A tunnel surrounding rock intelligent prediction method and system based on point cloud fusion

Abstract

The application discloses a kind of intelligent prediction method and system of roadway surrounding rock based on point cloud fusion, it is related to underground engineering safety technical field, and prediction method includes: obtaining to multi-parameter monitoring database;Construct point cloud data processing model, establish roadway main shaft identification and path extraction equation and boundary condition correction model and comprehensive deformation evaluation model, the above-mentioned model is analyzed, obtains multi-source boundary constraint condition;Based on double model network architecture and multi-source boundary constraint condition obtains double model constraint loss function, fitting analysis is carried out to double model constraint loss function, and comprehensive constraint loss function is obtained;Conversion memory prediction-pulse collaborative neural network architecture and deformation evolution analysis model are constructed;Comprehensive constraint loss function, conversion memory prediction-pulse collaborative neural network architecture and deformation evolution analysis model are combined multi-parameter monitoring database and are linked modeling analysis, realize the prediction and monitoring of roadway surrounding rock.The application significantly improves early warning accuracy.

Description

Technical Field

[0001] This invention relates to the field of underground engineering safety technology, and in particular to an intelligent prediction method and system for roadway surrounding rock based on point cloud fusion. Background Technology

[0002] Roof collapse and surrounding rock deformation in coal mine roadways are among the major hazards in coal mine safety production. Especially under thick coal seam mining conditions, due to the large thickness of the roof coal, the weakness of the coal seam, and the potential presence of unknown hidden structures, even with scientifically designed anchor bolt and cable support schemes, localized surrounding rock instability and roof collapse accidents can still occur. Traditional methods for monitoring surrounding rock deformation in coal mine roadways mainly employ mechanical monitoring, laser ranging, optical mapping, and photogrammetry. Mechanical monitoring, using instruments such as measuring guns, measuring rods, roof dynamic monitoring instruments, convergence meters, and displacement gauges, can only obtain relative displacements at a limited number of points and cannot measure asymmetric deformation of the roadway. Laser ranging uses lasers to accurately determine the distance to the target, achieving high-precision non-contact measurement; however, its measurement range is small, making regional observation difficult, and its real-time performance is poor. Optical mapping uses equipment such as mine theodolites and total stations to measure roadway deformation, but this method is costly, complex, inefficient, and requires professional surveyors. Photogrammetry uses specialized cameras to photograph the same location in a coal mine roadway and obtains deformation data through image processing. However, it is inferior to some traditional measurement methods in terms of absolute observation accuracy and has stringent environmental requirements, necessitating good lighting conditions. Overall, traditional measurement methods suffer from problems such as scattered monitoring points, limited coverage, poor real-time performance, weak predictive capabilities, and reliance on empirical parameter settings.

[0003] Three-dimensional laser scanning technology, as a non-contact measurement technique capable of rapidly acquiring spatial three-dimensional point cloud data, boasts advantages such as high measurement speed, high accuracy, and large data volume, providing a new technical approach for monitoring surrounding rock in coal mine roadways. However, how to effectively apply three-dimensional laser scanning technology to the monitoring of surrounding rock in thick coal seam roadways, achieving intelligent and automated processes from point cloud data acquisition and processing to roadway deformation prediction, and establishing a scientific system of critical and early warning values ​​for monitoring parameters, remains a pressing technical challenge.

[0004] Traditional statistical methods, such as time series analysis, regression analysis, and grey prediction models, are used in the field of coal mine roadway deformation prediction. These methods are typically based on the statistical characteristics of historical data and are computationally simple, but they struggle to capture nonlinear relationships and complex spatiotemporal features in the data, resulting in limited prediction accuracy. Classic machine learning algorithms, such as Support Vector Machines (SVM), Random Forests (RF), and Decision Trees (DT), can handle multi-dimensional feature inputs, but their generalization ability is insufficient when processing time-series data, especially when dealing with data like roadway surrounding rock, which has significant time-series dependence and abrupt change characteristics, making it difficult to guarantee prediction accuracy. Single neural network models, such as Backpropagation Neural Networks (BP), Recurrent Neural Networks (RNN), and Long Short-Term Memory (LSTM), have certain advantages in time-series prediction, but they are usually based on single data features and lack an effective fusion mechanism for multi-source heterogeneous monitoring data. Furthermore, their simple model structure makes it difficult to simultaneously model long-term trends and identify sudden deformations, and they are prone to gradient vanishing during deep network training, affecting model stability and prediction accuracy.

[0005] Given that existing methods for monitoring deformation of surrounding rock in roadways suffer from problems such as discrete monitoring points, limited coverage, poor real-time performance, and weak predictive capabilities, and that traditional prediction algorithms have shortcomings in multi-parameter fusion, response to abrupt events, and balance between long-term and short-term predictions, resulting in low prediction accuracy and intelligence, this invention proposes an intelligent prediction method and system for surrounding rock in roadways based on point cloud fusion. Summary of the Invention

[0006] This solution addresses the problems and needs raised above by proposing an intelligent prediction method and system for roadway surrounding rock based on point cloud fusion. The method achieves the above-mentioned technical objectives and brings about several other technical benefits by adopting the following technical features.

[0007] One objective of this invention is to propose an intelligent prediction method for roadway surrounding rock based on point cloud fusion, comprising the following steps: S10: Obtain basic data and deformation monitoring data of the surrounding rock of the roadway, and construct a data acquisition and monitoring mechanism to process the basic data and deformation monitoring data to obtain a multi-parameter monitoring database; S20: Based on the multi-parameter monitoring database, a point cloud data processing model is constructed, a roadway main axis identification and path extraction model, a boundary condition correction model, and a comprehensive deformation evaluation model are established. Multi-source data unified boundary conditions are established to analyze the point cloud data processing model, the roadway main axis identification and path extraction equation, the boundary condition correction model, and the comprehensive deformation evaluation model, and obtain multi-source boundary constraint conditions. S30: Design a dual-model network architecture, obtain the dual-model constraint loss function based on the dual-model network architecture and multi-source boundary constraints, perform fitting analysis on the dual-model constraint loss function, and obtain the comprehensive constraint loss function of deformation trend-sudden event index. S40: Based on the initial deformation state encoding vector, deformation history feature vector, and environmental process feature vector, the temporal input feature vector is obtained. Based on the temporal input feature vector, a transformation memory prediction-pulse collaborative neural network architecture and a deformation evolution analysis model are constructed. S50: Combining a multi-parameter monitoring database, prediction results are obtained based on a comprehensive constraint loss function, a transformation memory prediction-pulse collaborative neural network architecture, and a deformation evolution analysis model. The prediction results are then dynamically adjusted through an early warning threshold to perform closed-loop verification and optimization, thereby enabling the prediction and monitoring of instantaneous deformation and long-term trends of the surrounding rock in the roadway.

[0008] In addition, the intelligent prediction method for roadway surrounding rock based on point cloud fusion according to the present invention may also have the following technical features: In one example of the present invention, in step S10, a data acquisition and monitoring model is constructed to process the basic data and deformation monitoring data to obtain a multi-parameter monitoring database, including the following steps: S11: Determine the scope of the calculation domain based on the tunnel model, and set the coordinate system spatiotemporal calculation domain and the boundary conditions of ground stress and deformation of the tunnel according to the scope of the calculation domain. S12: Determine the sampling point density weights based on the spatiotemporal computational domain and the boundary conditions of geostress and deformation; S13: Establish a sampling encryption algorithm and coordinate normalization mapping function for deformation-sensitive areas; S14: The basic data and deformation monitoring data are adjusted by combining the computational domain range, sampling point density weight, deformation-sensitive area sampling densification algorithm and coordinate standardization mapping function to obtain the adjusted basic data and deformation monitoring data. S15: Use a data acquisition and monitoring model to fuse the adjusted basic data and deformation monitoring data to obtain a multi-parameter monitoring database.

[0009] In one example of the present invention, step S20 specifically includes the following steps: S21: Establish a point cloud data processing model based on statistical outlier filtering and curvature-preserving downsampling; S22: Establish equations for identifying the main axis of the roadway and extracting the center path, and construct a boundary condition correction model based on deformation influencing factors; wherein, the equations for identifying the main axis of the roadway and extracting the center path include: the formula for identifying the main axis of the roadway and the formula for extracting the center path; S23: Establish a comprehensive deformation evaluation model based on the experimental information of the surrounding rock deformation in the tunnel and the time-accumulated deformation model; S24: By coupling multi-source data and unifying boundary conditions, the parameter boundary conditions of the point cloud data processing model, the roadway main axis identification and path extraction equation and boundary condition correction model and the comprehensive deformation evaluation model are analyzed to obtain deformation boundary conditions, trend boundary conditions and sudden boundary conditions. S25: Combining deformation boundary conditions, trend boundary conditions, and sudden boundary conditions yields multi-source boundary constraint conditions.

[0010] In one example of the present invention, in step S22, the formula for identifying the main shaft of the roadway is:

[0011] in, The main axis direction vector For the point cloud dataset, the first Each point is a three-dimensional vector. This represents the total number of points in the point cloud. For the centroid of the point cloud, For vectors transpose; The center path is fitted using RANSAC to the roadway centerline. The formula for the distance from a point to a line is:

[0012] in, This represents the straight line that best fits the point cloud data. For the shape of The linear model, Let be the distance from the point to the line. The threshold value is used.

[0013] In one example of the present invention, in step S24, the deformation boundary condition, the trend boundary condition, and the sudden boundary condition are respectively: The deformation boundary conditions satisfy the following relationship:

[0014] In the formula, For the surrounding rock point p during the monitoring time The amount of deformation at that time; To monitor the surrounding rock point p at time t, and the reference outline of the roadway at time t. The minimum distance; Based on the standard outline and the surrounding rock monitoring points Feature matching corresponds to the pairing points; This serves as the standard reference outline for the roadway at the base time. Based on the standard outline and the surrounding rock monitoring points Feature matching corresponding pairing points Standard reference profile of the roadway at the reference time The minimum distance; The trend boundary conditions satisfy the following relationship:

[0015] In the formula, The current cell state is used to control the flow of information; For the forget gate, multiply by To determine how much information is forgotten; This represents the cell state at the previous moment; The output of the input gate determines the update of the cell state at the current moment; Representing the candidate cell state, it is the alternative feature information to be included in the memory unit at the current time step; This is element-wise multiplication; Sudden boundary conditions satisfy the following relationship:

[0016] In the formula, The membrane potential time constant is used to control the potential recovery rate; Current time Membrane potential at time t; This is the resting potential, which is also the potential of a neuron when it is not stimulated. For membrane resistance; For synaptic input current; For adaptive current weighting coefficients; For adaptive current.

