Coal mining deformation real-time monitoring system and use method

Abstract

The invention discloses a coal mining deformation real-time monitoring system which comprises a physical layer, a transmission layer, a processing layer and an application layer. The use method of the system comprises the following steps: laying the main optical fiber and the branch optical fibers, connecting the data processing unit and the optical fiber sensing head and acquiring strain data, performing denoising, normalization and sequence construction on the data, predicting optical fiber strain time sequence data, performing discriminant analysis on a predicted value and an actual value residual error, and performing graded early warning. And triggering the three-dimensional visualization platform to update the abnormal label, querying the historical deformation record, and drilling the emergency response plan. The system has the characteristics of high continuous monitoring capability, high real-time performance, high deployment flexibility and good visualization effect, and the safety and the production efficiency of a coal mine are remarkably improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of coal mine safety monitoring, relates to a real-time monitoring system for coal mine mining deformation, and also relates to a method for using the real-time monitoring system for coal mine mining deformation. Background Art

[0002] Coal mining is a high-risk industrial activity, often accompanied by geological deformation phenomena such as surface subsidence, rock displacement, and tunnel deformation. These deformations can easily lead to structural instability in mines, causing accidents such as landslides and gas explosions, seriously threatening the lives of miners and the integrity of production equipment. Traditional deformation monitoring methods use leveling, GPS positioning, and inclinometers, but these methods have many drawbacks. For example, the monitoring coverage is insufficient, relying on discrete monitoring points with large spacing between points, and the underreporting rate of deformation at the edge of the goaf is high. For example, when the dust concentration is greater than 50mg / m³, the measurement error of the traditional laser rangefinder increases to ±5mm (data source: "Technical Specifications for Mine Surveying" GB / T51256-2017); the real-time performance is insufficient, data acquisition and processing take a long time, and the traditional threshold alarm system has a long response delay to slow deformation. For example, the response delay to roof subsidence of 0.1mm / hour exceeds 6 hours, which is difficult to meet the real-time warning needs; the environmental adaptability is poor, when the humidity is greater than 90%, the risk of circuit board corrosion increases by 47% (according to the 2022 accident analysis report of the State Administration of Mine Safety Supervision). When the methane concentration is greater than 1%, the probability of false triggering of electromagnetic sensors reaches 8.3%; there are safety hazards. Some traditional equipment requires electric drive and is prone to explosions due to sparks in high-gas environments, which does not meet the explosion-proof requirements of coal mines.

[0003] While distributed fiber-optic sensing technology has been successfully applied in recent years in fields like bridge monitoring and pipeline inspection, owing to its advantages of high precision, high sensitivity, and resistance to electromagnetic interference, its application in coal mining faces challenges. For example, the multipath effect in mines can increase optical signal attenuation by 3-5 dB / km; fiber microbend losses caused by mining stresses interfere with signal demodulation; and existing algorithms fail to fully account for the coupled effects of gas adsorption and desorption on rock deformation. Therefore, developing a deformation monitoring system optimized for the complex working conditions of coal mines is of great significance, both technically valuable and practically relevant. Summary of the Invention

[0004] The purpose of the present invention is to provide a real-time monitoring system for coal mine deformation, which solves the problems of poor monitoring space continuity, weak real-time performance, low anti-interference ability and delayed early warning mechanism in the existing technology. The present invention also provides a method for using the real-time monitoring system for coal mine deformation.

[0005] The first technical solution adopted by the present invention is a real-time monitoring system for coal mining deformation, comprising a physical layer configured with a tree-topology optical fiber network and a composite encapsulated optical fiber for constructing a physical carrier for signal transmission; The transmission layer is equipped with an optical signal demodulation unit and a data transmission unit to realize optical signal conversion and data transmission; The processing layer is equipped with a data pre-processing module and a core algorithm module to receive, store and analyze the transmitted data; The application layer includes an ASIL-D warning terminal and a 3D visualization platform, which are aimed at actual scenario applications. The ASIL-D warning terminal is used to meet high-safety-level warning requirements, and the 3D visualization platform is used for data visualization.

[0006] The characteristics of the present invention are: The tree-like topology fiber optic network includes a trunk fiber and several branch fibers. The trunk fiber is laid axially along the main tunnel of the coal mine. The pre-tension of the trunk fiber is 0.1%-0.3% during the laying process. The branch optical fibers and the trunk optical fibers form a tree-like topology. Several branch optical fibers are connected to the goaf boundary, the front edge of the working face, the fault and the turning point of the tunnel. Several branch optical fibers extend into the goaf boundary at an inclination angle of 15°±2° relative to the main tunnel direction. The length of several branch optical fibers to the trunk optical fiber is 1:(0.35-0.48).

[0007] The composite encapsulated optical fiber is coated on the outside of the trunk optical fiber and the branch optical fiber. The composite encapsulated optical fiber is composed of a 316L stainless steel corrugated protective tube and a polytetrafluoroethylene-ceramic double-layer coating. The polytetrafluoroethylene-ceramic double-layer coating is arranged on both the inside and outside of the 316L stainless steel corrugated protective tube. The dielectric strength of the polytetrafluoroethylene-ceramic double-layer coating is not less than 30kV / mm.

