Underground stoping ore body sinking construction method and system

Abstract

The invention relates to an underground stoping ore body sinking construction method and system, and belongs to the technical field of underground mining engineering. The method comprises the following steps: constructing a multi-dimensional feature system, and collecting stoping parameters, geological conditions and monitoring data feature parameters to form a time sequence sample set; training a subsidence prediction model based on bidirectional LSTM and an attention mechanism, inputting time sequence features, and outputting a future daily average subsidence prediction value; implementing hierarchical regulation and control based on a prediction result; establishing a model incremental learning mechanism, dynamically freezing bottom network parameters and finely tuning a top layer to realize dynamic adaptation; the system comprises a multi-source data acquisition unit, a feature engineering terminal, an intelligent prediction module, a regulation and control execution unit and a visual platform. Intelligent prediction, dynamic regulation and control and multi-system cooperation are fused, and accurate control over ore body sinking is achieved.

Description

Technical Field

[0001] This invention relates to a method and system for sinking underground mining bodies, belonging to the field of underground mining engineering technology. Background Technology

[0002] During underground ore body mining, ore body subsidence is a key factor leading to surface collapse, tunnel deformation, and even safety accidents. Traditional mining operations primarily rely on empirical formulas or static monitoring to control ore body subsidence, which has the following technical limitations: Low prediction accuracy: Traditional models such as linear regression or ARIMA are difficult to capture the nonlinear time-series relationship between mining parameters, geological conditions and subsidence, and the prediction error often exceeds 10 mm / day. Delayed control: Relying on manual analysis of monitoring data to adjust the mining progress or filling parameters, the response time from the discovery of anomalies to the implementation of control measures exceeds 24 hours, which may cause the best control opportunity to be missed. Poor adaptability: When geological conditions (such as joint density and ground stress) change, the model cannot dynamically adapt, resulting in a control failure probability of up to 30% in high-risk areas (such as subsidence ≥ 500 mm). Insufficient coordination: The mining, backfilling and monitoring systems operate independently, lacking a "prediction-decision-execution" closed loop, resulting in unreasonable mortar ratios or a mismatch between the mining rate and the subsidence trend, increasing costs by 15-20%.

[0003] For example, Chinese Patent Publication No. CN103758519A discloses a mining design and implementation method for staged subsequent backfilling in large and thick mineral deposits. This method selects the stope structure type and optimizes parameters sequentially from top to bottom based on the current stage's ore and rock conditions. It employs orthogonal numerical experiments and nonlinear regression fitting, using the amount of cementitious material used for backfilling in the staged subsequent backfilling mining panel as the optimization objective, and using the stope stability safety factor, surface rock movement parameters, and ore recovery rate as constraints. A staged subsequent backfilling stope design optimization model is established and solved to obtain the current stage's subsequent backfilling mining design parameters. This method can adapt to the differences in ore and rock stability at different stages and the impact of previously mined stages on the stability of the current stage. The selected stope structure type is adapted to the stage's ore and rock stability conditions, and the obtained stage stope design parameters and cemented backfill strength are coordinated and matched with stope stability, rock movement control, mining efficiency, and environmental protection. However, it still uses a traditional model and suffers from the aforementioned technical shortcomings.

[0004] Therefore, there is an urgent need for a technical solution that integrates intelligent prediction, dynamic regulation and multi-system collaboration to achieve precise control of ore body subsidence. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for underground mining body subsidence construction, which integrates intelligent prediction, dynamic regulation and multi-system collaboration to achieve precise control of ore body subsidence.

[0006] The method for sinking construction of an underground mining body according to the present invention includes: A multi-dimensional feature system was constructed, and feature parameters of mining parameters, geological conditions and monitoring data were collected to form a time-series sample set. Train a sinking prediction model based on bidirectional LSTM and attention mechanism, input time series features, and output the predicted value of the average daily sinking amount in the future; Based on the prediction results, implement tiered control: when the predicted average daily subsidence is greater than or equal to the maximum subsidence, start infilling and reduce the mining rate; when it is between the maximum and minimum subsidence, maintain the current parameters and increase the monitoring frequency; when it is less than the minimum subsidence, mine at the normal pace; the maximum subsidence is preferably 10mm, but can also be 8mm, 14mm, 20mm, etc. depending on the situation; the minimum subsidence is preferably 5mm, but can also be 3mm, 7mm, 9mm, etc. depending on the situation; when the predicted average daily subsidence is greater than or equal to the maximum subsidence, start infilling and reduce the mining rate to preferably 0.8m / day, but can also be 0.5m / day, 0.7m / day, 1m / day, etc. depending on the situation.