[0017] In one example of the present invention, step S30 specifically includes the following steps: S31: Set up a multi-dimensional parameter vector input mechanism and an input parameter standardization processing mechanism in the network architecture; S32: Based on the multi-dimensional parameter vector input mechanism, the input parameter standardization processing mechanism, and the multi-source boundary constraints, the output results of the trend prediction layer, the sudden event detection layer, and the fusion prediction layer are obtained; S33: Construct a dual-model constrained loss function based on the output results of the trend prediction layer, the sudden event detection layer, and the fusion prediction layer; S34: Perform fitting analysis on the dual-model constraint loss function to obtain the comprehensive constraint loss function of deformation trend-emergency index; The comprehensive constraint loss function satisfies the following relationship:

[0018] In the formula, The current parameters are used to measure the overall loss function of the model. The overall prediction error, The mean squared error loss penalizes the squared error between the predicted and actual values. To weight safety risks, higher weights are given to errors in high-risk areas. The weighting coefficients are used to control the balance between ordinary mean square error and risk-weighted loss.

[0019] In one example of the present invention, step S40, obtaining the temporal input feature vector based on the initial deformation state encoding vector, the deformation history feature vector, and the environmental process feature vector, specifically includes the following steps: S41: Based on the network architecture and multi-parameter monitoring database, obtain the spatial deformation distribution characteristics, and obtain the initial deformation state encoding vector according to the spatial deformation distribution characteristics; S42: Construct characteristic expressions for the deformation history and stress process of the surrounding rock, and use the characteristic expressions to obtain the deformation history feature vector; S43: Set up a coding mechanism for environmental conditions and monitoring cycles, and obtain the environmental process feature vector through the coding mechanism; S44: Combine time-varying environmental conditions, initial deformation state encoding vector, deformation history feature vector, and environmental process feature vector to obtain the time-series input feature vector.

[0020] In one example of the present invention, step S40 involves constructing a transformation memory prediction-pulse collaborative neural network architecture and a deformation evolution analysis model based on the temporal input feature vector, specifically including the following steps: S401: Establish a sequence length determination model based on the time-series input feature vector, and determine the optimal input sequence length through the sequence length determination model; S402: A transformation memory prediction-spiking collaborative neural network architecture is constructed based on the time scale of deformation evolution features and the optimal input sequence length. The transformation memory prediction-spiking collaborative neural network architecture includes a transformation memory predictor and a spiking neural network. The transformation memory predictor captures long-term deformation trends through multi-layer memory units, multi-head attention mechanisms, and multi-scale time window fusion. The spiking neural network uses an improved leakage integral discharge neuron model combined with a dynamic threshold mechanism to identify sudden deformation events. The two networks work together through a dynamic feature gating fusion framework. The dynamic feature gating fusion framework satisfies the following relationship:

[0021] In the formula, For the final prediction result, To transform the prediction output of the memory predictor model, This is the predicted output of the spiking cooperative neural network model. For fusion function, For time-varying context information; S403: Based on the time-series input feature vector, calculate the deformation state encoding vector, deformation propagation velocity vector, and key time nodes respectively, and fuse the deformation state encoding vector, deformation propagation velocity vector, and key time nodes to obtain the deformation evolution analysis model.

[0022] In one example of the present invention, step S50 specifically includes the following steps: S51: Combining historical data, simulation data, and expert knowledge from the multi-parameter monitoring database, historical data analysis, simulation results integration, and expert experience introduction are performed sequentially to obtain the initial comprehensive early warning threshold; then, an adaptive threshold calculation strategy is adopted to dynamically adjust the initial comprehensive early warning threshold based on the early warning feedback effect during operation; The initial comprehensive early warning threshold satisfies the following relationship:

[0023] In the formula, For the first The warning threshold for final fusion at the level, , , The fusion weighting coefficients for historical data. For the first The initial threshold for level-one early warning based on statistical results from historical data. The simulation results are as follows: A critical deformation index. The first expert consensus Level 1 warning threshold; The adaptive threshold calculation strategy involves iteratively updating the threshold based on the initial comprehensive early warning threshold and the actual early warning effect, satisfying the following relationship:

[0024] in, The sensitivity coefficient for threshold adjustment. For the first The number of false alarms in the level-one warning For the first The number of missed reports in the level-one early warning system. For the first Total number of alarms at the Level 1 warning level; S52: Based on the obtained early warning threshold, establish a prediction, verification and optimization loop mechanism to output, evaluate and optimize the prediction results. In the prediction phase, the prediction result is obtained based on the fusion output of the transition memory predictor model and the spiking neural network model, satisfying the following relationship:

[0025] In the formula, Input features In prediction depth The following prediction results These are the prediction functions for the transition memory predictor model and the spiking neural network model, respectively. , For the parameter sets of their respective models, and To predict depth-related dynamic fusion weights; The verification phase compares the predicted results with the actual observed values, constructs a closed-loop verification mechanism, and evaluates the current prediction accuracy, satisfying the following relationship:

[0026] In the formula, To comprehensively assess the error, , Here, represents the weighting coefficients for short-term and long-term evaluations, and N and M represent the number of sampling points for short-term and long-term evaluations, respectively. , Assign higher weights to recent data in the time-weighted function. , These are the predicted value and the actual observed value, respectively.

[0027] Another objective of this invention is to propose an intelligent prediction system for roadway surrounding rock based on point cloud fusion, comprising: The parameter database module is configured to acquire basic data and deformation monitoring data of the surrounding rock of the roadway, and to build a data acquisition and monitoring mechanism to process the basic data and deformation monitoring data to obtain a multi-parameter monitoring database. The boundary constraint module is configured to construct a point cloud data processing model based on a multi-parameter monitoring database, establish equations for roadway main axis identification and path extraction, boundary condition correction models, and comprehensive deformation evaluation models, and establish unified boundary conditions for multi-source data to analyze the point cloud data processing model, roadway main axis identification and path extraction equations, boundary condition correction models, and comprehensive deformation evaluation models to obtain multi-source boundary constraint conditions. The loss function module is configured to design a dual-model network architecture, obtain a dual-model constraint loss function based on the dual-model network architecture and multi-source boundary constraints, perform fitting analysis on the dual-model constraint loss function, and obtain a comprehensive constraint loss function of deformation trend-sudden event index. The architecture and model building module is configured to obtain the temporal input feature vector based on the initial deformation state encoding vector, deformation history feature vector, and environmental process feature vector, and to build a transformation memory prediction-pulse collaborative neural network architecture and deformation evolution analysis model based on the temporal input feature vector; The prediction and monitoring module is configured to combine a multi-parameter monitoring database, obtain prediction results based on a comprehensive constraint loss function, a transformation memory prediction-pulse collaborative neural network architecture, and a deformation evolution analysis model, and perform closed-loop verification and optimization of the prediction results by dynamically adjusting the early warning threshold, thereby realizing the prediction and monitoring of instantaneous deformation and long-term trends of the surrounding rock in the roadway.

[0028] Compared with the prior art, the present invention has the following beneficial effects: This invention, through the scientific and rational arrangement of a three-dimensional laser scanner, a target network, and traditional monitoring devices, and the establishment of an efficient data fusion mechanism, can comprehensively acquire the surrounding rock deformation data, mechanical parameters, and environmental condition data required by the transformation memory predictor and the pulse collaborative neural network algorithm.

[0029] This invention deploys displacement sensors at key locations in the roadway to monitor changes in surrounding rock deformation in real time; deploys stress sensors at typical cross-sections to accurately monitor the stress state of the surrounding rock; and deploys crack monitoring equipment in deformation-sensitive areas to track crack development dynamics. Simultaneously, it establishes an optimized monitoring point layout and data quality assessment mechanism to ensure the validity and reliability of the monitoring data.

[0030] This invention achieves continuous monitoring of the entire cross-section of the surrounding rock in a roadway through three-dimensional laser scanning technology; it enables the identification of the main axis of the roadway and the extraction of the time series of roof displacement through point cloud data processing and feature extraction; it constructs a dual-model collaborative mechanism of conversion memory predictor and spiking neural network to capture long-term deformation trends and sudden event characteristics respectively; and it establishes scientific monitoring parameter critical values ​​and early warning values ​​based on numerical simulation technology and historical data analysis to predict the deformation trend of the surrounding rock.

[0031] This invention achieves continuous full-section monitoring in complex tunnel environments by collaboratively deploying a 3D laser scanner, target network, and traditional monitoring devices. The system relies on a point cloud processing module for real-time data processing, employing various filtering and feature extraction algorithms to achieve efficient tunnel main axis identification and cross-sectional analysis. Spatial geometric constraints are introduced to ensure data processing conforms to the actual tunnel morphology. A dual-model prediction mechanism optimizes the early warning strategy, and an early warning feedback module enables dynamic adjustment of monitoring parameter thresholds and visualized output of prediction results. The combined design of a transformation memory predictor and a spiking neural network ensures comprehensive capture of long-term deformation trends and sudden events, while adaptive threshold calculation provides a scientific basis for early warning.

[0032] This invention features a flexible layout that can be adjusted according to the progress of tunnel excavation, and it possesses online learning and model updating capabilities. Practical application verification shows that this invention significantly improves monitoring accuracy and early warning precision, effectively preventing roof collapse accidents in thick coal seams and providing reliable technical support for safe coal mine production.

[0033] The preferred embodiments of the invention will be described in more detail below with reference to the accompanying drawings, so as to facilitate an understanding of the features and advantages of the invention. Attached Figure Description

[0034] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings of the embodiments of the present invention will be briefly described below. The drawings are merely illustrative of some embodiments of the present invention and are not intended to limit the scope of the present invention to all embodiments.