[0008] The signal attenuation rate of the composite encapsulated optical fiber is less than 0.5dB / km under the working conditions of ambient temperature of 60~80℃ and relative humidity of 90~100%.

[0009] The data preprocessing module includes a long short-term memory network (LSTM) prediction model and a random forest (RF) anomaly detection algorithm. The LSTM prediction model consists of an input layer and a hidden layer. The input layer includes 12-dimensional real-time feature parameters. The input layer predicts the strain trend value within the next 30 to 60 minutes, which is used to provide early warning of impending slow deformation. The hidden layer contains two layers of LSTM units and a Dropout regularization layer.

[0010] The 12-dimensional real-time characteristic parameters include current strain value, strain change rate, historical strain sequence of the past 12 hours, corresponding node temperature, corresponding node humidity, surrounding support structure status mark, day / night cycle characteristics, shift cycle characteristics, blasting disturbance log, vibration disturbance log, coal seam thickness, and geological fault distance.

[0011] The hidden layer of the LSTM prediction model contains 128 neurons. Specifically, the loss function constructed by combining MAE and Smooth L1 is used for model training. The loss function expression is: (1) Where, is the weighting coefficient, The value of is 0.3–0.7, MAE is the mean absolute error, and Smooth L1 is the smooth L1 loss function.

[0012] The formula for calculating the mean absolute error is: (2) Where, represents the true value, represents the predicted value, represents the number of samples; The calculation formula of the smooth L1 loss function is: (3).

[0013] The random forest (RF) anomaly detection algorithm sets a dual-threshold recognition mechanism to identify anomalies in the residual between the prediction results of the long short-term memory (LSTM) prediction model and the real-time strain value, output the anomaly level and recommended warning level, and trigger the 3D visualization platform to update the anomaly annotation.

[0014] Another technical solution adopted by the present invention is a method for using a real-time monitoring system for coal mining deformation, which is characterized by comprising: Step 1: laying trunk optical fibers and branch optical fibers axially along the main tunnel of the coal mine to form a tree-like topology; Step 2: Connect the data processing unit to the optical fiber sensor head, start the terminal fault self-diagnosis system, and connect to the central monitoring platform; Step 3: The optical fiber sensing node collects strain data, synchronously records 12-dimensional feature parameters, and uploads them after edge processing for denoising, normalization, and sequence construction. Step 4, predicting the optical fiber strain time series data to obtain a predicted value; Step 5: Perform discriminant analysis on the residuals between the predicted value and the actual value and issue graded warnings, triggering the 3D visualization platform to update the abnormality annotations; Step 6: Start the mine digital twin model built with the Unity 3D engine and load real-time strain data and warning results; Step 7: Query historical deformation records and practice emergency response plans.

[0015] The beneficial effects of the present invention are: The real-time monitoring system for coal mine deformation of the present invention realizes all-weather, full-space monitoring and early warning of the deformation status of key areas of coal mines by constructing a high-density, high-sensitivity, and highly environmentally adaptable distributed optical fiber network and combining it with advanced data processing and three-dimensional display technology. It can effectively improve the safety production level of mines, reduce the incidence of safety accidents, and has good engineering applicability and promotion value. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 This is a module distribution diagram of the real-time monitoring system for coal mining deformation of the present invention; Figure 2 This is an operation flow chart of the real-time monitoring system for coal mining deformation according to the present invention. DETAILED DESCRIPTION

[0017] The present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0018] Coal mining deformation real-time monitoring system, such as Figure 1 As shown, it includes a physical layer, which is equipped with a tree-topology optical fiber network and composite packaged optical fiber to build a physical carrier for signal transmission; a transmission layer, which has an optical signal demodulation unit and a data transmission unit to realize optical signal conversion and data transmission; a processing layer, which has a data preprocessing module and a core algorithm module to receive, store and analyze the transmitted data; an application layer, which includes an ASIL-D level early warning terminal and a three-dimensional visualization platform, and is oriented to actual scenario applications. The ASIL-D level early warning terminal is used to meet high-security level early warning requirements, and the three-dimensional visualization platform is used for data visualization.

[0019] The tree-like topology fiber optic network includes a trunk fiber and several branch fibers. The trunk fiber is laid axially along the main tunnel of the coal mine and is used for the main channel transmission of deformation signals. The trunk fiber is applied with a pre-tension of 0.1%-0.3% during the laying process to ensure its mechanical stability and deformation response sensitivity; the branch fibers and the trunk fiber form a tree-like topology structure to achieve multi-area distributed layout. Several branch fibers are connected to the goaf boundary, the front edge of the working face, the fault and the tunnel inflection point. Several branch fibers extend into the goaf boundary at an inclination angle of 15°±2° relative to the main tunnel direction. The length of several branch fibers to the trunk fiber is 1:(0.35-0.48) to optimize the balance between coverage range and signal attenuation performance.