[0007] An incremental learning mechanism is established to dynamically freeze the parameters of the underlying network and fine-tune the parameters of the top layer, achieving dynamic adaptation. Ideally, the underlying network parameters are frozen and the top layer is fine-tuned every 100 new samples or when geological conditions change beyond a threshold (KS test p < 0.05). Alternatively, depending on the specific circumstances, the parameters can be frozen and the top layer fine-tuned every 200 new samples or when geological conditions change beyond a threshold (KS test p < 0.1), every 50 new samples or when geological conditions change beyond a threshold (KS test p < 0.03), or every 500 new samples or when geological conditions change beyond a threshold (KS test p < 0.2).

[0008] By integrating the characteristics of mining parameters, geological conditions, and monitoring data, a time-series sample set is constructed to achieve quantitative characterization of subsidence driving factors; the bidirectional LSTM + attention mechanism model can capture long-term time-series dependencies, with small prediction error for daily average subsidence; the three-level control strategy forms a closed-loop control of a multi-dimensional feature system and dynamic control methods, which reduces the exposed area through intensified filling, improves the timeliness of early warning through dynamic monitoring, and optimizes construction efficiency through conventional mining.

[0009] Preferably, the mining parameters, geological conditions, and monitoring data characteristic parameters are as follows: Mining parameters: daily advance, amount of explosive per shot, exposed area of ​​goaf, filling rate, and mortar ratio; Geological parameters: ore body compressive strength, surrounding rock integrity coefficient, joint density, in-situ stress value, dip angle, and burial depth; Monitoring data: cumulative subsidence over the previous 7 days, daily subsidence rate, strain of the filling material, frequency of micro-vibrations, and pore water pressure.

[0010] It covers 5 types of mining parameters (daily footage / explosive charge / exposed area, etc.) + 5 types of geological parameters (Kv / joint density, etc.) + 5 types of monitoring indicators (7-day cumulative subsidence, etc.); the parameter combination covers the three major elements of ore body, surrounding rock and process, with a feature dimension of up to 15 dimensions, solving the one-sidedness problem of traditional single parameter prediction.

[0011] Preferably, the time-series sample set is generated through the following steps: Outlier handling: Outliers are identified using the extended IQR method and corrected by replacing them with the nearest mean; Normalization: Non-dimensional features are normalized to [0,1] using min-max, and dimensional features are normalized using Z-score. Time series construction: Generate 7-day sliding window features, including mean, variance, and peak statistics.

[0012] The extended IQR method has high anomaly detection accuracy, and the replacement of the nearest mean avoids data gaps; the hybrid normalization strategy (min-max+Z-score) adapts to dimensionless / dimensional features and improves the model convergence speed; the 7-day sliding window constructs third-order statistics including mean / variance / peak, which enhances the expressive power of time series patterns.

[0013] Preferably, the sinking prediction model based on bidirectional LSTM and attention mechanism includes: Input layer: 16 features × 7 time steps; 3 layers of Bi-LSTM stacked: first layer with 128 neurons, second layer with 64 neurons, and third layer with 32 neurons; Attention mechanism layer: The output of the third layer is weighted by a 32×1 weight matrix to highlight the features of key time steps; Fully connected layer: 16 → 8 nodes, LeakyReLU activated; Output layer: 1 node, linear activation, outputs the average daily sinking amount.

[0014] A 3-layer Bi-LSTM stacked structure (128-64-32) enables layer-by-layer abstraction of spatiotemporal features and backpropagation gradient stabilization; attention mechanism weight allocation increases the contribution of key time steps and improves the ability to predict mutation events; LeakyReLU activation avoids neuron death, and the linear output layer ensures the consistency of the physical meaning of the sinking amount.

[0015] Preferably, the hierarchical regulation further includes: Prediction confidence interval calculation: Estimate the 95% confidence interval based on the normal distribution, and output "predicted value ± upper limit of error"; Bias correction: The systematic error is fitted using linear regression, and the correction formula is as follows: Correction value = Predicted value + Geological complexity coefficient × Deviation coefficient.

