to the rule correction module, is used to perform full life-cycle cost-benefit analysis, risk analysis, and sensitivity analysis on technically feasible solutions. The monitoring, early warning and feedback closed-loop module, connected to the rule correction module, is used for post-construction monitoring and early warning, on-site verification and incremental model learning, forming a continuously optimized decision-making closed loop; The human-computer interaction module is connected to the monitoring, early warning and feedback closed-loop module and the economic benefit evaluation module, respectively, and is used to provide a visual interface, parameter configuration and decision report generation functions.

[0008] As a further improvement of the present invention, a data augmentation module is also included; the data augmentation module is connected to the hierarchical evaluation module and is used to generate high-quality synthetic samples that meet physical constraints and expand the training dataset when there are insufficient samples of historical mining data or uneven distribution of categories.

[0009] As a further improvement of the present invention, the hierarchical evaluation module adopts the fuzzy comprehensive evaluation method to configure a state-variable weighting function for each evaluation index, supporting adaptive parameter fitting. The calculation process is as follows: Input the original index x of each parameter i Perform extreme value testing and calculate the state-weighted function: ; Perform weight normalization: ; Perform a weighted comprehensive calculation to output a risk level score R, and determine the risk level based on the range of values ​​for R: ; in, The mean; w i The initial weights for the i-th indicator are assigned by experts; w' i The adjusted weights; S(x) i ) is the state-varying weighting function; μ i σ is the membership degree of the i-th indicator; f is the standard deviation; f is the specific functional form of the state-variable weighting function; j is the evaluation level number; n is the total number of evaluation indicators. The fuzzy comprehensive evaluation method supports Gaussian, exponential decay, linear, sigmoid, and trapezoidal function forms, and can automatically optimize the morphological parameters of membership functions using historical labeled data.

[0010] As a further improvement of the present invention, in the graded evaluation module, the mine-level macro-evaluation unit comprehensively considers factors such as reserves, grade, ore body structure and goaf distribution to calculate the mine-level recoverability index, which is divided into four levels: high recovery potential, medium potential, low recovery potential and not yet available. In the intermediate-level evaluation unit, the calculation formula for the intermediate-level evaluation indicators is as follows: M 0= Σ(w i × μ i ) × λ s × λ a ; Among them, M0 is the mid-level evaluation indicator; w i The weights for each indicator; μ i λ represents the membership degree of each indicator. s λ is a correction factor for the mining sequence, ranging from 0.8 to 1.2. When the upper and middle sections have been filled, λ... s =1.2, when the upper and middle sections have not yet been mined. s =0.8; λ a The influence coefficient for adjacent intermediate sections is calculated based on the volume and distance of the goaf in adjacent intermediate sections, with a value ranging from 0.7 to 1.0. The intermediate section-level evaluation results are divided into four levels, as follows: When M0 ≥ 0.8, priority is given to mining, and it is set as Level I; When 0.6≤M0<0.8, normal mining is performed and is set to Level II; When 0.4 ≤ M0 < 0.6, cautious mining should be adopted, and the level should be set to III. When M0 < 0.4, mining is temporarily suspended and set to Level IV; The results of the intermediate-level evaluation are used to determine the priority order of mining in each intermediate level and to provide boundary condition constraints for the stope-level evaluation.

[0011] As a further improvement of the present invention, the mining area-level evaluation unit of the graded evaluation module introduces the Mathews stability diagram method as a comparative verification mechanism; the Mathews stability diagram method is expressed by the formula... Calculate the stability coefficient; in, N' Q' is the corrected stability coefficient; Q' is the corrected Q value, Q' = (RQD / J n ) × (J r / J a ), J n J is the number of joint groups. r J is the roughness coefficient of the most vulnerable joint. a RQD represents the degree of erosion or infill at the weakest joint surface; A is the rock integrity index; and A is the stress coefficient, A = f(σ c / σ1), σ c σ1 is the uniaxial compressive strength of the rock; B is the joint orientation coefficient; C is the gravity adjustment coefficient; HR is the hydraulic radius, HR = S / P = excavation face area / excavation face perimeter; Based on the calculated N' and HR values, the location of the goaf on the Mathews stability map is determined to be either a stable zone, a transition zone, or a collapse zone. The evaluation results of the fuzzy comprehensive evaluation method are checked for consistency with the Mathews stability graph determination results: if the determination results of the two methods are consistent, a high-confidence evaluation result is output; if the determination results are different, the system triggers a manual review prompt and simultaneously outputs the detailed calculation process of the two methods for engineers to refer to and make decisions.

[0012] As a further improvement of the present invention, the hierarchical evaluation module also includes a method weight adaptive unit; the method weight adaptive unit uses historical verification data to dynamically adjust the weight ratio of fuzzy evaluation and Mathews stable graph method in the final fusion decision through a Bayesian update mechanism; the multi-method fusion evaluation formula for the mining area is: ; Where a, b, and c are the fusion weights of fuzzy comprehensive evaluation, Mathews stable graph method, and RMR score, respectively, a+b+c = 1, and the initial weights can be set to a=0.4, b=0.4, and c=0.2. The weights can be adaptively adjusted according to historical validation data. The stability score obtained from the fuzzy comprehensive evaluation at the mining site level is normalized to 0-1. The stability level corresponding to the Mathews stability graph method is mapped to a value between 0 and 1; For RMR rock mass quality grade normalization scoring, =RMR / 100, where RMR (Rock Mass Rating) is the rock mass geomechanical classification index value; R1 is the final comprehensive stability score after the fusion of multiple methods at the stope level, with a value range of 0~1.

[0013] As a further improvement of the present invention, the data augmentation module includes an SMOTE algorithm unit (Synthetic Minority Over-sampling Technique) and a PI-MAC-GAN unit (Physics-informed Multi-Attribute Conditional Generative Adversarial Network), employing a strategy combining the SMOTE algorithm and PI-MAC-GAN; it takes as input a concatenated vector of random noise vector and conditional vector, and outputs filling recipe parameters.

[0014] As a further improvement of the present invention, the neural network prediction module adopts a graph neural network encoder and a multi-task cascaded prediction head, including a GCN layer (Graph Convolutional Network), a GAT layer (Graph Attention Network), and an SE-Block channel attention module (Squeeze-and-Excitation Block). The multi-task cascaded prediction sequentially realizes the prediction of filling method type, material ratio, and process parameters. The cascaded structure utilizes the inter-task dependency relationship, and the upstream prediction result serves as the condition input for the downstream task. Before outputting the backfill material ratio, the neural network prediction module needs to be verified by the strength verification unit. Based on Mitchell's strength design theory, the minimum strength requirement of the backfill body is calculated according to the geometric parameters of the stope. The calculation formula is as follows: ; Where, σ c,min Where γ is the minimum uniaxial compressive strength required for the filling material (kPa), γ is the bulk density of the filling material (kN / m³), H is the height of the filling material (m), L is the exposed length of the filling material (m), W is the width of the filling material (m), and φ is the internal friction angle of the filling material (°). In practical applications, it is usually necessary to multiply by a safety factor FS, which is generally taken as 1.5~2.0. The expected strength σ corresponding to the matching scheme predicted by the neural network c,pred σ must be satisfied c,pred ≥ σ c,min If the conditions are met, the system will automatically adjust the mixing parameters or trigger manual review.