[0020] Composite encapsulated optical fiber is coated on the outside of the trunk optical fiber and branch optical fiber to achieve highly corrosion-resistant and high-insulation optical signal transmission. The composite encapsulated optical fiber is composed of a 316L stainless steel corrugated protective tube and a polytetrafluoroethylene-ceramic double-layer coating. The polytetrafluoroethylene-ceramic double-layer coating is arranged on both the inside and outside of the 316L stainless steel corrugated protective tube. The dielectric strength of the polytetrafluoroethylene-ceramic double-layer coating is not less than 30kV / mm to improve the stability and safety of the system in humid and high-gas environments, and to improve the corrosion resistance and signal transmission stability of the optical fiber in extreme mining environments.

[0021] Under working conditions of ambient temperature of 60~80℃ and relative humidity of 90~100%, the signal attenuation rate of the composite encapsulated optical fiber is less than 0.5dB / km, ensuring its long-term stable transmission performance in extremely high temperature and high humidity underground environments.

[0022] The data preprocessing module includes a long short-term memory network (LSTM) prediction model and a random forest (RF) anomaly detection algorithm. The LSTM prediction model consists of an input layer and a hidden layer. The input layer includes 12-dimensional real-time feature parameters. The input layer predicts the strain trend value within the next 30 to 60 minutes, which is used to provide early warning of impending slow deformation. The hidden layer contains two layers of LSTM units and a Dropout regularization layer.

[0023] The 12-dimensional real-time characteristic parameters include current strain value, strain change rate, historical strain sequence of the past 12 hours, corresponding node temperature, corresponding node humidity, surrounding support structure status mark, day / night cycle characteristics, shift cycle characteristics, blasting disturbance log, vibration disturbance log, coal seam thickness, and geological fault distance.

[0024] The hidden layer of the LSTM prediction model contains 128 neurons. Specifically, the loss function constructed by combining MAE and Smooth L1 is used for model training. The loss function expression is: (1) Where, is the weighting coefficient, The value of is 0.3–0.7, MAE is the mean absolute error, and Smooth L1 is the smooth L1 loss function.

[0025] The mean absolute error is a commonly used indicator to measure the deviation between the model's predicted value and the true value. It is calculated as the average of the absolute values of the difference between the predicted value and the true value. MAE has good interpretability and is insensitive to outliers. It is suitable for trend assessment of medium-amplitude deformation variables in mining areas. The calculation formula for the mean absolute error is: (2) Where, represents the true value, represents the predicted value, represents the number of samples; The Smooth L1 loss function combines the advantages of the Mean Squared Error (MSE) and Mean Absolute Error (MAE). It is often used in prediction scenarios that are sensitive to small errors but robust to large errors. This function has smaller gradient fluctuations when the error is small, which helps the model learn subtle changes. However, when the error is large, it degenerates to the MAE, avoiding excessive penalty for outliers. In this system, Smooth L1 effectively improves the fitting accuracy for small deformations of 0.1–1 mm. The calculation formula for the Smooth L1 loss function is: (3).

[0026] This paper systematically improves the Long Short-Term Memory (LSTM) prediction model and the Random Forest (RF) anomaly detection algorithm. First, a 12-dimensional fusion feature, including strain value, rate of change, temperature and humidity, support status, and historical trends, is introduced into the LSTM prediction model, enhancing the model's ability to recognize deformation patterns under complex coal mine conditions. Simultaneously, a dual-scale prediction mechanism, combining short-term and medium-term predictions, improves the ability to simultaneously respond to sudden deformation and slow settlement. Furthermore, a combined MAE and Smooth L1 loss function is introduced for model training, significantly improving the sensitivity of detecting small, gradual deformations (<0.3 mm / h). For the Random Forest (RF) anomaly detection algorithm, a recursive feature elimination method is used to select the optimal feature combination, and the tree depth and sample subset distribution are optimized to enhance the stability and robustness of anomaly detection. Furthermore, the system incorporates a dual-threshold graded warning mechanism to achieve differentiated responses to slow and sudden changes. It also integrates with a three-dimensional visualization platform to enable dynamic visualization of deformation hotspots. Finally, the model integrates empirical rules to effectively improve the recognition accuracy of high-risk areas such as goaf boundaries and fault zones, and realizes the efficient combination of data-driven and knowledge-driven approaches.

[0027] The random forest RF anomaly detection algorithm is used to identify anomalies in the residual between the prediction results of the long short-term memory network LSTM prediction model and the real-time strain value. It includes three parts: input features, model construction and judgment mechanism. Among them, the input features are 8 key features screened out from the original feature set through the recursive feature elimination (RFE) method, including current strain value, residual between predicted value and actual value, abnormal residual frequency, ambient temperature change rate, ambient humidity change rate, working surface activity status, node position code, and strain gradient within the previous 1-2 hours; the model is constructed by constructing a random forest model containing 100 decision trees, each tree is trained with different feature subsets and sample subsets, and finally the abnormality type is judged by voting, including normal, slow change and mutation; the judgment mechanism is a dual-threshold recognition mechanism, first-level abnormality: predicted residual >1.5mm and lasting ≥30 minutes, second-level abnormality: mutation rate >0.5mm / min or surrounding node resonance abnormality >3; it is used to identify abnormalities of the residual between the prediction result of the long short-term memory network LSTM prediction model and the real-time strain value, output the abnormality level and recommended warning level, and trigger the 3D visualization platform to update the abnormality label.