[0016] The 95% confidence interval provides an error boundary of ±8mm, reducing the probability of over-adjustment; the geological complexity coefficient (0.5-1.2) quantifies the stability of the surrounding rock, correcting the deviation value to ±12mm, which is suitable for complex geological scenarios.

[0017] Preferably, the geological complexity coefficient is dynamically adjusted according to the following rules: When the surrounding rock integrity coefficient Kv ≥ 0.8, the coefficient = 0.5; When 0.5 ≤ Kv < 0.8, the coefficient = 0.8; When Kv < 0.5 or joint density ≥ 5 joints / m, the coefficient = 1.2.

[0018] When Kv≥0.8, a conservative correction is made (coefficient 0.5); when Kv<0.5 or joint density≥5 joints / m, an enhanced correction is made (coefficient 1.2); a graded response from stable rock mass to fractured rock mass is achieved, and the amount of filling material is optimized.

[0019] The present invention provides a system for implementing the above-described method for sinking underground mining bodies, comprising: Multi-source data acquisition unit: used for real-time acquisition of mining parameters, geological conditions, and monitoring data characteristic parameters; Feature engineering terminal: Deployed at downhole edge nodes to realize outlier correction, normalization and sliding window feature generation; Intelligent prediction module: includes a pre-trained Bi-LSTM model library and an incremental learning engine; Control and execution unit: In conjunction with the filling system and mining equipment, it automatically adjusts the mortar ratio and mining rate based on prediction results; Visualization platform: Displays forecast curves, confidence intervals, and control instructions in real time.

[0020] Edge node feature engineering reduces the central computing load and minimizes data preprocessing latency; integrated prediction-control-execution linkage shortens the system response time; and the visualization platform enhances decision-making transparency and improves the efficiency of key parameter monitoring.

[0021] Preferably, the intelligent prediction module includes: Model deployment unit: TensorRT quantization acceleration is used, improving inference speed by 3 times at INT8 precision; Incremental learning engine: Freeze the parameters of the first two LSTM layers and fine-tune only the fully connected layers; Model evaluation unit: calculates prediction error in real time and automatically triggers retraining when the error exceeds the threshold.

[0022] TensorRT quantization enables INT8 inference speed to reach 120fps (compared to 40fps for the original FP32), meeting the requirements for real-time prediction; incremental learning freezes the first two LSTM layers and only fine-tunes the fully connected layers, completing model updates within 30 minutes, improving data drift adaptation speed by 4 times.

[0023] Preferably, the weights of the attention mechanism layer are calculated as follows: A linear transformation is performed on the 32-dimensional feature vector output by the third LSTM layer to obtain the attention score; The score is normalized using the softmax function, and the sum of the weights is 1. The weight of key time steps is ≥30%.

[0024] The linear transformation plus softmax weight calculation method ensures the enhancement of features at key time steps (such as the day of sudden change in microseismic frequency); the attention mechanism improves the model's F1-score in sudden change scenarios.

[0025] Preferably, the model evaluation unit further includes: Distribution difference detection: The KS test is used to compare the distribution of the new sample with that of the training set. Incremental learning is triggered when p < 0.05. Robustness test: Periodically inject 10% outlier data, and the model is considered qualified when the accuracy drops by ≤10%.

[0026] The KS test (p<0.05) automatically identifies changes in data distribution, triggering incremental learning with high accuracy; 10% outlier data injection testing ensures model robustness and low failure rate in industrial scenarios.

[0027] Compared with existing technologies, the beneficial effects of the method and system for underground mining body sinking construction of the present invention are as follows: Through innovation across the entire chain of "data-model-control-optimization", we achieve high accuracy in predicting daily average subsidence, save on filling material usage, reduce construction safety accident rate, and improve overall efficiency; by integrating intelligent prediction, dynamic control, and multi-system collaboration, we achieve precise control of ore body subsidence. Attached Figure Description

[0028] Figure 1 This is a structural block diagram of a system for underground mining body sinking construction according to the present invention; Figure 2 This is a flowchart of a method for sinking underground mining bodies according to the present invention. Detailed Implementation

[0029] The technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0030] Example 1 like Figure 1 As shown in the figure, this embodiment discloses a system for the sinking construction of underground mining bodies, including a multi-source data acquisition unit, a feature engineering terminal, an intelligent prediction module, a control and execution unit, and a visualization platform.