[0035] Figure 1 This is a flowchart of an intelligent prediction method for roadway surrounding rock based on point cloud fusion according to an embodiment of the present invention; Figure 2 This is a 3D view of a slice of point cloud after noise reduction processing at X=172 and Y=-227 according to an embodiment of the present invention. Figure 3 This is a 3D view of a slice of point cloud after noise reduction processing at X=189 and Y=-279 according to an embodiment of the present invention. Figure 4 This is a 3D view of a slice of point cloud after noise reduction processing at X=107 and Y=-31.5 according to an embodiment of the present invention. Figure 5 This is a schematic diagram of point cloud slicing processing according to an embodiment of the present invention; Figure 6 This is a schematic diagram of the connection of the two-dimensional projection plane point cloud at X=107 and Y=-31.5 according to an embodiment of the present invention. Figure 7 This is a schematic diagram of the connection of the two-dimensional projection plane point cloud at X=172 and Y=-227 according to an embodiment of the present invention; Figure 8 This is a schematic diagram of the connection of the two-dimensional projection plane point cloud at X=189 and Y=-279 according to an embodiment of the present invention; Figure 9 This is a schematic diagram of point cloud top panel recognition according to an embodiment of the present invention. Detailed Implementation

[0036] To make the objectives, technical solutions, and advantages of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. The same reference numerals in the drawings represent the same components. It should be noted that the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the described embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0037] Unless otherwise defined, the technical or scientific terms used herein shall have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms “first,” “second,” and similar terms used in this patent application specification and claims do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, “an” or “a” and similar terms do not necessarily indicate a quantity limitation. Terms such as “comprising” or “including” mean that the element or object preceding the word encompasses the element or object listed following the word and its equivalents, without excluding other elements or objects. Terms such as “connected” or “linked” are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as “upper,” “lower,” “left,” and “right” are used only to indicate relative positional relationships; these relative positional relationships may change accordingly when the absolute position of the described object changes.

[0038] According to a first aspect of the present invention, a method for intelligent prediction of roadway surrounding rock based on point cloud fusion is provided, such as... Figure 1 As shown, it includes the following steps: S10: Obtain basic data and deformation monitoring data of the surrounding rock of the roadway, and construct a data acquisition and monitoring mechanism to process the basic data and deformation monitoring data to obtain a multi-parameter monitoring database; S20: Based on the multi-parameter monitoring database, a point cloud data processing model is constructed, and equations for roadway main axis identification and path extraction, boundary condition correction models, and comprehensive deformation evaluation models are established. Unified boundary conditions for multi-source data are established to analyze the point cloud data processing model, the roadway main axis identification and path extraction equations and boundary condition correction models, and the comprehensive deformation evaluation models, and obtain multi-source boundary constraint conditions. S30: Design a dual-model network architecture, obtain the dual-model constraint loss function based on the dual-model network architecture and multi-source boundary constraints, perform fitting analysis on the dual-model constraint loss function, and obtain the comprehensive constraint loss function of deformation trend-sudden event index. S40: Based on the initial deformation state encoding vector, deformation history feature vector, and environmental process feature vector, the temporal input feature vector is obtained. Based on the temporal input feature vector, the Transformer Memory Predictor-Spiking Neural Network (TMP-SNN) architecture and deformation evolution analysis model are constructed. S50: Combining a multi-parameter monitoring database, prediction results are obtained based on a comprehensive constraint loss function, a transformation memory prediction-pulse collaborative neural network architecture, and a deformation evolution analysis model. The prediction results are then dynamically adjusted through an early warning threshold to perform closed-loop verification and optimization, thereby enabling the prediction and monitoring of instantaneous deformation and long-term trends of the surrounding rock in the roadway.

[0039] This invention, through the scientific and rational arrangement of a three-dimensional laser scanner, a target network, and traditional monitoring devices, and the establishment of an efficient data fusion mechanism, can comprehensively acquire the surrounding rock deformation data, mechanical parameters, and environmental condition data required by the transformation memory predictor and the pulse collaborative neural network algorithm.

[0040] This prediction method deploys displacement sensors at key locations in the roadway to monitor changes in surrounding rock deformation in real time; stress sensors are deployed at typical cross-sections to accurately monitor the stress state of the surrounding rock; and crack monitoring equipment is deployed in deformation-sensitive areas to track crack development dynamics. Simultaneously, a mechanism for optimizing the layout of monitoring points and evaluating data quality is established to ensure the effectiveness and reliability of the monitoring data.

[0041] This prediction method achieves continuous monitoring of the entire cross-section of the surrounding rock in the roadway through three-dimensional laser scanning technology; it identifies the main axis of the roadway and extracts the time series of roof displacement through point cloud data processing and feature extraction; it constructs a dual-model collaborative mechanism of transformation memory predictor and spiking neural network to capture long-term deformation trends and sudden event characteristics respectively; and it establishes scientific monitoring parameter critical values ​​and early warning values ​​based on numerical simulation technology and historical data analysis to predict the deformation trend of the surrounding rock.

[0042] This prediction method achieves continuous full-section monitoring in complex tunnel environments by collaboratively deploying a 3D laser scanner, target network, and traditional monitoring devices. The system relies on a point cloud processing module for real-time data processing, employing various filtering and feature extraction algorithms to achieve efficient tunnel main axis identification and cross-sectional analysis. Spatial geometric constraints are introduced to ensure data processing conforms to the actual tunnel morphology. A dual-model prediction mechanism optimizes the early warning strategy, and an early warning feedback module enables dynamic adjustment of monitoring parameter thresholds and visualization of prediction results. The combined design of a transformation memory predictor and a spiking neural network ensures comprehensive capture of long-term deformation trends and sudden events, while adaptive threshold calculation provides a scientific basis for early warning.

[0043] This prediction method is flexible in layout and can be adjusted according to the progress of roadway excavation. It also has online learning and model update capabilities. Practical application verification shows that this invention significantly improves monitoring accuracy and early warning accuracy, effectively preventing roof collapse accidents in thick coal seam roadways, and providing reliable technical support for safe coal mine production.

[0044] In one example of the present invention, in step S10, acquiring basic data and deformation monitoring data of the surrounding rock of the roadway specifically includes: acquiring mechanical parameters of the surrounding rock of the coal seam through coal and rock sample collection and mechanical property testing; analyzing the stress distribution based on a dual-loop hydraulic fracturing stress measurement system; establishing a three-dimensional laser scanning monitoring system; acquiring initial point cloud data through target network and coordinate system deployment; and acquiring deformation monitoring data based on periodic scanning and traditional monitoring devices.

[0045] Specifically, basic data on the surrounding rock of the roadway were obtained: To ensure the comprehensiveness and accuracy of the mechanical and geological parameters of the surrounding rock, coal and rock samples were collected in representative coal seams, roof, and floor areas, and systematic data acquisition was carried out in conjunction with on-site geological survey data. Standardized tests were conducted on the collected samples using rock mechanics testing equipment, including uniaxial compression tests, Brazilian splitting indirect tensile tests, and direct shear tests, to obtain key mechanical parameters such as elastic modulus, Poisson's ratio, compressive strength, tensile strength, cohesion, and internal friction angle. These parameters serve as the basic inputs for subsequent numerical analysis and surrounding rock deformation assessment.

[0046] Simultaneously, to obtain the in-situ in-situ stress state of the surrounding rock in the tunnel, no fewer than three measuring points were set up in the test area, with a spacing of no less than 30 m between the measuring points. A dual-loop hydraulic fracturing stress measurement system was used for on-site testing. By applying pressure in stages using packers and recording the fracturing pressure curve and re-tensioning pressure curve, the values ​​and directions of the maximum and minimum horizontal principal stresses were determined, and parameters such as rock fracture pressure, re-tensioning pressure, and in-situ tensile strength were obtained. The in-situ stress test results were integrated with the rock mechanics test results to form a dataset of surrounding rock geomechanical parameters.

[0047] Acquiring deformation monitoring data of roadway surrounding rock: In order to achieve continuous monitoring of the entire cross section of roadway surrounding rock, the embodiment adopts a combination of three-dimensional laser scanning technology and traditional point monitoring equipment to acquire deformation monitoring data.

[0048] In terms of 3D laser scanning, the test area is comprehensively scanned and collected according to a preset monitoring cycle. Each scan is calibrated with the initially deployed fixed reference target as a benchmark to ensure that the point cloud data from different periods are under a unified mine coordinate system. Spatial point cloud coordinate information is recorded during the scanning process to form the original 3D point cloud dataset.

[0049] To ensure the accuracy of monitoring at key locations, traditional monitoring equipment such as delamination meters and anchor bolt force gauges are also deployed in deformation-sensitive areas such as the top plate and sidewalls to collect local displacement and stress data in real time, thus achieving complementarity between spatial scanning data and point monitoring data.

[0050] The collected raw point cloud data is processed. The raw scan data is imported into the point cloud processing system, and the spatial position parameters of the scanner in the mine coordinate system are calculated using the resection algorithm. The discrete point cloud is then subjected to a unified coordinate transformation to generate a three-dimensional digital point cloud model in a unified coordinate system, providing basic data for subsequent point cloud processing and deformation analysis.

[0051] In one example of the present invention, in step S10, a data acquisition and monitoring model is constructed to process the basic data and deformation monitoring data to obtain a multi-parameter monitoring database, including the following steps: S11: Determine the scope of the calculation domain based on the tunnel model, and set the coordinate system spatiotemporal calculation domain and the boundary conditions of ground stress and deformation of the tunnel according to the scope of the calculation domain. S12: Determine the sampling point density weights based on the spatiotemporal computational domain and the boundary conditions of geostress and deformation; S13: Establish a sampling encryption algorithm and coordinate normalization mapping function for deformation-sensitive areas; S14: The basic data and deformation monitoring data are adjusted by combining the computational domain range, sampling point density weight, deformation-sensitive area sampling densification algorithm and coordinate standardization mapping function to obtain the adjusted basic data and deformation monitoring data. S15: Use a data acquisition and monitoring model to fuse the adjusted basic data and deformation monitoring data to obtain a multi-parameter monitoring database.