[0028] ASIL-D warning terminals are used for real-time display of deformation status in mining areas, risk warning, and coordinated control. They support visualization based on a three-dimensional visualization platform and have fault self-diagnosis capabilities. They can detect system operating status and provide prompts and processing for abnormal conditions to meet the requirements for stable operation in high-safety scenarios. The ASIL-D warning terminal is equipped with a fault self-diagnosis system that performs system health status checks every 30 to 60 seconds. Monitoring indicators include memory usage, CPU temperature, communication link integrity, and data transmission latency. When the detection value exceeds the set threshold, a fault report is automatically generated and uploaded to the central monitoring platform. The three-dimensional visualization platform, based on the Unity 3D engine, builds a digital twin model of the mine, supports deformation thermal map display with an accuracy of 0.1mm, and supports at least five years of historical data backtracking and emergency plan simulation.

[0029] like Figure 2 As shown in the figure, the implementation of a real-time monitoring system for coal mining deformation includes monitoring scheme design, on-site deployment and installation, data collection and synchronous transmission, model training and trend prediction, and early warning and response linkage. Before using the real-time monitoring system for coal mining deformation, a monitoring scheme design is required. This design requires selecting typical monitoring areas based on the coal mining plan, geological structure, and the distribution of historical disaster areas. A layered deployment is then performed based on stress concentration areas, weak support areas, goaf boundaries, and surface subsidence zones. The system simulates underground stress transmission paths and subsidence trends, and determines the length and path of the sensor cable wiring.

[0030] The method for using the real-time monitoring system for coal mining deformation is based on the real-time monitoring system for coal mining deformation, including: Step 1: A main optical fiber is laid axially along the main roadway of the coal mine, and a pre-tension of 0.1% to 0.3% is applied. Branch optical fibers are laid at an inclination angle of 15°±2° with the main optical fiber toward the goaf, working face, and roadway turning point. The length ratio of the branch optical fiber to the main optical fiber is 1:(0.35-0.48), forming a tree-like topology. Step 2: Connect the data processing unit and the fiber optic sensor head, check communication integrity, start the terminal fault self-diagnosis system and test the memory, CPU temperature, communication link and delay parameters to ensure the healthy operation of the system. Then connect to the central monitoring platform and complete basic configuration and data calibration. Step 3: The fiber optic sensing node collects strain data every 10-15 minutes and simultaneously records 12-dimensional feature parameters. The 12-dimensional feature parameters are processed at the edge and uploaded to the data processing unit for denoising, normalization, and sequence construction. Step 4: Use the long short-term memory network (LSTM) prediction model to predict the optical fiber strain time series data to obtain the predicted value; Step 5: Use the random forest RF anomaly detection algorithm to perform discriminant analysis on the residuals between the predicted value and the actual value, issue a graded warning based on the dual threshold recognition mechanism, and trigger the 3D visualization platform to update the anomaly label; Step 6: Start the mine digital twin model built with the Unity 3D engine, load real-time strain data and warning results, and display a thermal map with 0.1mm accuracy; Step 7: Query historical deformation records. The 3D visualization platform supports data backtracking for no less than 5 years, exercises emergency response plans, and assists on-site decision-making.

[0031] In the later stages, the system needs to be maintained and safety tested, including system health testing, which automatically performs fault self-diagnosis every 30 to 60 seconds. If the detection indicators such as memory usage and communication delay exceed the set threshold, a report will be automatically generated and uploaded to the monitoring center. The system supports online firmware upgrades and self-recovery mechanisms to ensure long-term operational safety. It also includes regular maintenance, monthly physical inspections and signal tests of fiber optic nodes, quarterly model retraining or parameter fine-tuning to adapt to changes in the mining environment, and an annual three-dimensional deformation assessment report for the entire mine, combined with the results of monitoring data interpretation.

[0032] Example 1 Coal mining deformation real-time monitoring system, such as Figure 1As shown, it includes a physical layer, which is equipped with a tree-topology optical fiber network and composite packaged optical fiber to build a physical carrier for signal transmission; a transmission layer, which has an optical signal demodulation unit and a data transmission unit to realize optical signal conversion and data transmission; a processing layer, which has a data preprocessing module and a core algorithm module to receive, store and analyze the transmitted data; an application layer, which includes an ASIL-D level early warning terminal and a three-dimensional visualization platform, and is oriented to actual scenario applications. The ASIL-D level early warning terminal is used to meet high-security level early warning requirements, and the three-dimensional visualization platform is used for data visualization.