[0031] The multi-source data acquisition unit performs the following tasks: Surface monitoring: GNSS monitoring stations (accuracy ±2mm, sampling frequency 1Hz), deployed at 200m intervals; Downhole monitoring: Fiber optic strain gauge (measuring range ±3000με, resolution 0.1με), microseismic monitor (16 channels, sampling rate 1MHz, positioning error ≤5m), pore water pressure sensor (range 0-5MPa, accuracy ±0.02MPa). Mining parameter acquisition: The drilling rig has a built-in encoder (records daily advance, accuracy ±0.01m), a blasting recorder (single-shot charge error ≤5%), and a filling pipeline flow meter (accuracy ±0.5%).

[0032] Specifically, five GNSS stations are deployed on the surface, and one fiber optic strain gauge is installed every 50m underground, for a total of 20 microseismic monitoring instruments.

[0033] The feature engineering terminal uses an industrial-grade server (CPU 2.5GHz, memory 32GB), which is deployed in the underground chamber. It mainly performs the following functions: real-time outlier correction, normalization and sliding window feature generation, with a processing latency of ≤1 second, and supports parallel processing of 1000+ monitoring points.

[0034] Specifically, the feature engineering terminal is deployed in the -350m middle section chamber to process data from 1200 monitoring points in real time.

[0035] The intelligent prediction module is as follows: Model deployment: TensorRT quantization acceleration is adopted, which improves inference speed by 3 times at INT8 accuracy, and the single-round prediction time is ≤0.5 seconds; Incremental learning engine: Automatically detects changes in sample distribution and triggers incremental training (time ≤ 30 minutes); the model library covers 5 typical geological conditions. Evaluation unit: RMSE / MAE is calculated in real time, and retraining is automatically triggered when RMSE > 5mm.

[0036] The control and execution unit has the following functions: The filling system is linked: the proportion of binder is adjusted by an electromagnetic flowmeter, with a response delay of ≤500ms and a binder-mortar ratio adjustment accuracy of ±0.1. Integrated operation of mining equipment: Sending advance commands to the drilling rig, with a rate control accuracy of ±0.1m / day; Emergency Mechanism: When the predicted cumulative subsidence over 7 days is ≥50mm, the mining power supply will be automatically cut off and emergency backfilling (quick-setting material, initial setting time ≤30 minutes) will be initiated.

[0037] The visualization platform has the following functions: 3D display: Real-time rendering of ore body subsidence cloud map (updated once every 2 hours), supporting immersive viewing on VR devices; Data traceability: Records the comparison between predicted curves and actual values ​​(error labeling accuracy ±0.1mm), control commands, and parameter adjustment history; Early warning function: Level 3 early warning audio and visual prompts (sound pressure level ≥ 110dB), which are simultaneously pushed to the management personnel terminal.

[0038] After three months of operation, the maximum subsidence of the ore body was controlled at 35mm, a 48% decrease compared to the same period last year (68mm), and the consumption of cementitious materials was reduced by 18%.

[0039] Example 2 like Figure 2 As shown in this embodiment, a method for sinking an underground mining body includes: A multi-dimensional feature system was constructed, and 16 types of feature parameters were collected, including mining parameters, geological conditions and monitoring data, to form a time-series sample set.

[0040] The mining parameters, geological conditions, and monitoring data are as follows: Mining parameters: daily advance (m / day), single-shot charge (kg), exposed area of ​​goaf (m²), filling rate (m³ / h), mortar ratio; Geological parameters: ore body compressive strength (MPa), surrounding rock integrity coefficient (Kv), joint density (strips / m), in-situ stress value (MPa), dip angle (°), burial depth (m); Monitoring data: cumulative subsidence over the previous 7 days (mm), daily subsidence rate (mm / day), strain of the filling material (με), frequency of micro-vibrations (times / day), and pore water pressure (MPa).

[0041] The time series sample set is generated through the following steps: Outlier handling: Outliers are identified using the extended IQR method and corrected by replacing them with the nearest mean; Normalization: Non-dimensional features are normalized to [0,1] using min-max, and dimensional features are normalized using Z-score. Time series construction: Generate 7-day sliding window features, including mean, variance, and peak statistics.