[0052] In other words, to achieve unified management and efficient utilization of multi-source data, time synchronization, coordinate unification, and feature standardization are performed on data from different sources and at different scales, establishing a multi-source data fusion mechanism. By constructing a unified data structure, coal and rock mechanical parameters, geostress parameters, three-dimensional point cloud data, and traditional monitoring data are integrated to form a multi-parameter monitoring database.

[0053] The multi-parameter monitoring database not only records various physical and mechanical parameters and their spatial distribution information, but also saves deformation evolution data under different monitoring periods, providing data support for subsequent point cloud feature extraction, trend prediction and sudden deformation identification.

[0054] Furthermore, the construction method of the multi-parameter monitoring database in this embodiment is only one optional implementation of the present invention. In other embodiments, the data acquisition method and fusion strategy can be adjusted according to the tunnel size, monitoring accuracy requirements and on-site construction conditions to improve the adaptability and engineering practicality of the database.

[0055] In one example of the present invention, step S20 specifically includes the following steps: S21: Establish a point cloud data processing model based on statistical outlier filtering and curvature-preserving downsampling; Specifically, in the process of three-dimensional scanning monitoring of the surrounding rock of the tunnel, in order to eliminate noise in the original point cloud data and improve processing efficiency, it is necessary to integrate statistical outlier filtering and curvature-preserving downsampling methods to construct a point cloud data processing model system. This model system not only covers the complex processing of noise elimination, but also fully considers the need to preserve key geometric features, providing a high-quality data foundation for subsequent main axis identification and deformation analysis.

[0056] In data denoising, the raw point cloud data often contains a large number of noise points due to factors such as measurement environment, equipment accuracy, and human factors. These noise points can seriously affect the accuracy of subsequent analysis. For the special environment of coal mine roadways, methods such as statistical outlier filtering, radius filtering, conditional filtering, and surface normal-based filtering are used to effectively eliminate noise points caused by measurement environment, equipment accuracy, and human factors.

[0057] In terms of point cloud simplification and optimization, the original point cloud data is huge (usually tens of millions of points are generated each time scan), and it needs to be effectively simplified to improve processing efficiency while maintaining key geometric features. This can be achieved by using methods such as uniform downsampling, curvature-preserving downsampling, importance-based simplification, and multi-resolution representation.

[0058] S22: Establish equations for identifying the main axis of the roadway and extracting the center path, and construct a boundary condition correction model based on deformation influencing factors; wherein, the equations for identifying the main axis of the roadway and extracting the center path include: the formula for identifying the main axis of the roadway and the formula for extracting the center path; To analyze the three-dimensional spatial morphological characteristics of the tunnel, it is necessary to establish equations for tunnel main axis identification and path extraction, and to establish a boundary condition correction model based on the effects of deformation influencing factors, so as to improve the accuracy and reliability of the main axis identification and path extraction results.

[0059] The formula for identifying the main axis of the tunnel is as follows:

[0060] in, The main axis direction vector For the point cloud dataset, the first Each point is a three-dimensional vector. This represents the total number of points in the point cloud. For the centroid of the point cloud, For vectors transpose; The center path is fitted using RANSAC to the roadway centerline, and the distance from the point to the line is calculated as follows:

[0061] in, This represents the straight line that best fits the point cloud data. For the shape of The linear model, Let be the distance from the point to the line. For the threshold; In the extraction of the top plate displacement time series, the deformation calculation satisfies:

[0062] in, For the region In time The average deformation, For the region The number of sampling points.

[0063] Based on the identified main axis of the tunnel, a slicing strategy with adjustable parameters is used to generate a series of cross-sections perpendicular to the main axis, in the following form:

[0064] in, This indicates that the parameter represents the data. The slice plane described; used to introduce the definition or constraints of the plane equation. For the parameterized values ​​of the slice, The centerline position of the slice plane. Point To the center line of the slice vector and normal vector Orthogonal.

[0065] Based on the identified main axis direction vector of the tunnel, a series of slicing planes perpendicular to the main axis are constructed. The point cloud is then processed by layering and cropping to form cross-sectional point sets at different locations. To visually demonstrate the slicing construction process and cross-sectional distribution, a schematic diagram of the point cloud slicing process is provided; please refer to [link / reference]. Figure 2 , Figure 3 , Figure 4 and Figure 5 .

[0066] The subset of point cloud contained in each slice is defined as follows:

[0067] in, To be included in the slice The set of points in, For point Belongs to point cloud dataset , The thickness of the slice. This represents the maximum tolerable distance of a point from the centerline.

[0068] Projecting each slice of point cloud onto a plane perpendicular to the principal axis enables standardized analysis of the tunnel cross-sectional morphology.

[0069] in, For point The projection point on the slice plane.

[0070] The sliced ​​point cloud is projected onto a two-dimensional plane perpendicular to the principal axis, and the projected point set is connected and its contour is constructed. To visually reflect the cross-sectional morphology features after two-dimensional projection, a schematic diagram of the connection of the point cloud on the two-dimensional projection plane is drawn; please refer to [link / reference]. Figure 6 , Figure 7 and Figure 8 .

[0071] The boundary condition correction model based on the effects of deformation influencing factors comprehensively considers the correction of deformation boundary conditions by factors such as changes in roadway geometry, mining stress, and support constraints, and can more accurately describe the roadway deformation characteristics under actual working conditions.

[0072] S23: Establish a comprehensive deformation evaluation model based on the experimental information of the surrounding rock deformation in the tunnel and the time-accumulated deformation model; Specifically, in the field of tunnel engineering technology, the deformation assessment of surrounding rock under complex working conditions such as mining is crucial for the structural safety of tunnels. In this embodiment, based on experimental data of tunnel surrounding rock deformation and a time-cumulative deformation model, a comprehensive deformation evaluation model system is constructed. This system can accurately predict the degree of deformation of the surrounding rock under different conditions, providing a scientific basis for the safe design and maintenance of tunnels.

[0073] Based on the extraction of tunnel cross-section contours, the convex hull algorithm and α-shape algorithm are used to obtain the accurate cross-section contours:

[0074] in, For slices The corresponding cross-sectional profile point set, For slices The set of regular cross-sectional contour points extracted by the convex hull algorithm For slices pass -The set of irregular cross-sectional contour points extracted by the shape algorithm For slices The point cloud data set on the above, for The parameters of the shape algorithm control the tightness of the contour boundaries. For point set implement Scanning algorithm, For execution on point sets The shape algorithm extracts contours suitable for concave features.

[0075] Roof area identification is based on the roadway's geometric characteristics and the actual locations of deployed targets. The roof area is automatically identified in a standardized coordinate system, and its normal vector is standardized as follows:

[0076] in, For the first The normal vector estimated at each point. Let the coordinates be the center coordinates of the point cloud. This is for dot product calculation.

[0077] The top plate point selection strategy is as follows:

[0078] in, For the selected set of top plate points, For the first The normal vector of each point Quantity, The cosine value of the corresponding angle. For the first points Coordinate values It represents the Xth percentile of the Z-coordinate.

[0079] For each cross-section location, calculate the difference between the current cross-section and the initial reference cross-section:

[0080] In the formula, For the surrounding rock point p during the monitoring time The amount of deformation at that time; To monitor the surrounding rock point p at time t, and the reference outline of the roadway at time t. The minimum distance; Based on the standard outline and the surrounding rock monitoring points Feature matching corresponds to the pairing points; This serves as the standard reference outline for the roadway at the base time. Based on the standard outline and the surrounding rock monitoring points Feature matching corresponding pairing points Standard reference profile of the roadway at the reference time The minimum distance.

[0081] After the cross-sectional contour is extracted, the top plate region is automatically identified based on normal vector unification and spatial position constraints. A point cloud diagram illustrating the top plate identification process is provided; please refer to [link to diagram]. Figure 9 .

[0082] The average deformation is defined separately for the top plate, bottom plate, and side panels:

[0083] in, For the region In time The average deformation, For the region The number of sampling points.

[0084] 3D Deformation Field Reconstruction:

[0085] in, This indicates the spatial interpolation method used.

[0086] S24: By coupling multi-source data and unifying boundary conditions, the parameter boundary conditions of the point cloud data processing model, the roadway main axis identification and path extraction equation and boundary condition correction model and the comprehensive deformation evaluation model are analyzed to obtain deformation boundary conditions, trend boundary conditions and sudden boundary conditions. S25: Combining deformation boundary conditions, trend boundary conditions, and sudden boundary conditions yields multi-source boundary constraint conditions.

[0087] In the case of multi-source data coupling analysis for roadway surrounding rock deformation monitoring, in order to comprehensively and accurately describe the deformation behavior of the system, it is necessary to comprehensively analyze multiple key processes such as point cloud geometric features, deformation time sequence laws, and sudden event detection. By unifying the boundary conditions through multi-source data coupling, the above-mentioned deformation boundary conditions, trend boundary conditions, and sudden boundary conditions are derived respectively.

[0088] The deformation boundary conditions indicate that the deformation of the surrounding rock approaches zero at the far-field boundary and cannot abruptly change at the boundary, suggesting that the boundary plays a certain role in constraining the propagation of deformation, and that deformation mainly develops and accumulates within the surrounding rock. The trend boundary condition represents the memory transmission mechanism of long-term deformation evolution; the sudden boundary condition means that the neuron dynamics model can sensitively respond to rapidly changing abrupt signals. These three boundary conditions influence and constrain each other, jointly determining the response characteristics of the roadway surrounding rock under the coupling effect of multi-source data.

[0089] Furthermore, the analysis method for multi-source boundary constraints in this embodiment is merely an optional condition of the present invention. In other embodiments, the analysis method for multi-source boundary constraints can be optimized based on the multi-parameter monitoring database and the target for monitoring and predicting deformation of the surrounding rock in the tunnel. In practical application scenarios, the size and completeness of the multi-parameter monitoring database vary, and the requirements for multi-source boundary constraints differ at different stages. Optimizing the analysis method for multi-source boundary constraints can more accurately determine the various parameters in the boundary conditions.