[0015] As a further improvement of the present invention, the hard constraint rules of the rule correction module refer to rules involving safety regulations, policy red lines, or equipment limits, which, once triggered, have a veto power over the solution; the soft constraint score refers to the addition and subtraction rules based on engineering experience; the final score of the solution is a weighted fusion of the neural network prediction probability and the rule score, as shown in the formula: ; In the formula, PNN is the confidence level of the scheme predicted by the neural network, with a value of 0~100; Sr is the soft constraint rule score, with a value of 0~100; λ is the model weight coefficient, the initial value can be set to 0.7, and the system can dynamically adjust it according to the actual verification data using the Bayesian update mechanism.

[0016] Secondly, embodiments of the present invention provide a method for backfilling and recycling residual ore based on multi-layer intelligent decision-making, which is implemented using the above-mentioned system and includes the following steps: S1, Data Acquisition and Preprocessing Steps: Collect mine geological, production and environmental data, perform standardization processing, and output a standard feature matrix; S2, graded evaluation steps: The mine is evaluated in three levels from the mine level, intermediate level to the stope level, and the three levels of recoverability indicators are calculated in turn to realize the quantitative evaluation of residual ore recovery potential and engineering geological conditions from macro to micro. S3, Intelligent Prediction Step: Based on the standard feature matrix of step S1 and the hierarchical evaluation results of step S2, a deep neural network is used to model the spatial relationship of the mining area, and multiple tasks are cascaded to predict suitable filling methods, material types and proportion parameters, and the safety of the scheme is verified by strength check. S4, Rule Correction Steps: Combining the expert rule base, the predicted schemes are verified and optimized for safety through hard constraint verification, soft constraint scoring, and comprehensive weighted fusion, and a corrected scheme ranking list is output. S5, Economic Benefit Evaluation Steps: Conduct a life-cycle cost-benefit analysis on the corrected scheme; S6, Monitoring and Feedback Steps: Output a visual decision report and achieve a closed loop for solution optimization through real-time monitoring and feedback mechanisms; In step S2, if the historical data samples of the mine are insufficient or the category distribution is uneven, data augmentation processing is required. A strategy combining oversampling technology and generative adversarial networks with physical information embedding is adopted to generate high-quality synthetic samples that meet physical constraints and expand the training dataset.

[0017] Beneficial effects: 1. This invention breaks through the subjective limitations of traditional backfill design, which heavily relies on the personal experience of engineers, by constructing a multi-layered decision-making framework encompassing mine-level macro-evaluation, intermediate-level evaluation, and stope-level evaluation. The system can automatically optimize evaluation parameters based on historical mine data, reducing errors from manually set thresholds and ensuring that decision-making results conform to the objective laws of geological engineering, flexibly adapting to the specific environments of different mines. It achieves step-by-step refinement and optimization from macro-level resource potential assessment to micro-level stope scheme design, solving the problem of the lack of systematic decision-making logic in existing technologies.

[0018] 2. This invention innovatively introduces multiple evaluation methods for parallel analysis of the same mining area, and identifies discrepancies between evaluation results through a consistency check mechanism, triggering verification or weight adjustment when necessary, effectively reducing the decision-making risk caused by the failure of a single method. It innovatively integrates fuzzy comprehensive evaluation method and Mathews stability graph method, reducing the inherent defects of a single evaluation method through consistency checks and Bayesian weight adaptive adjustment mechanisms. It introduces state-varying weight functions and parameter adaptive fitting functions to automatically optimize evaluation criteria, reducing subjective errors from manually set thresholds and making evaluation results more consistent with the actual geological laws of the mine.

[0019] 3. This invention innovatively introduces a data augmentation strategy combining SMOTE and PI-MAC-GAN, which can expand the minority class samples and generate high-quality simulation data that conforms to the real physical distribution. Combined with a quality filter to remove noise, it improves the prediction accuracy and robustness of the system under complex and variable geological conditions. It not only expands the training sample size but also ensures the engineering rationality of the synthesized samples through a physical constraint verification layer. The quality filter further removes noisy samples, significantly improving the model's generalization ability and prediction accuracy under complex geological conditions.

[0020] 4. After the intelligent model outputs the filling scheme, this invention introduces engineering strength verification and safety threshold criteria to automatically correct or eliminate schemes that do not meet engineering safety requirements. This avoids engineering risks caused by relying solely on data-driven models and improves the safety and reliability of the filling scheme. The rule correction module has a built-in hard constraint rule library based on national standards, which can veto high-risk schemes. At the same time, it incorporates engineering experience through soft constraint scoring, forming a triple safety guarantee of model prediction, strength verification, and rule validation, effectively avoiding engineering risks.

[0021] 5. This invention incorporates a full life-cycle economic benefit assessment module, supporting the calculation of economic indicators such as Net Present Value (NPV) and dynamic sensitivity analysis. It provides mining enterprises with comprehensive decision support that meets both technical and safety requirements while achieving optimal economic benefits, thus improving management efficiency. Through a six-level economic benefit assessment indicator system, it comprehensively considers financial returns, cost structure, risk level, and environmental benefits, supporting Monte Carlo risk analysis and dynamic sensitivity analysis to help mines select the optimal economic solution while ensuring safety and compliance. Recommendations for tailings resource utilization and low-carbon backfilling schemes further reduce environmental impact and align with the requirements for green mine construction.

[0022] 6. This invention incorporates a monitoring, early warning, and feedback closed-loop module, constructing a PDCA cycle (Plan-Do-Check-Act) for construction monitoring, result verification, and model updates. Through a three-level early warning mechanism, it responds promptly to abnormal situations, and the incremental learning mechanism transforms measured data into the driving force for model optimization, enabling the decision-making system to continuously accumulate experience and improve performance as the mining process progresses.

[0023] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, and in order to make the above and other objects, features and advantages of the present invention more apparent and understandable, specific embodiments of the present invention are described below. Attached Figure Description

[0024] To more clearly illustrate the technical solution of the present invention, the accompanying drawings used in the present invention will be briefly described below. Obviously, the drawings described below are merely some embodiments of the present invention, and those skilled in the art can obtain other drawings based on these drawings without any creative effort.

