Coal mine tunnel laser radar three-dimensional dynamic monitoring method

Abstract

The invention relates to the technical field of coal mine safety monitoring, and discloses a coal mine tunnel laser radar three-dimensional dynamic monitoring method. Aiming at the problems that in the prior art, data of a laser radar is easy to interrupt or unreliable under a severe working condition, and monitoring continuity is poor due to insufficient multi-source data fusion, the method is characterized in that laser radar surface data and optical strain gauge point data are synchronously acquired; constructing a roadway digital twinborn model; carrying out fusion and assimilation on the multi-source data by utilizing a Transform mechanism to obtain initial roadway deformation data; judging the credibility of the laser radar data based on Bayesian estimation; when the data are missing or uncredible, deducing full-field deformation by utilizing a gating mechanism in combination with optical strain gauge point data and a digital twin model; and finally, generating and outputting a three-dimensional dynamic deformation field. The system is mainly used for real-time, continuous and high-reliability deformation monitoring and safety early warning of the coal mine tunnel.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of optical instrument monitoring roadway deformation, and particularly relates to a coal mine roadway laser radar three-dimensional dynamic monitoring method. BACKGROUND

[0002] In the field of coal mine safety production, roadway deformation monitoring is a key technical means to prevent roof collapse, surrounding rock instability and other accidents. With the development of optical monitoring technology, laser radar (LiDAR) has become an important tool for roadway deformation monitoring because it can non-contact and efficiently obtain high-density three-dimensional point cloud data of the roadway surface. Laser radar can reconstruct the three-dimensional geometry of the roadway by emitting laser beams and receiving reflected signals, realizing dynamic acquisition of full-field deformation information. In addition, optical strain gauges based on fiber Bragg grating (FBG) technology can accurately measure the strain changes of local key points, have advantages such as anti-electromagnetic interference and good stability, and are suitable for coal mine harsh environments. However, the existing monitoring system has deficiencies in the collaborative use of multi-source data, and it is difficult to ensure the continuity and reliability of the data in complex working conditions.

[0003] At present, the common roadway deformation monitoring scheme mainly relies on a single laser radar system or a simple combination of multiple sensors. For example, some existing technologies use laser radar to periodically scan the roadway and extract deformation information through point cloud data processing algorithms. This kind of scheme can provide relatively accurate full-field data in good environmental conditions, but in extreme conditions underground coal mines (such as high concentration of dust, water vapor interference, equipment failure or severe deformation of surrounding rock), laser radar data is prone to interruption, increased noise or significant decrease in reliability, resulting in monitoring blind spots or false positives. In addition, a single laser radar system lacks verification and supplementation of local high-precision data, making it difficult to provide reliable deformation evaluation in key areas such as stress concentration points.

[0004] Some other existing technologies attempt to use laser radar in combination with other sensors such as optical strain gauges, but usually only perform data parallel acquisition and independent display without constructing a deep multi-source data fusion and complementary mechanism. For example, optical strain gauges can provide high-precision strain data at sparse points, but their spatial coverage is limited and cannot directly reflect the full-field deformation distribution; while laser radar data covers the full field, but the local precision may be insufficient when affected by environmental interference. Existing technologies have not effectively solved the problems of time sequence alignment, data correlation, weight distribution and data assimilation under physical constraints between laser radar and optical strain gauges, especially when laser radar data is missing or unreliable, lacking the ability to deduce full-field deformation based on optical strain gauge points and roadway mechanical models, resulting in insufficient data integrity of the monitoring system in extreme working conditions, affecting the continuity of safety decisions.

[0005] The present application inventors find that the above technology has the following technical problem: how to construct a multi-source data fusion and complementary mechanism that fuses laser radar and other sparse optical strain gauges in the roadway, to ensure the integrity and high reliability of deformation monitoring data under any working condition. SUMMARY

[0006] The present application provides a coal mine roadway laser radar three-dimensional dynamic monitoring method, and aims to solve the technical problem of how to construct a multi-source data fusion and complementary mechanism that fuses laser radar and other sparse optical strain gauges in the roadway, to ensure the integrity and high reliability of deformation monitoring data under any working condition.