[0090] In one example of the present invention, in step S23, the deformation boundary condition, the trend boundary condition, and the sudden boundary condition are respectively: The parametric boundary conditions of the comprehensive deformation evaluation model are analyzed by coupling multi-source data and unifying boundary conditions to obtain the deformation boundary conditions; among them, the deformation boundary conditions satisfy the following relationship:

[0091] In the formula, For the surrounding rock point p during the monitoring time The amount of deformation at that time; To monitor the surrounding rock point p at time t, and the reference outline of the roadway at time t. The minimum distance; Based on the baseline standard outline and the surrounding rock monitoring points Feature matching corresponds to the pairing points; This serves as the standard reference outline for the roadway at the base time. Based on the baseline standard outline and the surrounding rock monitoring points Feature matching corresponding pairing points Standard reference profile of the roadway at the reference time The minimum distance; By analyzing the parameter boundary conditions of the memory unit in the transformation memory predictor model through multi-source data coupling and unified boundary conditions, trend boundary conditions are obtained; among them, the trend boundary conditions satisfy the following relationship:

[0092] In the formula, The current cell state is used to control the flow of information; For the forget gate, multiply by To determine how much information is forgotten; This represents the cell state at the previous moment; The output of the input gate determines the update of the cell state at the current moment; Representing the candidate cell state, it is the alternative feature information to be included in the memory unit at the current time step; This is element-wise multiplication; By analyzing the parametric boundary conditions of neuron dynamics in the spiking cooperative neural network model through multi-source data coupling and unified boundary conditions, burst boundary conditions are obtained. These burst boundary conditions satisfy the following relationship:

[0093] In the formula, The membrane potential time constant is used to control the potential recovery rate; Current time Membrane potential at time t; This is the resting potential, which is also the potential of a neuron when it is not stimulated. For membrane resistance; For synaptic input current; For adaptive current weighting coefficients; For adaptive current.

[0094] In one example of the present invention, step S30 specifically includes the following steps: S31: Set up a multi-dimensional parameter vector input mechanism and an input parameter standardization processing mechanism in the network architecture; Specifically, in order to eliminate the differences in units and numerical ranges between different parameters and improve the training efficiency and stability of the model, the input parameters are standardized, where feature standardization satisfies the following relationship:

[0095] in, For the standardized features to the input features The results after standardization For input features, Input features The mean, Input features The standard deviation is used for data scaling. To prevent the denominator from being zero and to ensure stability, a small constant is used. The scaling factor is used to control the scale of the standardized data. The offset is used to adjust the standardized data to fit the model's learning.

[0096] Temporal difference features are used to capture dynamic changes in time series data, such as deformation rate and acceleration. These dynamic features are of great significance for predicting and monitoring disasters such as roof collapse.

[0097] in, The first-order difference feature represents the current time step. and the previous moment The difference between them is used to capture the rate of change of the data. The feature value at the current time. The eigenvalue is the feature value from the previous time step.

[0098]

[0099] in, The second-order difference feature represents the current time step. And the previous two moments , The difference between 2 and 3 is used to capture changes in the acceleration of the data. These are the characteristic values ​​from the first two time points.

[0100] S32: Based on the multi-dimensional parameter vector input mechanism, the input parameter standardization processing mechanism, and the multi-source boundary constraints, the output results of the trend prediction layer, the sudden event detection layer, and the fusion prediction layer are obtained; S33: Construct a dual-model constrained loss function based on the outputs of the trend prediction layer, the sudden event detection layer, and the fusion prediction layer. Based on the outputs of the trend prediction layer, the sudden event detection layer, and the fusion prediction layer, construct a dual-model constrained loss function system. To ensure the physical rationality and engineering reliability of the prediction results from the dual-model collaborative prediction framework, the prediction equation controlling the deformation evolution process is embedded into the loss function of the neural network.

[0101] S34: Perform fitting analysis on the dual-model constraint loss function to obtain the comprehensive constraint loss function of deformation trend-emergency index; The comprehensive constraint loss function satisfies the following relationship:

[0102] In the formula, The current parameters are used to measure the overall loss function of the model. The overall prediction error, The mean squared error loss penalizes the squared error between the predicted and actual values. To weight safety risks, higher weights are given to errors in high-risk areas. The weighting coefficients are used to control the balance between ordinary mean square error and risk-weighted loss; The risk weight function in the safety risk-weighted loss satisfies the following relationship:

[0103] In the formula, For risk weighting factors of individual observations, The magnitude coefficient used to weight risk determines the degree to which the weight is amplified when the risk approaches a safety threshold. A coefficient to control the rate of weight descent. As a safety threshold, This represents the current observed value.

[0104] Adam optimizer update:

[0105] in, For the first Model parameters at the next iteration For the first Model parameters after the next iteration The update magnitude is controlled by the current step size. For first-order moment estimation, For second-order moment estimation, It is a small constant.

[0106] In terms of model optimization and tuning, specific aspects include parameter importance assessment, model pruning and compression, multi-scale feature fusion, and adversarial training.

[0107] The importance of parameters is assessed using the following formula:

[0108] in, For the first The importance score of each parameter Let be the overall loss function of the model. For the first Each weight parameter, This is the gradient of the loss function with respect to this parameter.

[0109] Knowledge distillation loss function:

[0110] in, The total loss from knowledge distillation. To adjust the ratio of the two losses using the balance coefficient, For real labels With lightweight model prediction Cross-entropy loss between For the Softmax function, Temperature parameters used to control the smoothness of softmax.

[0111] Multi-scale feature fusion methods:

[0112] in, This represents the hidden features after fusion. As a set of time scales, For a specific scale, the feature transformation function, To be consistent with the scale The relevant weight matrix, For delay Data input at each time step.

[0113] The process of combat training:

[0114] in, The optimization goal is to find the optimal parameters. To minimize the loss, For sample pairs under the data distribution Take the expected value. To limit the disturbance amplitude to no more than , For the perturbed samples The loss value.

[0115] Furthermore, the method for constructing the comprehensive constraint loss function of deformation trend-emergency index in this embodiment is only an optional condition of the present invention. In other embodiments, the method for constructing the comprehensive constraint loss function can be optimized according to the multi-source boundary constraint conditions and the deformation prediction target of the surrounding rock of the roadway, so that the deformation trend and the emergency index are closer to the real situation, thereby reducing the error of the prediction result.

[0116] In one example of the present invention, step S40, obtaining the temporal input feature vector based on the initial deformation state encoding vector, the deformation history feature vector, and the environmental process feature vector, specifically includes the following steps: S41: Based on the network architecture and multi-parameter monitoring database, obtain the spatial deformation distribution characteristics, and obtain the initial deformation state encoding vector according to the spatial deformation distribution characteristics; S42: Construct characteristic expressions for the deformation history and stress process of the surrounding rock, and use the characteristic expressions to obtain the deformation history feature vector; S43: Set up a coding mechanism for environmental conditions and monitoring cycles, and obtain the environmental process feature vector through the coding mechanism; S44: Combine time-varying environmental conditions, initial deformation state encoding vector, deformation history feature vector, and environmental process feature vector to obtain the time-series input feature vector.

[0117] First, construct a Spiking Neural Network (SNN) model.

[0118] Spiking neural networks (SNNs) encompass neuron model design, data encoding, network structure, prediction algorithms, and collaborative mechanisms with Transformer Memory Predictor (TMP) models. They complement the Transformer Memory Predictor (TMP) model effectively.

[0119] This invention employs an improved Leakage Integral Discharge (LIF) model as the basic neuron model, which accurately captures the dynamic characteristics of time-series data while maintaining computational efficiency. The dynamic equations of the improved LIF model are defined as follows:

[0120] in, The membrane potential time constant is used to control the potential recovery rate. Current time Membrane potential at time t, The resting potential is the potential of a neuron when it is not stimulated. For membrane resistance, For synaptic input current, For adaptive current weighting coefficient, For adaptive current.

[0121] The discrete-time-step membrane potential update strategy realizes potential evolution through the following formula:

[0122] in, For time step.

[0123] When the membrane potential exceeds the threshold, the neuron generates a pulse and resets:

[0124] in, A binary output indicating whether a neuron fires a pulse. This is the pulse firing threshold.

[0125] Dynamic adjustment of neuron firing threshold to capture key changes in roadway surrounding rock deformation:

[0126] in, The dynamic threshold for the current time. Based on the static threshold, This is a dynamically adjustable amount.

[0127]

[0128] in, The time constant for dynamic threshold adjustment. The amount of threshold increase caused by each pulse. To be at the pulse generation time The Dirac function at that location.

[0129] Synaptic weights represent the strength of connections between neurons and determine the efficiency of information transmission. This invention employs a two-factor synaptic plasticity rule:

[0130] in, For from the first The first neuron to the second Synaptic weights of individual neurons The learning rate is used to control the speed of weight adjustment. The activity level of preneurons, The target pulse signal for the postneuron. Weight decay coefficient.

[0131] Expression for the pulse trajectory of the preneuron:

[0132] in, This is the time constant of the presynaptic neuron's activity. For the first The activity level of each neuron.

[0133] Then, the temporal input feature vector is obtained based on the initial deformation state encoding vector, the deformation history feature vector, and the environmental process feature vector.

[0134] Existing transformation memory predictor models can predict long-term deformation trends thanks to their memory mechanisms. However, predicting deformation of roadway surrounding rock is a highly complex process involving the transformation of multiple physical mechanisms. In the early stages of mining, deformation is mainly driven by deterministic stress redistribution and loosening of the surrounding rock. In long-term monitoring, deformation evolution is influenced by a combination of time-varying factors such as support effectiveness, environmental changes, and creep effects. Therefore, it is necessary to construct a Transformer Memory Predictor-Spiking Neural Network (TMP-SNN) collaborative prediction framework to achieve the transition from instantaneous deformation analysis to long-term trend prediction.