[0025] Figure 1 This is a schematic diagram of the overall logical architecture and data flow of the residual ore filling and recycling system based on multi-layer intelligent decision-making provided in Embodiment 1 of the present invention.

[0026] Figure 2 This is a schematic diagram of the hierarchical evaluation module (three-level architecture and variable weight fuzzy evaluation process) provided in Embodiment 1 of the present invention.

[0027] Figure 3 This is a schematic diagram of the PI-MAC-GAN network structure (including the physical constraint layer) provided in Embodiment 1 of the present invention.

[0028] Figure 4 This is a schematic diagram of a neural network prediction module based on graph neural networks and multi-task cascaded prediction provided in Embodiment 1 of the present invention.

[0029] Figure 5 The flowchart of the three-stage fusion of "hard constraints + soft scoring" for the rule correction module provided in Embodiment 1 of the present invention is shown.

[0030] Figure 6 The multi-level indicator system for economic evaluation and the Monte Carlo simulation flowchart provided in Embodiment 1 of the present invention.

[0031] Figure 7 This is a schematic diagram of the monitoring, early warning, and feedback closed-loop module provided in Embodiment 1 of the present invention.

[0032] Figure 8 This is a schematic diagram of the human-computer interaction interface design provided in Embodiment 1 of the present invention. Detailed Implementation

[0033] The embodiments of the technical solution of the present invention will now be described in detail with reference to the accompanying drawings. These embodiments are merely illustrative of the technical solution of the present invention and are therefore intended to limit the scope of protection of the present invention.

[0034] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains; the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the invention; the terms “comprising” and “having”, and any variations thereof, in the specification, claims, and foregoing description of the invention, are intended to cover non-exclusive inclusion.

[0035] In the description of the embodiments of this invention, technical terms such as "first" and "second" are used only to distinguish different objects and should not be construed as indicating or implying relative importance or implicitly specifying the number, specific order, or primary and secondary relationship of the indicated technical features. In the description of the embodiments of this invention, "multiple" means two or more, unless otherwise explicitly defined.

[0036] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of the invention. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0037] In the description of the embodiments of this invention, the term "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. Additionally, the character " / " in this document generally indicates that the preceding and following related objects have an "or" relationship.

[0038] In the description of the embodiments of the present invention, the term "multiple" refers to two or more (including two), similarly, "multiple groups" refers to two or more (including two groups), and "multiple pieces" refers to two or more (including two pieces).

[0039] In the description of the embodiments of the present invention, the technical terms "center," "longitudinal," "lateral," "length," "width," "thickness," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," "outer," "clockwise," "counterclockwise," "axial," "radial," and "circumferential" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing the embodiments of the present invention and simplifying the description, and are not intended to indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the embodiments of the present invention.

[0040] In the description of the embodiments of the present invention, unless otherwise explicitly specified and limited, the technical terms such as "installation," "connection," "joining," and "fixing" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication of two components or the interaction between two components. Those skilled in the art can understand the specific meaning of the above terms in the embodiments of the present invention according to the specific circumstances.

[0041] To address the technical problems of relying on engineers' experience in formulating residual ore backfilling and recovery schemes, which are characterized by strong subjectivity, significant uncertainty, lack of unified scientific basis, and the absence of tiered decision-making and verification integration, making it difficult to coordinate safety and benefits, this invention provides a method and system for residual ore backfilling and recovery based on multi-level intelligent decision-making. It quantifies the potential for residual ore recovery and geological conditions through a three-level evaluation system at the mine, intermediate, and stope levels; employs data augmentation strategies to overcome the bottleneck of small samples; utilizes graph neural networks for multi-task cascaded prediction of backfilling schemes and performs strength verification; optimizes the scheme through hard constraints + soft scoring rules correction and full life-cycle economic benefit assessment; and forms a closed loop based on monitoring and early warning and incremental learning, achieving scientific, intelligent, safe, and economical decision-making for residual ore backfilling and recovery.

[0042] Example 1 Please refer to Figures 1 to 8 As shown in Embodiment 1 of the present invention, a modularly designed multi-layer intelligent decision-making residual ore backfilling and recovery system is proposed. Its overall structure includes the following modules: data acquisition and preprocessing module, hierarchical evaluation module, data augmentation module, neural network prediction module, rule correction module, economic benefit evaluation module, monitoring, early warning and feedback closed-loop module, and human-computer interaction module.

[0043] The data acquisition and preprocessing module serves as the data starting point, responsible for importing mine geological, production, and environmental data, performing intelligent field mapping, missing value imputation, and outlier detection, and outputting a standard feature matrix. This matrix is ​​then input into the hierarchical evaluation module for layer-by-layer evaluation from the mine level, intermediate level, to the stope level, calculating recoverability indicators to provide a basis for subsequent decision-making. If historical data is insufficient, the data augmentation module intervenes to generate synthetic samples to ensure data quality. The neural network prediction module predicts filling schemes based on preprocessed data and evaluation indicators using a graph neural network model. The rule correction module performs hard and soft constraint verification on the prediction results to ensure safety. The economic benefit evaluation module conducts a full life-cycle cost-benefit analysis on the corrected prediction scheme. The monitoring, early warning, and feedback closed-loop module monitors construction parameters in real time, triggers a three-level early warning mechanism, and uses feedback data for incremental model learning to achieve optimization closed loop. The human-computer interaction module runs throughout the entire process, providing a visual interface and report generation function, supporting user interaction and parameter adjustment. All modules are seamlessly connected through standardized interfaces or communication links, forming a collaborative workflow from data input to decision output. Its overall architecture and data flow are as follows: Figure 1 As shown.

[0044] The data acquisition and preprocessing module imports geological, production, and environmental data from the mine, performs intelligent field recognition and standardized mapping, completes missing value imputation, outlier detection, and data normalization, and outputs a standard feature matrix to provide high-quality input for subsequent analysis.

[0045] Specifically, the data acquisition and preprocessing module can import various basic data from the mining residual ore field through structured tables such as Excel and CSV, including geological structure and surrounding rock mechanical parameters (such as rock integrity index, rock strength, geostress distribution, number of joint groups, joint roughness, joint alteration coefficient, etc.), ore body occurrence conditions (ore body shape and size, burial depth, location and geometric parameters of adjacent goaf areas), goaf geometric parameters (span, height, length, hydraulic radius), ore grade and reserves, current mining support status and mining process conditions, etc.

[0046] In some specific implementations, the data preprocessing module incorporates an intelligent field mapping function. This function includes a built-in thesaurus of mining terminology and a string matcher based on edit distance. When a user-imported column name (e.g., Cu%) differs from the system's standard field name (e.g., average grade), the system automatically identifies their similarity and completes the mapping, eliminating the need for manual adjustment by the user.