[0033] The tree-like topology fiber optic network includes a trunk fiber and several branch fibers. The trunk fiber is laid axially along the main tunnel of the coal mine and is used for the main channel transmission of deformation signals. The trunk fiber is applied with a pre-tension of 0.1%-0.3% during the laying process to ensure its mechanical stability and deformation response sensitivity; the branch fibers and the trunk fiber form a tree-like topology structure to achieve multi-area distributed layout. Several branch fibers are connected to the goaf boundary, the front edge of the working face, the fault and the tunnel inflection point. Several branch fibers extend into the goaf boundary at an inclination angle of 15°±2° relative to the main tunnel direction. The length of several branch fibers to the trunk fiber is 1:(0.35-0.48) to optimize the balance between coverage range and signal attenuation performance.

[0034] Composite encapsulated optical fiber is coated on the outside of the trunk optical fiber and branch optical fiber to achieve highly corrosion-resistant and high-insulation optical signal transmission. The composite encapsulated optical fiber is composed of a 316L stainless steel corrugated protective tube and a polytetrafluoroethylene-ceramic double-layer coating. The polytetrafluoroethylene-ceramic double-layer coating is arranged on both the inside and outside of the 316L stainless steel corrugated protective tube. The dielectric strength of the polytetrafluoroethylene-ceramic double-layer coating is not less than 30kV / mm to improve the stability and safety of the system in humid and high-gas environments, and to improve the corrosion resistance and signal transmission stability of the optical fiber in extreme mining environments.

[0035] Under working conditions of ambient temperature of 60~80℃ and relative humidity of 90~100%, the signal attenuation rate of the composite encapsulated optical fiber is less than 0.5dB / km, ensuring its long-term stable transmission performance in extremely high temperature and high humidity underground environments.

[0036] Example 2 Based on the real-time monitoring system for coal mine deformation provided in Example 1, the real-time monitoring system for coal mine deformation provided in this embodiment has a data preprocessing module including a long short-term memory network LSTM prediction model and a random forest RF anomaly detection algorithm. The long short-term memory network LSTM prediction model includes an input layer and a hidden layer. The input layer includes 12-dimensional real-time feature parameters. The input layer predicts the strain trend value within the next 30 to 60 minutes, which is used to provide early warning of impending slow deformation. The hidden layer includes 2 layers of LSTM units and a Dropout regularization layer.

[0037] The 12-dimensional real-time characteristic parameters include current strain value, strain change rate, historical strain sequence of the past 12 hours, corresponding node temperature, corresponding node humidity, surrounding support structure status mark, day / night cycle characteristics, shift cycle characteristics, blasting disturbance log, vibration disturbance log, coal seam thickness, and geological fault distance.

[0038] The hidden layer of the LSTM prediction model contains 128 neurons. Specifically, the loss function constructed by combining MAE and Smooth L1 is used for model training. The loss function expression is: (1) Where, is the weighting coefficient, The value of is 0.3–0.7, MAE is the mean absolute error, and Smooth L1 is the smooth L1 loss function.

[0039] The mean absolute error is a commonly used indicator to measure the deviation between the model's predicted value and the true value. It is calculated as the average of the absolute values of the difference between the predicted value and the true value. MAE has good interpretability and is insensitive to outliers. It is suitable for trend assessment of medium-amplitude deformation variables in mining areas. The calculation formula for the mean absolute error is: (2) Where, represents the true value, represents the predicted value, represents the number of samples; The Smooth L1 loss function combines the advantages of the Mean Squared Error (MSE) and Mean Absolute Error (MAE). It is often used in prediction scenarios that are sensitive to small errors but robust to large errors. This function has smaller gradient fluctuations when the error is small, which helps the model learn subtle changes. However, when the error is large, it degenerates to the MAE, avoiding excessive penalty for outliers. In this system, Smooth L1 effectively improves the fitting accuracy for small deformations of 0.1–1 mm. The calculation formula for the Smooth L1 loss function is: (3).

[0040] This paper systematically improves the Long Short-Term Memory (LSTM) prediction model and the Random Forest (RF) anomaly detection algorithm. First, a 12-dimensional fusion feature, including strain value, rate of change, temperature and humidity, support status, and historical trends, is introduced into the LSTM prediction model, enhancing the model's ability to recognize deformation patterns under complex coal mine conditions. Simultaneously, a dual-scale prediction mechanism, combining short-term and medium-term predictions, improves the ability to simultaneously respond to sudden deformation and slow settlement. Furthermore, a combined MAE and Smooth L1 loss function is introduced for model training, significantly improving the sensitivity of detecting small, gradual deformations (<0.3 mm / h). For the Random Forest (RF) anomaly detection algorithm, a recursive feature elimination method is used to select the optimal feature combination, and the tree depth and sample subset distribution are optimized to enhance the stability and robustness of anomaly detection. Furthermore, the system incorporates a dual-threshold graded warning mechanism to achieve differentiated responses to slow and sudden changes. It also integrates with a three-dimensional visualization platform to enable dynamic visualization of deformation hotspots. Finally, the model integrates empirical rules to effectively improve the recognition accuracy of high-risk areas such as goaf boundaries and fault zones, and realizes the efficient combination of data-driven and knowledge-driven approaches.