[0042] Train a sinking prediction model based on bidirectional LSTM and attention mechanism, input the time series features of the previous 7 days, and output the predicted value of the average daily sinking amount for the next 7 days.

[0043] Sinking prediction models based on bidirectional LSTM and attention mechanisms include: Input layer: 16 features × 7 time steps; Three layers of Bi-LSTM stacked: the first layer has 128 neurons, the second layer has 64 neurons, and the third layer has 32 neurons (no sequence return). Attention mechanism layer: The output of the third layer is weighted by a 32×1 weight matrix to highlight the features of key time steps; Fully connected layer: 16→8 nodes, LeakyReLU activation (α=0.01); Output layer: 1 node, linear activation, outputs the average daily sinking amount (mm / day).

[0044] Based on the model prediction results, implement graded control: when the predicted daily average subsidence is ≥10mm, start densified backfilling (increase the mortar ratio to 1:6) and reduce the mining rate to 0.8m / day; when it is 5-10mm, maintain the current parameters and increase the monitoring frequency; when it is <5mm, mine according to the normal schedule.

[0045] Tiered regulation also includes: Prediction confidence interval calculation: Estimate the 95% confidence interval based on the normal distribution, and output "predicted value ± upper limit of error"; Deviation correction: The systematic error is fitted by linear regression, and the correction formula is: Correction value = Predicted value + Geological complexity coefficient × Deviation coefficient.

[0046] An incremental learning mechanism for the model is established. When 100 new samples are added or the geological conditions change beyond a threshold (KS test p<0.05), the parameters of the bottom layer network are frozen and the top layer is fine-tuned to achieve dynamic adaptation.

[0047] The geological complexity coefficient is dynamically adjusted according to the following rules: When the surrounding rock integrity coefficient Kv ≥ 0.8, the coefficient = 0.5; When 0.5 ≤ Kv < 0.8, the coefficient = 0.8; When Kv < 0.5 or joint density ≥ 5 joints / m, the coefficient = 1.2.

[0048] The method for underground mining body subsidence construction described in this embodiment is applied to the mining of an iron ore body. This iron ore body has a burial depth of 350m, a dip angle of 45°, a surrounding rock integrity coefficient Kv=0.6, and a joint density of 4 joints / m. The staged open-cut subsequent backfilling method is used for mining. Specifically, the following steps are included: Feature collection: Sixteen parameters were collected according to the method of this invention, including a daily advance of 1.5m / day, a single-shot charge of 300kg, a filling speed of 100m³ / h, and an initial mortar ratio of 1:10. Model training: A bidirectional LSTM model was trained using 3 years of historical data (5000 samples) from this mine. The training set RMSE=2.1mm and the validation set RMSE=2.8mm. Forecasting and Regulation: The first round of forecasts predicts an average daily subsidence of 8.5 mm over the next 7 days (medium risk). The current parameters are maintained, and the monitoring frequency is adjusted to once every 4 hours. The new data on day 3 triggered incremental learning (KS test p=0.03<0.05), and the model prediction was updated to 9.2mm after fine-tuning; On the 5th day, the predicted value rose to 10.3mm (high risk), and the mortar ratio was automatically adjusted to 1:6, reducing the mining rate to 0.8m / day.

[0049] Results verification: The actual daily average subsidence over 7 days was 9.8 mm, with a prediction error of 0.5 mm, which is 88% higher than the traditional method (error 4.2 mm). No subsidence exceeding the threshold occurred.

[0050] The above description is only a preferred embodiment of this practice, but the scope of protection of this embodiment is not limited thereto. Any equivalent substitutions or changes made by those skilled in the art within the scope of the technology disclosed in this embodiment, based on the technical solution and inventive concept of this embodiment, should be covered within the scope of protection of this embodiment.

Claims

1. A method for sinking construction of an underground mining body, characterized in that, include: A multi-dimensional feature system was constructed, and feature parameters of mining parameters, geological conditions and monitoring data were collected to form a time-series sample set. Train a sinking prediction model based on bidirectional LSTM and attention mechanism, input time series features, and output the predicted value of the average daily sinking amount in the future; Based on the prediction results, implement tiered control: when the predicted average daily subsidence is greater than or equal to the maximum subsidence, start intensified backfilling and reduce the mining rate; when it is between the maximum and minimum subsidence, maintain the current parameters and increase the monitoring frequency; when it is less than the minimum subsidence, mine according to the normal schedule. An incremental learning mechanism for the model is established, which dynamically freezes the parameters of the underlying network and fine-tunes the parameters of the top layer to achieve dynamic adaptation.