[0047] Furthermore, the data acquisition and preprocessing module uses the MissForest algorithm (random forest missing value imputation algorithm) or multiple imputation method (MICE) to process missing data, and uses the interquartile range (IQR) method to automatically identify and correct outliers.

[0048] The graded evaluation module is used to evaluate the overall remaining resources of the mine, including a mine-level macro-evaluation unit, a mid-level evaluation unit, and a stope-level evaluation unit. The mine-level macro-evaluation unit comprehensively considers factors such as reserves, grade, ore body structure, and goaf distribution to calculate the mine-level recoverability index (R0) and assess the recovery potential of the entire mine's residual ore resources. The mid-level evaluation unit quantifies the geological conditions and mining sequence of each mid-level section and calculates the mid-level evaluation index (M0). The stope-level evaluation unit quantifies the engineering geological conditions of a specific stope and calculates the stope-level evaluation index (R1).

[0049] In some specific implementations, the hierarchical evaluation module employs a fuzzy comprehensive evaluation method to configure a state-variable weighting function for each evaluation index, supporting adaptive parameter fitting. The core calculation process is as follows: Input the original index x of each parameter i Perform extreme value testing and calculate the state-weighted function: ; Perform weight normalization: ; Perform a weighted comprehensive calculation to output a risk level score R, and determine the risk level based on the range of values ​​for R: ; in, The mean; w i The initial weights for the i-th indicator are assigned by experts; w' i The adjusted weights; S(x) i ) is the state-varying weighting function; μ i σ represents the membership degree of the i-th indicator; σ represents the standard deviation; f represents the specific functional form of the state-variable weighting function; j represents the evaluation level number; and n represents the total number of evaluation indicators.

[0050] The state-weighted function is used to dynamically adjust the weights according to the degree to which the index value deviates from the normal range. This invention supports multiple function forms (Gaussian, exponential decay, linear, S-shaped, and trapezoidal), which users can choose according to the actual data characteristics.

[0051] The hierarchical evaluation module can utilize historical labeled data to automatically optimize the morphological parameters (such as mean and variance) of the membership function by minimizing the classification error objective function, thereby reducing the subjective error of manually setting thresholds.

[0052] The classification criteria for the mine-level recoverability index R0 are as follows: The intermediate-level evaluation unit is used to establish a transition layer between mine-level and stope-level evaluations, enabling a progressively refined evaluation from macro to micro levels. The intermediate-level evaluation index M0 comprehensively considers the following four types of factors: The formula for calculating the intermediate-level evaluation index M0 is: M 0 = Σ(w i × μ i ) × λ s × λ a ; Among them, w i The initial weights for the i-th indicator are assigned by experts; μ i λ represents the membership degree of the i-th index. s λ is a correction factor for the mining sequence, ranging from 0.8 to 1.2. When the upper and middle sections have been filled, λ... s =1.2, when the upper and middle sections have not yet been mined. s 0.8; λ a The influence coefficient of adjacent middle sections is calculated based on the volume and distance of the goaf in adjacent middle sections, and the value ranges from 0.7 to 1.0.

[0053] The intermediate-level evaluation results are divided into four levels: Level I (M0≥0.8) priority mining, Level II (0.6≤M0<0.8) normal mining, Level III (0.4≤M0<0.6) cautious mining, and Level IV (M0<0.4) temporary mining. The intermediate-level evaluation results are used to determine the mining priority order of each intermediate level and to provide boundary condition constraints for the stope-level evaluation.

[0054] The graded evaluation module introduces the Mathews stability graph method as a comparative verification mechanism in the stope-level evaluation (R1) to improve the system's acceptance in the field of mining engineering and ensure the reliability of the system's calculation results.

[0055] The core calculation formula for the Mathews stability graph method is: ; The meanings of each parameter are as follows: The system locates the goaf on the Mathews stability map based on the calculated N' and HR values, and determines whether the goaf belongs to the stable zone, transition zone, or collapse zone.

[0056] The consistency test between the evaluation result R1 of the fuzzy comprehensive evaluation method and the Mathews stability graph determination result is performed: If the two methods yield the same result, then output a high-confidence evaluation result; If there is a discrepancy in the judgment results, the system will trigger a manual review prompt and simultaneously output the detailed calculation process of both methods for engineers to refer to and make decisions.

[0057] The graded evaluation module establishes a complete evaluation index system for goaf areas, including the following four categories of indicators: The hierarchical evaluation module also includes a method weight adaptive unit. This unit uses historical validation data and a Bayesian update mechanism to dynamically adjust the weight ratio of fuzzy evaluation and the Mathews method in the final fusion decision. The multi-method fusion evaluation formula for the mining area is: ; Where a, b, and c are the fusion weights of fuzzy comprehensive evaluation, Mathews stable graph method, and RMR score, respectively, a+b+c = 1, and the initial weights can be set to a=0.4, b=0.4, and c=0.2. The weights can be adaptively adjusted according to historical validation data. The stability score obtained from the fuzzy comprehensive evaluation at the mining site level (normalized to 0-1); The stability level corresponding to the Mathews stability graph method is mapped to a value between 0 and 1; RMR rock mass quality grade normalization score ( =RMR / 100), RMR (Rock Mass Rating) is the geomechanical classification index value of rock mass.

[0058] The data augmentation module addresses the issues of limited historical mine data samples and uneven category distribution by generating high-quality synthetic samples that conform to the true distribution, thereby expanding the training dataset.

[0059] In some specific implementations, the data augmentation module addresses the problem of insufficient data samples in the filling scheme by using SMOTE (Synthetic Minority Oversampling) to interpolate and augment imbalanced samples, and combines it with a multi-attribute conditional generative adversarial network (PI-MAC-GAN) embedded with physical information to generate new simulation samples, thereby expanding the training dataset and improving the model's generalization ability.

[0060] The neural network prediction module uses a deep neural network to predict the most suitable filling method and filling material type based on the input standard features and evaluation features.

[0061] The neural network prediction module uses a graph neural network (GNN) to model the spatial relationships of the mining area. Node features include surrounding rock strength, RQD, grade, thickness, burial depth, risk score, and void ratio; edge features include distance, same-layer relationship, and void spacing.

[0062] The network structure employs a graph neural network encoder and a multi-task cascaded prediction head, including GCN (Graph Convolutional Network) layers, GAT (Graph Attention Network) layers, and SE-Block (Compression-Excitation) channel attention modules. It utilizes progressive multi-task learning: the first level predicts the filling method type, the second level predicts material proportions based on the method, and the third level predicts detailed process parameters based on both the method and materials. The cascaded structure leverages inter-task dependencies, with upstream prediction results serving as conditional inputs for downstream tasks.