[0041] The random forest RF anomaly detection algorithm is used to identify anomalies in the residual between the prediction results of the long short-term memory network LSTM prediction model and the real-time strain value. It includes three parts: input features, model construction and judgment mechanism. Among them, the input features are 8 key features screened out from the original feature set through the recursive feature elimination (RFE) method, including current strain value, residual between predicted value and actual value, abnormal residual frequency, ambient temperature change rate, ambient humidity change rate, working surface activity status, node position code, and strain gradient within the previous 1-2 hours; the model is constructed by constructing a random forest model containing 100 decision trees, each tree is trained with different feature subsets and sample subsets, and finally the abnormality type is judged by voting, including normal, slow change and mutation; the judgment mechanism is a dual-threshold recognition mechanism, first-level abnormality: predicted residual >1.5mm and lasting ≥30 minutes, second-level abnormality: mutation rate >0.5mm / min or surrounding node resonance abnormality >3; it is used to identify abnormalities of the residual between the prediction result of the long short-term memory network LSTM prediction model and the real-time strain value, output the abnormality level and recommended warning level, and trigger the 3D visualization platform to update the abnormality label.

[0042] Example 3 Based on the real-time monitoring system for coal mining deformation provided in Example 2, the real-time monitoring system for coal mining deformation provided in this embodiment, the ASIL-D level early warning terminal is used for real-time display of the deformation state of the mining area, risk early warning and linkage control, the ASIL-D level early warning terminal supports visualization based on a three-dimensional visualization platform, has a fault self-diagnosis function, can detect the system operation status and prompt and process abnormal conditions to meet the stable operation requirements under high safety level scenarios. The ASIL-D level early warning terminal is equipped with a fault self-diagnosis system, which is set to perform a system health status detection every 30 to 60 seconds. The monitoring indicators include memory usage, CPU temperature, communication link integrity and data transmission delay. When the detection value exceeds the set threshold, a fault report is automatically generated and uploaded to the central monitoring platform. Among them, the three-dimensional visualization platform builds a digital twin model of the mine based on the Unity3D engine, supports deformation thermal map display with an accuracy of 0.1mm, and historical data backtracking and emergency plan simulation for no less than 5 years.

[0043] Example 4 like Figure 2 As shown in the figure, the implementation of a real-time monitoring system for coal mining deformation includes monitoring scheme design, on-site deployment and installation, data collection and synchronous transmission, model training and trend prediction, and early warning and response linkage. Before using the real-time monitoring system for coal mining deformation, a monitoring scheme design is required. This design requires selecting typical monitoring areas based on the coal mining plan, geological structure, and the distribution of historical disaster areas. A layered deployment is then performed based on stress concentration areas, weak support areas, goaf boundaries, and surface subsidence zones. The system simulates underground stress transmission paths and subsidence trends, and determines the length and path of the sensor cable wiring.

[0044] On-site layout and installation involve fixing the optical fiber with special clamps at locations such as the tunnel roof, side walls, and surface cracks to ensure tight coupling of the sensor response, and introducing high-temperature resistant sealing boxes and flame-retardant sheaths to prevent the impact of secondary disasters such as underground fires. Among them, the high-temperature resistant sealing box is installed at the connection point between the branch optical fiber and the trunk optical fiber, the demodulation instrument interface, the key turning node, or the humid area prone to water accumulation. It is especially used to protect the bare ends of optical fibers, fusion points, and other locations susceptible to environmental influences. The specific method of laying out the high-temperature resistant sealing box is to pre-dig grooves or hanging points of corresponding sizes according to the installation location, neatly coil the optical fiber connection points, and place them inside the sealing box. The optical fiber connector is fixed with high-temperature silicone or heat shrink tubing, and the sealing box is filled with halogen-free flame-retardant colloids such as epoxy potting glue, and the outer shell is sealed with stainless steel buckles. The outer wall of the sealing box is coated with an anti-corrosion layer, and a numbered label is affixed. It is fixed to the tunnel side wall or roof, and reinforced with metal brackets or expansion bolts. The performance requirements of the high-temperature resistant sealed box are that the sealing level is not less than IP68, the temperature resistance range is -20℃ to 120℃, and the impact resistance and pressure resistance are not less than 10MPa.

[0045] The specific method for laying out the flame-retardant sheath is to lay it out along the entire length of the fiber trunk and branch lines, especially covering high-risk areas such as those crossing high-temperature working surfaces, cable intersections, and areas affected by blasting. A sleeve-type laying method is used: the optical cable is passed through a flexible flame-retardant sheath, such as a glass fiber-coated silicone sheath or a polyurethane composite sheath. Stainless steel clamps are used to secure the sheath to the optical cable every 1.5-2 meters to prevent loosening and slippage. Flame-retardant heat shrink tubing is used to seal both ends of the sheath, and laser identification signs are placed to indicate the cable number and laying direction. Fluorescent reflective identification tape is added to the outer layer of the sheath every 5 meters to facilitate inspection and identification in dim environments. At corners or intersections, flexible corrugated sheaths or metal bracket guides are used to prevent the sheath from stretching or twisting.