[0135] Based on the network architecture and multi-parameter monitoring database, spatial deformation distribution characteristics are obtained. An initial deformation state encoding vector is then obtained based on these characteristics, satisfying the following relationship:

[0136] in, This is the initial deformation state encoding vector. For the maximum deformation, This represents the average deformation. For the maximum deformation gradient, For residual stress, This represents the maximum deformation rate.

[0137] Construct characteristic expressions for the deformation history and stress process of the surrounding rock, and obtain the deformation history feature vector using these expressions. The specific expressions are as follows:

[0138] in, The feature vector of the deformation process. This represents the peak deformation. The duration of the impact of mining activities, The average deformation rate, The time integral represents the deformation.

[0139] Simultaneously, a coding mechanism for environmental conditions and monitoring cycles is established. By setting this coding mechanism, environmental process feature vectors can be obtained, with the following specific form:

[0140] in, For environmental process feature vectors, Code the support method. The duration of the impact of mining activities, For monitoring duration.

[0141] Finally, the initial state features, deformation history features, and environmental process features, combined with time-varying environmental conditions, are combined to form a complete time-series input feature vector, the expression of which is:

[0142] in, For the first The temporal input feature vector at time step 1. For the first The environmental feature vector at time t, For the first External load at any given time.

[0143] In one example of the present invention, step S40 involves constructing a transformation memory prediction-pulse collaborative neural network architecture and a deformation evolution analysis model based on the temporal input feature vector, specifically including the following steps: S401: Establish a sequence length determination model based on the time-series input feature vector, and determine the optimal input sequence length through the sequence length determination model; S402: A transformation memory prediction-spiking collaborative neural network architecture is constructed based on the deformation evolution feature time scale and the optimal input sequence length. The transformation memory prediction-spiking collaborative neural network architecture includes a transformation memory predictor and a spiking neural network. The transformation memory predictor captures long-term deformation trends through multi-layer memory units, multi-head attention mechanisms, and multi-scale time window fusion. In short, the multi-head attention mechanism is used as the backbone of temporal feature transformation, combined with external multi-layer memory units, to capture the temporal features of long-term creep deformation of the surrounding rock in the roadway. The spiking neural network uses an improved leakage integral discharge neuron model combined with a dynamic threshold mechanism to identify sudden deformation events. The transformation memory predictor outputs the long-term deformation trend of the surrounding rock, and the spiking neural network identifies instantaneous sudden surrounding rock disasters. Intelligent prediction of roadway surrounding rock deformation is achieved through gating fusion. The two networks work collaboratively through a dynamic feature gating fusion framework. The dynamic feature gating fusion framework satisfies the following relationship:

[0144] In the formula, For the final prediction result, To transform the prediction output of the memory predictor model, This is the predicted output of the spiking cooperative neural network model. For fusion function, For time-varying context information; S403: Based on the time-series input feature vector, calculate the deformation state encoding vector, deformation propagation velocity vector, and key time nodes respectively, and fuse the deformation state encoding vector, deformation propagation velocity vector, and key time nodes to obtain the deformation evolution analysis model.

[0145] Specifically, a sequence length determination model is established based on the time-series input feature vector, and the optimal input sequence length is determined through the sequence length determination model.

[0146] To determine the optimal input sequence length corresponding to the time-series input feature vector, a specialized sequence length determination model needs to be constructed. In the context of predicting the long-term deformation evolution of roadway surrounding rock, due to the multi-timescale characteristics and strong temporal dependence of the deformation evolution, a sequence length determination model is constructed by analyzing the characteristic timescales of the deformation evolution process, and it satisfies the following relationship:

[0147] in, For the optimal input sequence length, The time scale represents the characteristics of elastic deformation. The time scale represents the characteristics of creep deformation. This represents the timescale of environmental impact characteristics.

[0148] A transformation memory prediction-pulse collaborative neural network architecture is constructed based on the time scale of deformation evolution characteristics and the optimal input sequence length.

[0149] In the Transitional Memory Prediction-Spiking Collaborative Neural Network architecture, the Transitional Memory Prediction Unit captures long-term deformation trends through multi-layer memory units, multi-head attention mechanisms, and multi-scale time window fusion. The Spiking Neural Network uses an improved leakage integral discharge neuron model combined with a dynamic threshold mechanism to identify sudden deformation events. The two networks work collaboratively through a dynamic feature gating fusion framework.

[0150] The dynamic feature gating fusion framework satisfies the following relationship:

[0151] in, For the final prediction result, To transform the prediction output of the memory predictor model, This is the predicted output of the spiking cooperative neural network model. For fusion function, This is time-varying context information.

[0152] The fusion function is implemented through a dynamic feature gating mechanism:

[0153] in, , This is a context-based gate function.

[0154] A deformation evolution analysis model is constructed based on the time-series input feature vector. The aforementioned deformation evolution analysis model includes a state coding vector analysis model based on key deformation indices, a propagation pattern recognition algorithm based on deformation evolution speed, and a key time node recognition algorithm based on the second derivative of deformation evolution.

[0155] Deformation evolution modeling, as a core component of the transformation memory prediction-pulse collaborative neural network architecture, can perform sequential characterization of the deformation state of the surrounding rock in the roadway, identification of propagation patterns, prediction of key time nodes, and quantification of deformation development speed, thereby enabling deep learning and accurate prediction of deformation evolution laws.

[0156] Construct a state coding vector analysis model based on key deformation indicators, the expression of which is:

[0157] in, Let be the deformation state encoding vector at time t. Let be the maximum deformation at time t. Let be the average deformation at time t. Let be the standard deviation of the deformation at time t. Let be the ratio of the deformation volume at time t.

[0158] A propagation pattern recognition algorithm based on deformation evolution velocity is constructed, and its expression is:

[0159] in, For position The deformation propagation velocity vector at time t, The rate of change over time is the deformation. The gradient vector in the deformation space. This is a numerical stability parameter.

[0160] A key time node identification algorithm based on the second derivative of deformation evolution is constructed, and its determination formula is as follows:

[0161] in, This is a key time point. Let be the overall deformation index at time t. For deformation acceleration, The rate of change of deformation acceleration, This is the threshold for the rate of change of acceleration.

[0162] The Transformer Memory Predictor (TMP) model and the Spiking Neural Network (SNN) model are deeply integrated. The two models capture the characteristics of roadway surrounding rock deformation from different perspectives. The TMP model focuses on extracting global trends and periodic changes in long-term series, while the Spiking Neural Network model is good at capturing sudden events and rapidly changing local features.

[0163] To improve system operating efficiency, this invention adopts a hierarchical parallel computing architecture, in which the conversion memory predictor and the spiking neural network model run in parallel on independent processing units, each processing the same input data stream. The large-scale monitoring data is processed in segments, with each segment independently fed into a subset of the model. The results are then merged, and a sliding window technique is used to maintain partial overlap between adjacent time windows, achieving a smooth transition for continuous prediction.

[0164] The basic ensemble method combines the prediction results of the two models through a weighted average:

[0165] in, , Each has a fixed weight.

[0166] Residual correction integration formula:

[0167] in, This represents the residual portion predicted by the spiking cooperative neural network.

[0168] Based on the time scale, the optimal strategy is applied to different prediction time periods:

[0169] in, The result is a weighted average of the outputs of the basic prediction transition memory predictor model and the spiking collaborative neural network model.

[0170] Introducing an adaptive weighting mechanism based on historical performance evaluation:

[0171] in, For the width of the sliding window, This is the loss function.

[0172]

[0173] in, Indicates the first A prediction model at time... The fusion weight, Indicates the first The model at time... The prediction error Adjust the hyperparameters for weight sensitivity. It is a natural exponential function. This represents the total number of models. Through exponential normalization, models with smaller errors are assigned higher weights, thereby improving the accuracy of the fusion prediction results.

[0174] In one example of the present invention, step S50 specifically includes the following steps: S51: Combining historical data, simulation data, and expert knowledge from the multi-parameter monitoring database, historical data analysis, simulation results integration, and expert experience introduction are performed sequentially to obtain the initial comprehensive early warning threshold; then, an adaptive threshold calculation strategy is adopted to dynamically adjust the initial comprehensive early warning threshold based on the early warning feedback effect during operation; The initial comprehensive early warning threshold satisfies the following relationship:

[0175] In the formula, For the first The warning threshold for final fusion at the level, , , The fusion weighting coefficients for historical data. For the first The initial threshold for level-one early warning based on statistical results from historical data. The simulation results are as follows: A critical deformation index. The first expert consensus Level 1 warning threshold; The adaptive threshold calculation strategy involves iteratively updating the threshold based on the initial comprehensive early warning threshold and the actual early warning effect, satisfying the following relationship:

[0176] in, The sensitivity coefficient for threshold adjustment. For the first The number of false alarms in the level-one warning For the first The number of missed reports in the level-one early warning system. For the first Total number of alarms at the Level 1 warning level; S52: Based on the obtained early warning threshold, establish a prediction, verification and optimization loop mechanism to output, evaluate and optimize the prediction results. In the prediction phase, the prediction result is obtained based on the fusion output of the transition memory predictor model and the spiking neural network model, satisfying the following relationship:

[0177] In the formula, Input features In prediction depth The following prediction results These are the prediction functions for the transition memory predictor model and the spiking neural network model, respectively. , For the parameter sets of their respective models, and To predict depth-related dynamic fusion weights; The verification phase compares the predicted results with the actual observed values, constructs a closed-loop verification mechanism, and evaluates the current prediction accuracy, satisfying the following relationship:

[0178] In the formula, To comprehensively assess the error, , Here, represents the weighting coefficients for short-term and long-term evaluations, and N and M represent the number of sampling points for short-term and long-term evaluations, respectively. , Assign higher weights to recent data in the time-weighted function. , These are the predicted value and the actual observed value, respectively.

[0179] Specifically, the early warning threshold is a key standard for judging the safety status of roadways. This invention adopts a dynamic adjustment strategy to ensure that the early warning threshold is scientifically reasonable and has real-time adaptability. This includes historical data analysis, integration of simulation results, introduction of expert experience, and adaptive threshold calculation.

[0180] Historical data analysis: The system collects and analyzes historical deformation data of the tunnel, calculates the deformation rate distribution characteristics at different time scales (hourly, daily, and weekly), and uses quantile statistics to determine the initial warning threshold range.