2. The method for sinking construction of an underground mining body according to claim 1, characterized in that, The mining parameters, geological conditions, and monitoring data characteristics are as follows: Mining parameters: daily advance, amount of explosive per shot, exposed area of ​​goaf, filling rate, and mortar ratio; Geological parameters: ore body compressive strength, surrounding rock integrity coefficient, joint density, in-situ stress value, dip angle, and burial depth; Monitoring data: cumulative subsidence over the previous 7 days, daily subsidence rate, strain of the filling material, frequency of micro-vibrations, and pore water pressure.

3. The method for sinking construction of an underground mining body according to claim 2, characterized in that, The time-series sample set is generated through the following steps: Outlier handling: Outliers are identified using the extended IQR method and corrected by replacing them with the nearest mean; Normalization: Non-dimensional features are normalized to [0,1] using min-max, and dimensional features are normalized using Z-score. Time series construction: Generate 7-day sliding window features, including mean, variance, and peak statistics.

4. The method for sinking construction of an underground mining body according to claim 3, characterized in that, The sinking prediction model based on bidirectional LSTM and attention mechanism includes: Input layer: 16 features × 7 time steps; 3 layers of Bi-LSTM stacked: first layer with 128 neurons, second layer with 64 neurons, and third layer with 32 neurons; Attention mechanism layer: The output of the third layer is weighted by a 32×1 weight matrix to highlight the features of key time steps; Fully connected layer: 16 → 8 nodes, LeakyReLU activated; Output layer: 1 node, linear activation, outputs the average daily sinking amount.

5. The method for sinking construction of an underground mining body according to claim 4, characterized in that, The tiered regulation also includes: Prediction confidence interval calculation: Estimate the 95% confidence interval based on the normal distribution, and output "predicted value ± upper limit of error"; Bias correction: The systematic error is fitted using linear regression, and the correction formula is as follows: Correction value = Predicted value + Geological complexity coefficient × Deviation coefficient.

6. The method for sinking construction of an underground mining body according to claim 5, characterized in that, The geological complexity coefficient is dynamically adjusted according to the following rules: When the surrounding rock integrity coefficient Kv ≥ 0.8, the coefficient = 0.5; When 0.5 ≤ Kv < 0.8, the coefficient = 0.8; When Kv < 0.5 or joint density ≥ 5 joints / m, the coefficient = 1.

2.

7. A system for implementing a method for sinking construction of an underground mining body as described in any one of claims 1-6, characterized in that, include: Multi-source data acquisition unit: used for real-time acquisition of mining parameters, geological conditions, and monitoring data characteristic parameters; Feature engineering terminal: Deployed at downhole edge nodes to realize outlier correction, normalization and sliding window feature generation; Intelligent prediction module: includes a pre-trained Bi-LSTM model library and an incremental learning engine; Control and execution unit: In conjunction with the filling system and mining equipment, it automatically adjusts the mortar ratio and mining rate based on prediction results; Visualization platform: Displays forecast curves, confidence intervals, and control instructions in real time.

8. The system for the underground mining body sinking construction method according to claim 7, characterized in that, The intelligent prediction module includes: Model deployment unit: TensorRT quantization acceleration is used, improving inference speed by 3 times at INT8 precision; Incremental learning engine: Freeze the parameters of the first two LSTM layers and fine-tune the fully connected layers; Model evaluation unit: calculates prediction error in real time and automatically triggers retraining when the error exceeds the threshold.

9. The system for a method of sinking construction of an underground mining body according to claim 7, characterized in that, The weights of the attention mechanism layer are calculated as follows: A linear transformation is performed on the 32-dimensional feature vector output by the third LSTM layer to obtain the attention score; The score is normalized using the softmax function, and the sum of the weights is 1. The weight of key time steps is ≥30%.

10. The system for a method of sinking construction of an underground mining body according to claim 8, characterized in that, The model evaluation unit also includes: Distribution difference detection: The KS test is used to compare the distribution of the new sample with that of the training set. Incremental learning is triggered when p < 0.

05. Robustness test: Periodically inject 10% outlier data, and the model is considered qualified when the accuracy drops by ≤10%.

Images