[0046] Example 5 Using the coal mining deformation real-time monitoring system provided in Example 3, this embodiment provides a method for using the coal mining deformation real-time monitoring system, based on the coal mining deformation real-time monitoring system, including: Step 1: A main optical fiber is laid axially along the main roadway of the coal mine, and a pre-tension of 0.1% to 0.3% is applied. Branch optical fibers are laid at an inclination angle of 15°±2° with the main optical fiber toward the goaf, working face, and roadway turning point. The length ratio of the branch optical fiber to the main optical fiber is 1:(0.35-0.48), forming a tree-like topology. Step 2: Connect the data processing unit and the fiber optic sensor head, check communication integrity, start the terminal fault self-diagnosis system and test the memory, CPU temperature, communication link and delay parameters to ensure the healthy operation of the system. Then connect to the central monitoring platform and complete basic configuration and data calibration. Step 3: The fiber optic sensing node collects strain data every 10-15 minutes and simultaneously records 12-dimensional feature parameters. The 12-dimensional feature parameters are processed at the edge and uploaded to the data processing unit for denoising, normalization, and sequence construction. Step 4: Use the long short-term memory network (LSTM) prediction model to predict the optical fiber strain time series data to obtain the predicted value; Step 5: Use the random forest RF anomaly detection algorithm to perform discriminant analysis on the residuals between the predicted value and the actual value, issue a graded warning based on the dual threshold recognition mechanism, and trigger the 3D visualization platform to update the anomaly label; Step 6: Start the mine digital twin model built with the Unity 3D engine, load real-time strain data and warning results, and display a thermal map with 0.1mm accuracy; Step 7: Query historical deformation records. The 3D visualization platform supports data backtracking for no less than 5 years, exercises emergency response plans, and assists on-site decision-making.

[0047] In the later stages, the system needs to be maintained and safety tested, including system health testing, which automatically performs fault self-diagnosis every 30 to 60 seconds. If the detection indicators such as memory usage and communication delay exceed the set threshold, a report will be automatically generated and uploaded to the monitoring center. The system supports online firmware upgrades and self-recovery mechanisms to ensure long-term operational safety. It also includes regular maintenance, monthly physical inspections and signal tests of fiber optic nodes, quarterly model retraining or parameter fine-tuning to adapt to changes in the mining environment, and an annual three-dimensional deformation assessment report for the entire mine, combined with the results of monitoring data interpretation.

[0048] Example 6 The method for using the real-time coal mining deformation monitoring system provided in Example 5 employs this method. This example evenly lays three main optical fibers axially within the main tunnel, spaced 2 meters apart. A tension control device applies a 0.2% pre-tension for fixation, ensuring the optical fibers' deformation response sensitivity and stability during long-term operation. Furthermore, branch optical fibers branch off from the main fibers and extend within 50 meters of the coal mining face. Stress monitoring nodes are located every 10 meters, achieving a high-density deployment to cover high-risk areas within the face. A laser interferometer (with a measurement accuracy of λ / 20) is used to initially calibrate the reference segment of the optical fiber sensing system to ensure that sensing accuracy meets the required standards. Furthermore, a temperature-strain compensation database covering the range of -20°C to 150°C is established in a laboratory environment for ambient temperature correction during field operation, improving the reliability of sensing data.

[0049] The real-time monitoring system for coal mining deformation collects 150,000 sets of measured deformation data from 10 coal mines, and manually reviews and labels 5% of suspected abnormal samples to supervise the anomaly identification part of model training. The long short-term memory network (LSTM) prediction model uses 200 iterations of training, with a mixed loss function of MAE and Smooth L1 as the optimization target, and ultimately achieves a prediction accuracy of 98.2% on the validation set, with the corresponding deformation error controlled within the range of ±0.3mm. The random forest (RF) anomaly detection algorithm uses recursive feature elimination to screen out eight key input parameters (including strain slope, historical residuals, temperature and humidity fluctuations, etc.). The model's AUC value for anomaly identification is 0.972, and the 95% confidence interval is 0.961 to 0.983, indicating its extremely high recognition accuracy and robustness.

[0050] Under artificially simulated tunnel collapse conditions, the system triggers an audible and visual alarm upon detecting a deformation of 8mm, with a response time of just 2.3 seconds, meeting the need for rapid response to sudden disasters. When monitoring slow subsidence, the system can detect subtle subsidence trends as small as 0.1mm / h 6.5 hours earlier than traditional GPS monitoring systems, providing more time for disaster prevention and response. After operating for 72 hours in an environment with a high temperature of 70°C and a relative humidity of 95%, the sensor light signal attenuation rate remained below 0.5dB / km, demonstrating excellent system stability. The system has also passed explosion-proof environment testing, including spark ignition tests at 1.5 times the lower explosive limit (LEL) of methane, with no failures or false alarms, meeting the explosion-proof requirements for underground coal mines.