[0181] in, For the first The initial threshold for Level 1 warning, The first of the historical deformation rate data Quantiles This is historical deformation data.

[0182] The simulation results are integrated into the process based on factors such as real tunnel geometry, rock strata parameters and load conditions to simulate the mechanical response of the tunnel under different degrees of deformation. The critical deformation value of the surrounding rock from elastic deformation to plastic instability is recorded during the simulation process, providing a theoretical basis for the early warning threshold.

[0183] An expert knowledge base was established, containing empirical thresholds and criteria under different coal mine geological conditions. The Delphi method was used to aggregate opinions from multiple experts to form an expert consensus threshold.

[0184] in, The first expert consensus Level 1 warning threshold The number of participating experts, For the first The weight of each expert For the first The expert gave the first Level 1 warning recommendation value.

[0185] Initial comprehensive threshold:

[0186] in, For the first The warning threshold for final fusion at the level, , , The fusion weighting coefficients for historical data. For the first The initial threshold for level-one early warning based on statistical results from historical data. The simulation results are as follows: A critical deformation index. The first expert consensus Level 1 warning threshold.

[0187] As the system operates and data accumulates, the thresholds are automatically adjusted based on the effectiveness of the early warning system.

[0188] in, For the current time step, For the first Level threshold at The adjustment ratio at different times.

[0189] The relationship between the threshold adjustment amount and the threshold adjustment amount is as follows:

[0190] in, The sensitivity coefficient for threshold adjustment. For the first The number of false alarms in the level-one warning For the first The number of missed reports in the level-one early warning system. For the first Total number of alarms at the Level 1 warning level.

[0191] Prediction results output and iterative optimization.

[0192] To ensure the system continuously improves prediction accuracy and early warning reliability, this invention establishes a complete prediction, verification, and optimization loop mechanism.

[0193] The prediction and verification cycle forms a closed-loop verification mechanism:

[0194] in, To comprehensively assess the error, , These are the weighting coefficients for short-term and long-term evaluations. , The number of sampling points for short-term and long-term assessments, respectively. , Assign higher weights to recent data in the time-weighted function. , These are the predicted value and the actual observed value, respectively.

[0195] Prediction result output expression:

[0196] in, Input features In prediction depth The following prediction results These are the prediction functions for the transition memory predictor model and the spiking neural network model, respectively. , For the parameter sets of their respective models, and To predict depth-related dynamic fusion weights.

[0197] The fusion weighting function can dynamically adjust the contribution weights of each sub-model based on the prediction depth:

[0198]

[0199] in, , These are control parameters.

[0200] According to a second aspect of the present invention, a roadway surrounding rock intelligent prediction system based on point cloud fusion includes: The parameter database module is configured to acquire basic data and deformation monitoring data of the surrounding rock of the roadway, and to build a data acquisition and monitoring mechanism to process the basic data and deformation monitoring data to obtain a multi-parameter monitoring database. The boundary constraint module is configured to construct a point cloud data processing model based on a multi-parameter monitoring database, establish equations for roadway main axis identification and path extraction, boundary condition correction models, and comprehensive deformation evaluation models, and establish unified boundary conditions for multi-source data to analyze the point cloud data processing model, roadway main axis identification and path extraction equations, boundary condition correction models, and comprehensive deformation evaluation models to obtain multi-source boundary constraint conditions. The loss function module is configured to design a dual-model network architecture, obtain a dual-model constraint loss function based on the dual-model network architecture and multi-source boundary constraints, perform fitting analysis on the dual-model constraint loss function, and obtain a comprehensive constraint loss function of deformation trend-sudden event index. The architecture and model building module is configured to obtain the temporal input feature vector based on the initial deformation state encoding vector, deformation history feature vector, and environmental process feature vector, and to build a transformation memory prediction-pulse collaborative neural network architecture and deformation evolution analysis model based on the temporal input feature vector; The prediction and monitoring module is configured to combine a multi-parameter monitoring database, obtain prediction results based on a comprehensive constraint loss function, a transformation memory prediction-pulse collaborative neural network architecture, and a deformation evolution analysis model, and perform closed-loop verification and optimization of the prediction results by dynamically adjusting the early warning threshold, thereby realizing the prediction and monitoring of instantaneous deformation and long-term trends of the surrounding rock in the roadway.

[0201] This prediction system, through the scientific and rational arrangement of three-dimensional laser scanners, target networks, and traditional monitoring devices, and the establishment of an efficient data fusion mechanism, can comprehensively acquire the surrounding rock deformation data, mechanical parameters, and environmental condition data required by the transformation memory predictor and the pulse collaborative neural network algorithm.

[0202] The prediction system deploys displacement sensors at key locations in the roadway to monitor changes in surrounding rock deformation in real time; stress sensors are deployed at typical cross-sections to accurately monitor the stress state of the surrounding rock; and crack monitoring equipment is deployed in deformation-sensitive areas to track crack development dynamics. Simultaneously, a mechanism for optimizing the layout of monitoring points and evaluating data quality is established to ensure the effectiveness and reliability of the monitoring data.

[0203] This prediction system uses 3D laser scanning technology to achieve continuous monitoring of the entire cross-section of the surrounding rock in the roadway; it uses point cloud data processing and feature extraction to identify the main axis of the roadway and extract the time series of roof displacement; it constructs a dual-model collaborative mechanism of conversion memory predictor and spiking neural network to capture long-term deformation trends and sudden event characteristics respectively; and it establishes scientific monitoring parameter critical values ​​and early warning values ​​based on numerical simulation technology and historical data analysis to predict the deformation trend of the surrounding rock.

[0204] This prediction system achieves continuous full-section monitoring in complex tunnel environments by collaboratively deploying a 3D laser scanner, target network, and traditional monitoring devices. The system relies on a point cloud processing module for real-time data processing, employing various filtering and feature extraction algorithms to achieve efficient tunnel main axis identification and cross-sectional analysis. Spatial geometric constraints are introduced to ensure data processing conforms to the actual tunnel morphology. A dual-model prediction mechanism optimizes the early warning strategy, and an early warning feedback module enables dynamic adjustment of monitoring parameter thresholds and visualized output of prediction results. The combined design of a transformation memory predictor and a spiking neural network ensures comprehensive capture of long-term deformation trends and sudden events, while adaptive threshold calculation provides a scientific basis for early warning.

[0205] This predictive system is flexible in layout and can be adjusted according to the progress of roadway excavation. It also has online learning and model updating capabilities. Practical application verification shows that this invention significantly improves monitoring accuracy and early warning accuracy, effectively preventing roof collapse accidents in thick coal seam roadways, and providing reliable technical support for safe coal mine production.

[0206] It should be noted that the intelligent prediction system for roadway surrounding rock based on point cloud fusion of the present invention can also perform any of the processing described in the previously described intelligent prediction method for roadway surrounding rock based on point cloud fusion, and the specific details are not repeated here.

[0207] The foregoing description, with reference to preferred embodiments, details the exemplary implementation of the intelligent prediction method and system for roadway surrounding rock based on point cloud fusion proposed in this invention. However, those skilled in the art will understand that various modifications and alterations can be made to the above specific embodiments without departing from the concept of this invention, and various combinations can be made to the various technical features and structures proposed in this invention without exceeding the protection scope of this invention, which is determined by the appended claims.

Claims

1. A method for intelligent prediction of roadway surrounding rock based on point cloud fusion, characterized in that, Includes the following steps: S10: Acquire basic data and deformation monitoring data of the surrounding rock of the roadway, and construct a data acquisition and monitoring mechanism to process the basic data and deformation monitoring data to obtain a multi-parameter monitoring database; S20: Based on the multi-parameter monitoring database, a point cloud data processing model is constructed, a roadway main axis identification and path extraction model, a boundary condition correction model, and a comprehensive deformation evaluation model are established. Multi-source data unified boundary conditions are established to analyze the point cloud data processing model, the roadway main axis identification and path extraction equation, the boundary condition correction model, and the comprehensive deformation evaluation model, and obtain multi-source boundary constraint conditions. S30: Design a dual-model network architecture, obtain the dual-model constraint loss function based on the dual-model network architecture and multi-source boundary constraints, perform fitting analysis on the dual-model constraint loss function, and obtain the comprehensive constraint loss function of deformation trend-sudden event index. S40: Based on the initial deformation state encoding vector, deformation history feature vector, and environmental process feature vector, a time-series input feature vector is obtained. A transformation memory prediction-pulse collaborative neural network architecture and a deformation evolution analysis model are then constructed based on this time-series input feature vector. Specifically, the construction of the transformation memory prediction-pulse collaborative neural network architecture and the deformation evolution analysis model based on the time-series input feature vector includes the following steps: S401: Establish a sequence length determination model based on the time-series input feature vector, and determine the optimal input sequence length through the sequence length determination model; S402: A transformation memory prediction-spiking collaborative neural network architecture is constructed based on the time scale of deformation evolution features and the optimal input sequence length. The transformation memory prediction-spiking collaborative neural network architecture includes a transformation memory predictor and a spiking neural network. The transformation memory predictor captures long-term deformation trends through multi-layer memory units, multi-head attention mechanisms, and multi-scale time window fusion. The spiking neural network uses an improved leakage integral discharge neuron model combined with a dynamic threshold mechanism to identify sudden deformation events. The two networks work together through a dynamic feature gating fusion framework. The dynamic feature gating fusion framework satisfies the following relationship: In the formula, For the final prediction result, To transform the prediction output of the memory predictor model, This is the predicted output of the spiking cooperative neural network model. For fusion function, This is time-varying context information; S403: Based on the time-series input feature vector, calculate the deformation state encoding vector, deformation propagation velocity vector, and key time nodes respectively, and fuse the deformation state encoding vector, deformation propagation velocity vector, and key time nodes to obtain the deformation evolution analysis model; S50: Combining a multi-parameter monitoring database, prediction results are obtained based on a comprehensive constraint loss function, a transformation memory prediction-pulse collaborative neural network architecture, and a deformation evolution analysis model. The prediction results are then dynamically adjusted through an early warning threshold to perform closed-loop verification and optimization, thereby enabling the prediction and monitoring of instantaneous deformation and long-term trends of the surrounding rock in the roadway.