[0063] Before outputting the backfill material ratio, the neural network prediction module needs to be verified by the strength verification unit. This unit, based on Mitchell's strength design theory, calculates the minimum strength requirement of the backfill body according to the stope geometry parameters. The calculation formula is as follows: ; Where, σ c,min γ is the minimum uniaxial compressive strength required for the filling material (kPa), γ is the bulk density of the filling material (kN / m³), H is the height of the filling material (m), L is the exposed length of the filling material (m), W is the width of the filling material (m), and φ is the internal friction angle of the filling material (°). In practical applications, it is usually necessary to multiply by a safety factor FS, which is generally taken as 1.5~2.0.

[0064] The expected strength σ corresponding to the matching scheme predicted by the neural networkc,pred σ must be satisfied c,pred ≥ σ c,min If the conditions are met, the system will automatically adjust the mixing parameters or trigger manual review.

[0065] The neural network prediction module integrates the Shapley Additive Explanations (SHAP) and Local Interpretable Model-agnostic Explanations (LIME) analysis units. This unit can output the contribution of each input feature to the prediction result, generate feature importance ranking charts and decision path visualizations, enabling engineers to understand the specific reasons why the model recommends a particular approach.

[0066] The neural network prediction module employs Monte Carlo or deep ensemble methods to implement Bayesian uncertainty quantification. The system outputs not only the point estimate prediction value but also the 95% confidence interval and failure probability of the prediction result. The output format is as follows: When the prediction uncertainty exceeds the set threshold (e.g., confidence interval width > 20%), the system will automatically trigger a manual review prompt and suggest supplementing the field survey data.

[0067] The rule correction module is used to verify the neural network prediction results in terms of safety, operability, and engineering experience. Its purpose is to ensure that the final recommended filling scheme not only has high model prediction confidence but also meets safety regulations, the suitability of filling materials, and the constraints of on-site mining conditions. The expert rule base contains experience-based decision-making rules based on geological characteristics and production requirements.

[0068] The rule correction module employs a three-stage correction strategy: hard constraints, soft constraint scoring, and comprehensive weighted fusion. Hard constraints refer to rules related to safety regulations, policy red lines, or equipment limits; once triggered, they have a veto power over the solution. Soft scoring refers to bonus and deduction rules based on engineering experience. The final solution score is a weighted fusion of the neural network prediction probability and the rule score, using the following formula: ; Wherein, PNN is the confidence level of the scheme predicted by the neural network (0~100), Sr is the soft constraint rule score (0~100), and λ is the model weight coefficient. The initial value can be set to 0.7, and the system can dynamically adjust it according to the actual verification data using a Bayesian update mechanism.

[0069] The hard constraint rule base is compiled based on national and industry standards such as the "Safety Regulations for Metal and Non-metal Mines" and the "Design Code for Non-ferrous Metal Mining".

[0070] The soft constraints refer to the scoring rules based on engineering experience, including recommended rules (S1-S7) and empirical rules (E1-E3).

[0071] The economic benefit assessment module is used to perform a full life-cycle cost-benefit analysis on technically feasible recommended solutions, calculate financial indicators, and evaluate parameter sensitivity.

[0072] The economic benefit assessment module includes a six-level indicator system, which comprehensively considers the cost of filling materials, construction costs, and mining revenue, providing an economic decision-making basis for the optimal selection of the scheme.

[0073] The economic benefit assessment module uses Monte Carlo simulation (N≥10,000 times) for risk analysis.

[0074] The economic assessment module supports dynamic sensitivity analysis. Users can select key parameters (such as metal market prices, cement costs, and labor costs) on the interface, and the system will automatically calculate the impact of these parameters on net present value (NPV) within a certain fluctuation range (such as ±10%), and generate a sensitivity analysis tornado diagram in real time.

[0075] The monitoring, early warning and feedback closed-loop module is used for continuous monitoring, early warning and scheme optimization after construction, forming a complete Plan-Do-Check-Act (PDCA) closed loop.

[0076] The monitoring, early warning, and feedback closed-loop module includes a three-level early warning mechanism, and the early warning mechanism and indicators are as follows.

[0077] The three-tiered early warning mechanism corresponds to different early warning response measures. Level 1: Increase monitoring frequency and strengthen patrols; Level 2: Local grouting reinforcement and parameter fine-tuning; Level 3: Trigger feedback optimization and redesign the plan.

[0078] The monitoring, early warning, and feedback closed-loop module also includes an on-site verification indicator system, as detailed below.

[0079] The monitoring, early warning, and feedback closed-loop module also includes a model incremental learning mechanism. When the deviation between the predicted result and the measured result exceeds a threshold (e.g., |predicted - measured| > 15%), the system triggers model retraining; otherwise, only the knowledge base weights are updated. Newly added validation data is automatically included in the training set to achieve continuous model optimization.

[0080] The human-computer interaction module is used to provide a visual operation interface, parameter configuration window, and standardized decision report generation function based on the system analysis results.

[0081] The human-computer interaction module supports templated report generation. The system provides three preset report templates: a simplified version (only calculation results), a standard version (including main parameters and calculation results), and an engineering version (including detailed calculation process, rule correction list, and sensitivity chart).

[0082] The human-computer interaction module automatically generates a filling decision report, which includes the evaluation results of each module, recommended filling schemes, economic indicators, etc., and is exported in text format for users to archive or further edit.

[0083] The human-computer interaction module provides a paginated graphical user interface that displays multiple charts in real time, including geological parameters, evaluation results, and prediction schemes for the mine and mining areas. The interface supports importing expert-weighted data from Excel files, allowing users to adjust the weights of fuzzy evaluation indicators based on their practical experience, enabling flexible interaction.

[0084] Example 2 Embodiment 2 of this invention proposes an intelligent backfilling method for residual ore in metal mines based on multi-layer intelligent decision-making. Relying on the modularly designed intelligent system described in Embodiment 1, the specific implementation process includes the following steps: S1: Users import basic mine data in Excel or CSV format through the human-computer interaction module interface. The data acquisition and preprocessing module performs intelligent field recognition, fills in missing data, corrects outliers that deviate from the normal range, and finally normalizes all continuous numerical data, outputting a standard feature matrix.

[0085] S2, the graded evaluation module first performs a macro-evaluation of the mine, calculates the mine-level recoverability index R0, then evaluates each intermediate section, calculates the intermediate section-level evaluation index M0 to determine the priority order of mining, and finally, for specific stope units to be filled, uses the fuzzy comprehensive evaluation method combined with the Mathews stability diagram method to output the stope-level evaluation index R1 and classification score.

[0086] If the historical backfilling case data of the current mine is insufficient or the category distribution is uneven, the data augmentation module is activated, and a high-quality expanded training dataset is constructed using a strategy combining SMOTE and PI-MAC-GAN.