Claims

1. Real-time monitoring system for coal mining deformation, characterized by: It includes a physical layer, which is equipped with a tree-topology optical fiber network and composite encapsulated optical fibers to construct a physical carrier for signal transmission; The transmission layer is equipped with an optical signal demodulation unit and a data transmission unit to realize optical signal conversion and data transmission; The processing layer is equipped with a data pre-processing module and a core algorithm module to receive, store and analyze the transmitted data; The application layer includes an ASIL-D warning terminal and a 3D visualization platform, which are aimed at actual scenario applications. The ASIL-D warning terminal is used to meet high-safety-level warning requirements, and the 3D visualization platform is used for data visualization.

2. The real-time monitoring system for coal mining deformation according to claim 1 is characterized in that: The tree-like topology optical fiber network includes a trunk optical fiber and a plurality of branch optical fibers. The trunk optical fiber is laid axially along the main tunnel of the coal mine. The trunk optical fiber is pre-tensioned at 0.1%-0.3% during the laying process. The branch optical fibers and the trunk optical fibers form a tree-like topology structure, and several of the branch optical fibers are connected to the boundary of the goaf, the front edge of the working surface, the fault and the turning point of the tunnel. Several of the branch optical fibers extend into the boundary of the goaf at an inclination angle of 15°±2° relative to the direction of the main tunnel, and the length of several of the branch optical fibers and the trunk optical fiber is 1:(0.35-0.48).

3. The real-time monitoring system for coal mining deformation according to claim 1, characterized in that: The composite encapsulated optical fiber is coated on the outside of the trunk optical fiber and the branch optical fiber. The composite encapsulated optical fiber is composed of a 316L stainless steel corrugated protective tube and a polytetrafluoroethylene-ceramic double-layer coating. The polytetrafluoroethylene-ceramic double-layer coating is arranged on both the inner and outer sides of the 316L stainless steel corrugated protective tube. The dielectric strength of the polytetrafluoroethylene-ceramic double-layer coating is not less than 30kV / mm.

4. The real-time monitoring system for coal mining deformation according to claim 3, characterized in that: The composite packaged optical fiber has a signal attenuation rate of less than 0.5 dB / km under working conditions of an ambient temperature of 60-80° C. and a relative humidity of 90-100%.

5. The real-time monitoring system for coal mining deformation according to claim 1, characterized in that: The data preprocessing module includes a long short-term memory network (LSTM) prediction model and a random forest (RF) anomaly detection algorithm. The long short-term memory network (LSTM) prediction model includes an input layer and a hidden layer. The input layer includes 12-dimensional real-time feature parameters. The input layer predicts the strain trend value within the next 30 to 60 minutes for early warning of impending slow deformation. The hidden layer includes two layers of LSTM units and a Dropout regularization layer.

6. The real-time monitoring system for coal mining deformation according to claim 5, characterized in that: The 12-dimensional real-time characteristic parameters include current strain value, strain change rate, historical strain sequence of the past 12 hours, corresponding node temperature, corresponding node humidity, surrounding support structure status mark, day / night cycle characteristics, shift cycle characteristics, blasting disturbance log, vibration disturbance log, coal seam thickness, and geological fault distance.

7. The real-time monitoring system for coal mining deformation according to claim 5, characterized in that: The hidden layer of the LSTM prediction model contains 128 neurons. Specifically, the loss function constructed by combining MAE and Smooth L1 is used for model training. The loss function expression is: （1） Where, is the weighting coefficient, The value of is 0.3–0.7, MAE is the mean absolute error, and Smooth L1 is the smooth L1 loss function.

8. The real-time monitoring system for coal mining deformation according to claim 7, characterized in that: The calculation formula of the mean absolute error is: （2） Where, represents the true value, represents the predicted value, represents the number of samples; The calculation formula of the smooth L1 loss function is: （3）。 9. The real-time monitoring system for coal mining deformation according to claim 5, characterized in that: The random forest RF anomaly detection algorithm sets a dual-threshold recognition mechanism to identify anomalies in the residual between the prediction results of the long short-term memory network (LSTM) prediction model and the real-time strain value, output the anomaly level and recommended warning level, and trigger the 3D visualization platform to update the anomaly label.

10. The method for using the real-time monitoring system for coal mining deformation is characterized by: include: Step 1: laying trunk optical fibers and branch optical fibers axially along the main tunnel of the coal mine to form a tree-like topology; Step 2: Connect the data processing unit to the optical fiber sensor head, start the terminal fault self-diagnosis system, and connect to the central monitoring platform; Step 3: The optical fiber sensing node collects strain data, synchronously records 12-dimensional feature parameters, and uploads them after edge processing for denoising, normalization, and sequence construction. Step 4, predicting the optical fiber strain time series data to obtain a predicted value; Step 5: Perform discriminant analysis on the residuals between the predicted value and the actual value and issue graded warnings, triggering the 3D visualization platform to update the abnormality annotations; Step 6: Start the mine digital twin model built with the Unity 3D engine and load real-time strain data and warning results; Step 7: Query historical deformation records and practice emergency response plans.

Images