2. The intelligent prediction method for roadway surrounding rock based on point cloud fusion according to claim 1, characterized in that, In step S10, a data acquisition and monitoring model is constructed to process the basic data and deformation monitoring data to obtain a multi-parameter monitoring database, including the following steps: S11: Determine the scope of the calculation domain based on the tunnel model, and set the coordinate system spatiotemporal calculation domain and the boundary conditions of ground stress and deformation of the tunnel according to the scope of the calculation domain. S12: Determine the sampling point density weights based on the spatiotemporal computational domain and the boundary conditions of geostress and deformation; S13: Establish a sampling encryption algorithm and coordinate normalization mapping function for deformation-sensitive areas; S14: The basic data and deformation monitoring data are adjusted by combining the computational domain range, sampling point density weight, deformation-sensitive area sampling densification algorithm and coordinate standardization mapping function to obtain the adjusted basic data and deformation monitoring data. S15: Use a data acquisition and monitoring model to fuse the adjusted basic data and deformation monitoring data to obtain a multi-parameter monitoring database.

3. The intelligent prediction method for roadway surrounding rock based on point cloud fusion according to claim 1, characterized in that, Step S20 specifically includes the following steps: S21: Establish a point cloud data processing model based on statistical outlier filtering and curvature-preserving downsampling; S22: Establish equations for identifying the main axis of the roadway and extracting the center path, and construct a boundary condition correction model based on deformation influencing factors; wherein, the equations for identifying the main axis of the roadway and extracting the center path include: the formula for identifying the main axis of the roadway and the formula for extracting the center path; S23: Establish a comprehensive deformation evaluation model based on the experimental information of the surrounding rock deformation in the tunnel and the time-accumulated deformation model; S24: By coupling multi-source data and unifying boundary conditions, the parameter boundary conditions of the point cloud data processing model, the roadway main axis identification and path extraction equation and boundary condition correction model and the comprehensive deformation evaluation model are analyzed to obtain deformation boundary conditions, trend boundary conditions and sudden boundary conditions. S25: Combining deformation boundary conditions, trend boundary conditions, and sudden boundary conditions yields multi-source boundary constraint conditions.

4. The intelligent prediction method for roadway surrounding rock based on point cloud fusion according to claim 3, characterized in that, In step S22, the formula for identifying the main shaft of the roadway is: in, The main axis direction vector For the point cloud dataset, the first Each point is a three-dimensional vector. This represents the total number of points in the point cloud. For the centroid of the point cloud, For vectors Transpose of; The center path is fitted using RANSAC to the roadway centerline. The formula for the distance from a point to a line is: in, This represents the straight line that best fits the point cloud data. For the shape of The linear model, Let be the distance from the point to the line. The threshold value is used.

5. The intelligent prediction method for roadway surrounding rock based on point cloud fusion according to claim 3, characterized in that, In step S24, the deformation boundary conditions, trend boundary conditions, and sudden boundary conditions are respectively: The deformation boundary conditions satisfy the following relationship: In the formula, For the surrounding rock point p during the monitoring time The amount of deformation at that time; To monitor the surrounding rock point p at time t, and the reference outline of the roadway at time t. The minimum distance; Based on the baseline standard outline and the surrounding rock monitoring points Feature matching corresponds to the pairing points; This serves as the standard reference outline for the roadway at the base time. Based on the baseline standard outline and the surrounding rock monitoring points Feature matching corresponding pairing points Standard reference profile of the roadway at the reference time The minimum distance; The trend boundary conditions satisfy the following relationship: In the formula, The current cell state is used to control the flow of information; For the forget gate, multiply by To determine how much information is forgotten; This represents the cell state at the previous moment; The output of the input gate determines the update of the cell state at the current moment; Representing the candidate cell state, it is the alternative feature information to be included in the memory unit at the current time step; This is element-wise multiplication; Sudden boundary conditions satisfy the following relationship: In the formula, The membrane potential time constant is used to control the potential recovery rate; Current time Membrane potential at time t; This is the resting potential, which is also the potential of a neuron when it is not stimulated. For membrane resistance; For synaptic input current; For adaptive current weighting coefficients; For adaptive current.

6. The intelligent prediction method for roadway surrounding rock based on point cloud fusion according to claim 1, characterized in that, Step S30 specifically includes the following steps: S31: Set up a multi-dimensional parameter vector input mechanism and an input parameter standardization processing mechanism in the network architecture; S32: Based on the multi-dimensional parameter vector input mechanism, the input parameter standardization processing mechanism, and the multi-source boundary constraints, the output results of the trend prediction layer, the sudden event detection layer, and the fusion prediction layer are obtained; S33: Construct a dual-model constrained loss function based on the output results of the trend prediction layer, the sudden event detection layer, and the fusion prediction layer; S34: Perform fitting analysis on the dual-model constraint loss function to obtain the comprehensive constraint loss function of deformation trend-emergency index; The comprehensive constraint loss function satisfies the following relationship: In the formula, The current parameters are used to measure the overall loss function of the model. The overall prediction error, The mean squared error loss penalizes the squared error between the predicted and actual values. To weight safety risks, higher weights are given to errors in high-risk areas. The weighting coefficients are used to control the balance between ordinary mean square error and risk-weighted loss.

7. The intelligent prediction method for roadway surrounding rock based on point cloud fusion according to claim 1, characterized in that, In step S40, the temporal input feature vector is obtained based on the initial deformation state encoding vector, the deformation history feature vector, and the environmental process feature vector. This specifically includes the following steps: S41: Based on the network architecture and multi-parameter monitoring database, obtain the spatial deformation distribution characteristics, and obtain the initial deformation state encoding vector according to the spatial deformation distribution characteristics; S42: Construct characteristic expressions for the deformation history and stress process of the surrounding rock, and use the characteristic expressions to obtain the deformation history feature vector; S43: Set up a coding mechanism for environmental conditions and monitoring cycles, and obtain the environmental process feature vector through the coding mechanism; S44: Combine time-varying environmental conditions, initial deformation state encoding vector, deformation history feature vector, and environmental process feature vector to obtain the time-series input feature vector.

8. The intelligent prediction method for roadway surrounding rock based on point cloud fusion according to claim 1, characterized in that, Step S50 specifically includes the following steps: S51: Combining historical data, simulation data, and expert knowledge from the multi-parameter monitoring database, historical data analysis, simulation results integration, and expert experience introduction are performed sequentially to obtain the initial comprehensive early warning threshold; then, an adaptive threshold calculation strategy is adopted to dynamically adjust the initial comprehensive early warning threshold based on the early warning feedback effect during operation; The initial comprehensive early warning threshold satisfies the following relationship: In the formula, For the first The warning threshold for final fusion at the level, , , The fusion weighting coefficients for historical data. For the first The initial threshold for level-one early warning based on statistical results from historical data. The simulation results are as follows: A critical deformation index. The first expert consensus Level 1 warning threshold; The adaptive threshold calculation strategy involves iteratively updating the threshold based on the initial comprehensive early warning threshold and the actual early warning effect, satisfying the following relationship: in, The sensitivity coefficient for threshold adjustment. For the first The number of false alarms at the level of warning For the first The number of missed reports in the level-one warning, For the first Total number of alarms at the Level 1 warning level; S52: Based on the obtained early warning threshold, establish a prediction, verification and optimization loop mechanism to output, evaluate and optimize the prediction results. In the prediction phase, the prediction result is obtained based on the fusion output of the transition memory predictor model and the spiking neural network model, satisfying the following relationship: In the formula, Input features In prediction depth The following prediction results These are the prediction functions for the transition memory predictor model and the spiking neural network model, respectively. , For the parameter sets of their respective models, and To predict depth-related dynamic fusion weights; The verification phase compares the predicted results with the actual observed values, constructs a closed-loop verification mechanism, and evaluates the current prediction accuracy, satisfying the following relationship: In the formula, To comprehensively assess the error, , Here, N and M are the weighting coefficients for short-term and long-term evaluations, respectively, and represent the number of sampling points for short-term and long-term evaluations. , Assign higher weights to recent data in the time-weighted function. , These are the predicted value and the actual observed value, respectively.

9. A roadway surrounding rock intelligent prediction system based on point cloud fusion, characterized in that, The method for implementing the intelligent prediction method for roadway surrounding rock based on point cloud fusion as described in any one of claims 1-8 includes: The parameter database module is configured to acquire basic data and deformation monitoring data of the surrounding rock of the roadway, and to build a data acquisition and monitoring mechanism to process the basic data and deformation monitoring data to obtain a multi-parameter monitoring database. The boundary constraint module is configured to construct a point cloud data processing model based on a multi-parameter monitoring database, establish equations for roadway main axis identification and path extraction, boundary condition correction models, and comprehensive deformation evaluation models, and establish unified boundary conditions for multi-source data to analyze the point cloud data processing model, roadway main axis identification and path extraction equations, boundary condition correction models, and comprehensive deformation evaluation models to obtain multi-source boundary constraint conditions. The loss function module is configured to design a dual-model network architecture, obtain a dual-model constraint loss function based on the dual-model network architecture and multi-source boundary constraints, perform fitting analysis on the dual-model constraint loss function, and obtain a comprehensive constraint loss function of deformation trend-sudden event index. The architecture and model building module is configured to obtain the temporal input feature vector based on the initial deformation state encoding vector, deformation history feature vector, and environmental process feature vector, and to build a transformation memory prediction-pulse collaborative neural network architecture and deformation evolution analysis model based on the temporal input feature vector; The prediction and monitoring module is configured to combine a multi-parameter monitoring database, obtain prediction results based on a comprehensive constraint loss function, a transformation memory prediction-pulse collaborative neural network architecture, and a deformation evolution analysis model, and perform closed-loop verification and optimization of the prediction results by dynamically adjusting the early warning threshold, thereby realizing the prediction and monitoring of instantaneous deformation and long-term trends of the surrounding rock in the roadway.

Images