[0087] In step S3, the standard feature matrix obtained in step S1 and the evaluation index obtained in step S2 are used as inputs and fed into the neural network prediction module. The graph neural network performs message passing and feature aggregation based on the spatial adjacency relationship between mining areas. The cascaded prediction heads sequentially output the predicted probabilities of the filling method, material type, and mix proportion parameters. The system outputs preliminary recommended schemes based on the probability levels. The recommended schemes include: filling method (such as cemented filling, waste rock filling, paste filling, etc.), filling material type (such as whole tailings, graded tailings, waste rock aggregate, etc.), and mix proportion parameters (ash-sand ratio, slurry concentration, admixture dosage, etc.), along with the prediction confidence level and uncertainty interval.

[0088] S4, the rule correction module receives the preliminary recommended schemes from the neural network and calls the expert rule base for verification. First, a hard constraint check is performed. If any hard constraint rule (such as H1-H7) is triggered, the corresponding scheme is forcibly excluded. Schemes that pass the hard constraints enter the soft scoring stage, where applicable soft constraint rules (such as S1-S6, E1-E3) are matched according to the mining conditions, and the rule score is calculated. Finally, the neural network prediction probability and the rule score are combined to output a corrected ranking list of recommended schemes.

[0089] S5, the economic benefit assessment module, performs a full life cycle cost analysis on the technically feasible solutions selected in step S5.

[0090] The S6 human-computer interaction module displays geological parameters, grading evaluation results, recommended schemes, and economic indicator charts in real time on a paginated graphical user interface, and generates decision reports in .txt or PDF format that include detailed calculation processes and sensitivity charts.

[0091] In this embodiment, a copper mine with an underground mining depth of 850m and an average ground temperature of 38℃ is taken as an example. Based on a multi-layer intelligent decision-making system for intelligent backfilling of residual ore in metal mines, the system performs data preprocessing, hierarchical evaluation, backfilling scheme prediction, rule correction, and economic benefit assessment for this mining area. The specific steps are as follows: S1: The user imports raw exploration data for stope No. 802. The system automatically cleans the data, fills in missing data through field mapping and the MissForest algorithm, and obtains a normalized matrix. The mine parameter conversion table is as follows: The missing RQD values ​​were imputed using the MissForest algorithm to obtain 45% of the data; the underground ambient temperature, ore body dip angle, and copper grade were normalized and then incorporated into the standard feature matrix. This standard feature matrix subsequently served as input to both the grading evaluation module and the neural network prediction module.

[0092] S2, import the normalized matrix into the graded evaluation module to perform mine-level evaluation ( R0) Intermediate level evaluation ( M 0) and mining area level evaluation ( R 1); Four indicators—reserves, grade, depth, and hydrological conditions—were selected, with a weight vector of W0=[0.3,0.3,0.2,0.2]. Based on normalization and membership calculations, the evaluation values ​​for the four indicators are 0.8, 0.75, 0.9, and 0.6, respectively. The calculated R0=0.3×0.8+0.3×0.75+0.2×0.9+0.2×0.6=0.765. According to the mine-level evaluation and grading standards, the area where mine No. 802 is located belongs to the medium recovery potential category.

[0093] Intermediate-level evaluation: Stope No. 802 is located in the -650m intermediate level. The upper intermediate level (-600m) has been filled, and the lower intermediate level (-700m) is currently being mined. The intermediate-level evaluation comprehensively considers four categories of indicators: geological conditions, resource endowment, mining sequence, and infrastructure. The weights and membership degrees are [0.3, 0.25, 0.25, 0.2] and [0.62, 0.75, 0.72, 0.685], respectively. The mining sequence correction coefficient λ is used. s =1.15, influence coefficient λ of adjacent middle sections a =0.92, the comprehensive evaluation M0=(0.3×0.62+0.25×0.75+0.25×0.72+0.2×0.685)×1.15×0.92=0.73, which belongs to Level II normal mining.

[0094] Using an adaptive Gaussian membership function, the surrounding rock stability membership function parameters obtained by fitting historical data of this mine are: mean =65, standard deviation σ=15. Input RQD=45 for mining area 802, and calculate membership degree using Gaussian membership function: ; Combining other indicators, the fuzzy comprehensive evaluation score at the mining area level is calculated as R. 1fuzzy =0.36; According to the system's built-in Mathews stability graph threshold library, R 1Mathews =0.40; RMR score is 35, then R 1RMR =0.35. With weights set as a=0.40, b=0.40, c=0.20, the comprehensive stability score for stope No. 802 is R1=0.40×0.36+0.40×0.40+0.20×0.35=0.38, classifying it as Level IV. This indicates that stope No. 802 has poor local stability and is classified as a transitionally unstable stope. Low-strength or non-cemented backfilling schemes are not suitable; high-strength cemented backfilling should be prioritized and safety checks should be conducted.

[0095] Since there are only 128 historical mine filling records, sample augmentation and network training are required. SMOTE and a multi-attribute conditional generative adversarial network (PI-MAC-GAN) with physical information embedding are used to generate 500 simulation samples. After screening and filtering, 420 high-quality samples are retained to expand the neural network training set. S3. Construct a stope relationship diagram, with stope 802 as the central node, and establish edge connections with adjacent stopes 801, 803, 702, and 902. The graph neural network is input with node features (15 dimensions including surrounding rock strength and RQD) and edge features (6 dimensions including stope spacing and intra-layer relationships). The graph neural network prediction results are as follows: Mitchell strength verification: Based on the stope geometry parameters (H=50m, L=30m, W=80m, tanφ=1.5, γ=18kN / m) 3 ), calculate the minimum strength requirement σ c,min = (18×50×30) / (2×1.5×80) = 0.1125MPa. The graph neural network predicts that the 28-day filling strength of this scheme is 2.5 MPa > 0.1125MPa, which meets the safety requirements.

[0096] S4, the rule correction module performs the following verification process.

[0097] Hard constraint check: The sulfur content of the ore was detected to be 0.8% (<2% threshold), and the H2 rule was not triggered; Mathews determined the transition zone and triggered the H7 rule, which forced high-strength cemented filling, consistent with the neural network prediction result, and passed the hard constraint check.

[0098] Soft constraint scoring: Soft constraint scoring starts at 0 points and accumulates according to trigger rules, with a maximum of 100 points. Trigger rules S2 (grade 1.45% > 1%, +15 points), S3 (pillar recovery condition, +20 points), S7 (Mathews determines it as a transition boundary, allowing for a high-strength cemented solution, +15 points), rule score S... r = 50 points. The prediction confidence of the neural network for the whole tailings cemented backfill scheme is: 92 points. The system adopts adaptive weight λ=0.72, and the comprehensive score is: S = 0.72×92+0.28×50=66.24+14 =80.24 points.

[0099] The final recommended solution is: full tailings cemented backfill, with a cement-sand ratio of 1:6, a slurry concentration of 72%, a predicted 28-day strength of 2.5 MPa, and a comprehensive score of 80.24.

[0100] S5, Economic Benefit Assessment Module Calculation Results: Net Present Value (NPV) = 12.5 million yuan, Internal Rate of Return (IRR) = 18.5%, Investment Recovery Period (Pt') = 4.2 years. Monte Carlo Simulation (N = 10,000 times) Results: Expected Net Present Value (NPV) E[NPV] = 11.8 million yuan, Probability of NPV > 0 = 94.2%, Value at Risk (VaR) (95%) = -1.2 million yuan. Sensitivity analysis shows that copper price fluctuations have the greatest impact on NPV (elasticity coefficient 1.8).

[0101] S6 generates an engineering version of the decision report, which includes a complete evaluation calculation process, a list of rule corrections, a summary of economic indicators, and a sensitivity analysis tornado diagram.

[0102] In summary, this invention discloses a method and system for residual ore backfilling and recovery based on multi-level intelligent decision-making, relating to the field of mining technology. The system comprises eight modules: data acquisition and preprocessing, hierarchical evaluation, data augmentation, neural network prediction, rule correction, economic benefit assessment, monitoring and early warning with feedback loop, and human-computer interaction. The method flow is as follows: importing basic mine data and outputting a standard feature matrix after preprocessing; quantifying the residual ore recovery potential and engineering geological conditions through a three-level evaluation system of mine-level, intermediate-level, and stope-level; expanding the dataset using a combination of SMOTE and PI-MAC-GAN when samples are insufficient; predicting backfilling schemes using a multi-task cascaded graph neural network and verifying their strength; correcting schemes through hard constraint verification and soft constraint scoring; conducting full life-cycle economic benefit analysis; outputting a visualized decision report and continuously optimizing the model through post-construction monitoring. This invention constructs a multi-level, multi-method integrated intelligent decision-making system, solving the problems of existing technologies relying on experience and lacking a systematic decision-making mechanism, and improving the scientific, safe, and economically rational nature of residual ore backfilling schemes.

[0103] It should be noted that the present invention is not limited to the above-described embodiments. The above embodiments are merely examples, and any embodiments that have the same structure and perform the same effects as the technical concept within the scope of the present invention are included within the scope of the present invention. Furthermore, various modifications that can be conceived by those skilled in the art to the embodiments, and other ways of constructing by combining some of the constituent elements of the embodiments, without departing from the spirit of the present invention, are also included within the scope of the present invention.

Claims

1. A residual ore backfilling and recovery system based on multi-layer intelligent decision-making, characterized in that, include: The data acquisition and preprocessing module is used to import geological, production, and environmental data from the mine, perform data preprocessing, and output a standard feature matrix. The hierarchical evaluation module, connected to the data acquisition and preprocessing module, includes mine-level macro evaluation units, intermediate-level evaluation units, and stope-level evaluation units. It is used to calculate mine-level recoverability indicators, intermediate-level evaluation indicators, and stope-level evaluation indicators in sequence, so as to realize the quantitative evaluation of residual ore recovery potential and engineering geological conditions from macro to micro. The neural network prediction module is connected to the data acquisition and preprocessing module and the hierarchical evaluation module, respectively. It is used to model the spatial relationship of the mining area and predict suitable filling schemes, material types and proportion parameters in a multi-task cascade based on the standard feature matrix and evaluation index. The rule correction module, connected to the neural network prediction module, has a built-in expert rule base, including a hard constraint verification unit, a soft constraint scoring unit, and a comprehensive weighted fusion unit. It is used to correct the safety and engineering feasibility of the predicted scheme through hard constraint verification, soft constraint scoring, and comprehensive weighted fusion, and output a corrected scheme ranking list. The economic benefit assessment module, connected to the rule correction module, is used to perform full life-cycle cost-benefit analysis, risk analysis, and sensitivity analysis on technically feasible solutions. The monitoring, early warning and feedback closed-loop module, connected to the rule correction module, is used for post-construction monitoring and early warning, on-site verification and incremental model learning, forming a continuously optimized decision-making closed loop; The human-computer interaction module is connected to the monitoring, early warning and feedback closed-loop module and the economic benefit evaluation module, respectively, and is used to provide a visual interface, parameter configuration and decision report generation functions.

2. The residual ore backfilling and recovery system based on multi-layer intelligent decision-making according to claim 1, characterized in that, It also includes a data augmentation module; the data augmentation module is connected to the hierarchical evaluation module and is used to generate high-quality synthetic samples that meet physical constraints and expand the training dataset when there are insufficient samples of historical mining data or uneven distribution of categories.

3. The residual ore backfilling and recovery system based on multi-layer intelligent decision-making according to claim 1, characterized in that, The hierarchical evaluation module uses the fuzzy comprehensive evaluation method to configure a state-variable weighting function for each evaluation index, supporting adaptive parameter fitting. The calculation process is as follows: Input the original index x of each parameter i Perform extreme value testing and calculate the state-weighted function: ； Perform weight normalization: ； Perform a weighted comprehensive calculation to output a risk level score R, and determine the risk level based on the range of values ​​for R: ； in, The mean; w i The initial weights for the i-th indicator are assigned by experts; w' i The adjusted weights; S(x) i ) is the state-varying weighting function; μ i σ is the membership degree of the i-th indicator; f is the standard deviation; f is the specific functional form of the state-variable weighting function; j is the evaluation level number; n is the total number of evaluation indicators. The fuzzy comprehensive evaluation method supports Gaussian, exponential decay, linear, sigmoid, and trapezoidal function forms, and can automatically optimize the morphological parameters of membership functions using historical labeled data.

4. A residual ore backfilling and recovery system based on multi-layer intelligent decision-making according to claim 3, characterized in that, In the graded evaluation module, the mine-level macro-evaluation unit comprehensively considers factors such as reserves, grade, ore body structure and goaf distribution to calculate the mine-level recoverability index, which is divided into four levels: high recovery potential, medium potential, low recovery potential and not yet available. In the intermediate-level evaluation unit, the calculation formula for the intermediate-level evaluation indicators is as follows: M 0 = Σ(w i × μ i ) × λ s × λ a ； Where M0 is the mid-level evaluation indicator; λ s λ is a correction factor for the mining sequence, ranging from 0.8 to 1.

2. When the upper and middle sections have been filled, λ... s =1.2, when the upper and middle sections have not yet been mined. s =0.8; λ a The influence coefficient for adjacent intermediate sections is calculated based on the volume and distance of the goaf in adjacent intermediate sections, with a value ranging from 0.7 to 1.

0. The intermediate section-level evaluation results are divided into four levels, as follows: When M0 ≥ 0.8, priority is given to mining, and it is set as Level I; When 0.6≤M0<0.8, normal mining is performed and is set to Level II; When 0.4 ≤ M0 < 0.6, cautious mining should be adopted, and the level should be set to III. When M0 < 0.4, mining is temporarily suspended and set to Level IV; The results of the intermediate-level evaluation are used to determine the priority order of mining in each intermediate level and to provide boundary condition constraints for the stope-level evaluation.

5. A residual ore backfilling and recovery system based on multi-layer intelligent decision-making according to claim 4, characterized in that, The hierarchical evaluation module incorporates the Mathews stability graph method as a comparative verification mechanism in its mining area-level evaluation unit; the Mathews stability graph method uses the formula... Calculate the stability coefficient; in, N' Q' is the corrected stability coefficient; Q' is the corrected Q value, Q' = (RQD / J n ) × (J r / J a ), J n J is the number of joint groups. r J is the roughness coefficient of the most vulnerable joint. a RQD represents the degree of erosion or infill at the weakest joint surface; A is the rock integrity index; and A is the stress coefficient, A = f(σ c / σ1), σ c σ1 is the uniaxial compressive strength of the rock; B is the joint orientation coefficient; C is the gravity adjustment coefficient; HR is the hydraulic radius, HR = S / P = excavation face area / excavation face perimeter; Based on the calculated N' and HR values, the location of the goaf on the Mathews stability map is determined to be either a stable zone, a transition zone, or a collapse zone. The consistency of the evaluation results of the fuzzy comprehensive evaluation method with the Mathews stability graph determination results is verified: If the two methods yield the same result, then output a high-confidence evaluation result; If there is a discrepancy in the judgment results, the system will trigger a manual review prompt and simultaneously output the detailed calculation process of both methods for engineers to refer to and make decisions.

6. A residual ore backfilling and recovery system based on multi-layer intelligent decision-making according to claim 5, characterized in that, The hierarchical evaluation module also includes a method weight adaptive unit; the method weight adaptive unit uses historical validation data and a Bayesian update mechanism to dynamically adjust the weight ratio of fuzzy evaluation and Mathews stable graph method in the final fusion decision; the multi-method fusion evaluation formula is: ； Where a, b, and c are the fusion weights of the fuzzy comprehensive evaluation method, the Mathews stable graph method, and the RMR score, respectively, and a+b+c = 1. The initial weights can be set to a=0.4, b=0.4, and c=0.2, and the weights can be adaptively adjusted according to historical verification data. The stability score obtained from the fuzzy comprehensive evaluation at the mining site level is normalized to 0-1. The stability level corresponding to the Mathews stability graph method is mapped to a value between 0 and 1; For RMR rock mass quality grade normalization scoring, =RMR / 100, where RMR is the rock mass geomechanical classification index value; R1 is the final comprehensive stability score after the fusion of multiple methods at the stope level, with a value range of 0~1.

7. A residual ore backfilling and recovery system based on multi-layer intelligent decision-making according to claim 2, characterized in that, The data augmentation module includes a SMOTE algorithm unit and a PI-MAC-GAN unit. It adopts a strategy that combines the SMOTE algorithm with the PI-MAC-GAN multi-attribute conditional generative adversarial network embedded in physical information. The input is a concatenated vector of random noise vector and conditional vector, and the output is filling recipe parameters.

8. A residual ore backfilling and recovery system based on multi-layer intelligent decision-making according to claim 1, characterized in that, The neural network prediction module uses a graph neural network encoder and a multi-task cascaded prediction head, including a GCN layer, a GAT layer and an SE-Block channel attention module. The multi-task cascaded prediction sequentially realizes the prediction of filling method type, material ratio and process parameters. The cascaded structure utilizes the inter-task dependency relationship, and the upstream prediction result serves as the condition input for the downstream task. Before outputting the backfill material ratio, the neural network prediction module needs to be verified by the strength verification unit. Based on Mitchell's strength design theory, the minimum strength requirement of the backfill body is calculated according to the geometric parameters of the stope. The calculation formula is as follows: ； Where, σ c,min γ is the minimum uniaxial compressive strength required for the filling, H is the filling height, L is the exposed length of the filling, W is the filling width, and φ is the internal friction angle of the filling. In practical applications, it is usually necessary to multiply by the safety factor FS, which is generally taken as 1.5~2.

0. The expected strength σ corresponding to the matching scheme predicted by the neural network c,pred σ must be satisfied c,pred ≥ σ c,min If the conditions are met, the system will automatically adjust the mixing parameters or trigger manual review.

9. A residual ore backfilling and recovery system based on multi-layer intelligent decision-making according to claim 1, characterized in that, The hard constraint rules in the rule correction module refer to rules involving safety regulations, policy red lines, or equipment limits, which, once triggered, have a veto power over the solution; the soft constraint score refers to the addition and subtraction rules based on engineering experience; the final solution score is a weighted fusion of the neural network prediction probability and the rule score, as shown in the formula: ； In the formula, PNN is the confidence level of the scheme predicted by the neural network, with a value of 0~100; Sr is the soft constraint rule score, with a value of 0~100; λ is the model weight coefficient, with an initial value of 0.7, which can be dynamically adjusted by the system using a Bayesian update mechanism based on the actual validation data.

10. A method for backfilling and recovering residual ore based on multi-layer intelligent decision-making, characterized in that, The system described in any one of claims 1 to 9 is used for residual ore backfilling and recovery, comprising the following steps: S1, Data Acquisition and Preprocessing Steps: Collect mine geological, production and environmental data, perform standardization processing, and output a standard feature matrix; S2, graded evaluation steps: The mine is evaluated in three levels from the mine level, intermediate level to the stope level, and the three levels of recoverability indicators are calculated in turn to realize the quantitative evaluation of residual ore recovery potential and engineering geological conditions from macro to micro. S3, Intelligent Prediction Step: Based on the standard feature matrix of step S1 and the hierarchical evaluation results of step S2, a deep neural network is used to model the spatial relationship of the mining area, and multiple tasks are cascaded to predict suitable filling methods, material types and proportion parameters, and the safety of the scheme is verified by strength check. S4, Rule Correction Steps: Combining the expert rule base, the predicted schemes are verified and optimized for safety through hard constraint verification, soft constraint scoring, and comprehensive weighted fusion, and a corrected scheme ranking list is output. S5, Economic Benefit Evaluation Steps: Conduct a life-cycle cost-benefit analysis on the corrected scheme; S6, Monitoring and Feedback Steps: Output a visual decision report and achieve a closed loop for solution optimization through real-time monitoring and feedback mechanisms; In step S2, if the historical data samples of the mine are insufficient or the category distribution is uneven, data augmentation processing is required. A strategy combining oversampling technology and generative adversarial networks with physical information embedding is adopted to generate high-quality synthetic samples that meet physical constraints and expand the training dataset.

Images