[0007] In order to achieve the above application purposes, the technical scheme adopted by the present application is as follows:

[0008] A coal mine roadway laser radar three-dimensional dynamic monitoring method, the method comprising: step S10, synchronously acquiring laser radar surface data collected by a laser radar arranged in a coal mine roadway and optical strain gauge point data collected by a plurality of optical strain gauges arranged at key points of the coal mine roadway; step S20, constructing a roadway digital twin model based on geometric parameters and mechanical parameters of the coal mine roadway, wherein the roadway digital twin model comprises a roadway geometric model and a roadway mechanical model; step S30, based on the roadway digital twin model, using a Transformer mechanism to perform multi-source data fusion and data assimilation processing on the laser radar surface data and the optical strain gauge point data, to obtain initial roadway deformation data; step S40, judging the reliability of the laser radar surface data based on Bayesian estimation theory; step S50, when the laser radar surface data is missing or the reliability value is lower than a preset threshold, using a gating mechanism to deduce the full-field deformation of the coal mine roadway based on the optical strain gauge point data and the roadway digital twin model, to obtain deduced roadway deformation data; step S60, when the laser radar surface data is not missing and the reliability value is not lower than the preset threshold, taking the initial roadway deformation data as output data; step S70, based on the output data or the deduced roadway deformation data, generating and outputting a three-dimensional dynamic deformation field of the coal mine roadway, wherein the three-dimensional dynamic deformation field represents the continuous deformation state of the coal mine roadway over time.

[0009] The present application has the following beneficial effects:

[0010] 1. The method synchronously collects laser radar surface data and optical strain gauge point data, constructs a digital twin model of the roadway, and uses the Transformer mechanism for multi-source data fusion and data assimilation. The reliability of the laser radar data is dynamically evaluated using Bayesian estimation. When the data is missing or unreliable, the gate mechanism is used to deduce the full-field deformation, thereby generating a three-dimensional dynamic deformation field. This realizes continuous, complete and high reliability of deformation monitoring data under any working condition, significantly improving the monitoring accuracy and decision support capability of coal mine safety production.

[0011] 2. The integrity and continuity of the monitoring data are improved: by constructing a digital twin model of the roadway that integrates geometric and mechanical properties, and using the Transformer mechanism for deep fusion of laser radar surface data and optical strain gauge point data and data assimilation under physical constraints, the invention overcomes the limitations of a single data source. Especially when the laser radar data is missing or unreliable, the full-field deformation can be intelligently deduced based on limited optical strain gauge point data and roadway mechanical models using the gate mechanism, thereby ensuring the integrity and continuity of the monitoring data under any working condition, effectively avoiding monitoring blind spots.

[0012] 3. The reliability and accuracy of the monitoring results are enhanced: the invention innovatively applies Bayesian estimation theory, using reliable optical strain gauge point data as prior information to dynamically evaluate the reliability of laser radar surface data. This mechanism can automatically identify and quantify data uncertainty caused by environmental interference and other factors, providing a scientific basis for subsequent data processing strategies (direct use or start deduction), thereby significantly improving the reliability and accuracy of the final output three-dimensional dynamic deformation field.

[0013] 4. An intelligent and adaptive monitoring process is achieved: the technical solution of the invention constitutes a closed-loop intelligent monitoring system. It not only adaptively switches working modes when data is abnormal, but also regularly updates the mechanical parameters of the digital twin model based on historical monitoring data using the Transformer and gate mechanisms, allowing the model to continuously evolve and better fit the actual long-term deformation behavior of the roadway. This self-learning and adaptive capability enables the monitoring system to maintain high accuracy over a long period, improving the intelligence level and long-term practicality of the system. BRIEF DESCRIPTION OF DRAWINGS

[0014] Figure 1 is a structural schematic diagram of the invention. DETAILED DESCRIPTION

[0015] The specific embodiments of the invention will be further described below in conjunction with the drawings. The same parts are denoted by the same reference numerals.

[0016] It should be noted that the words "front", "back", "left", "right", "up" and "down" used in the following description refer to the directions in the drawings, and the words "in" and "out" refer to the directions towards or away from the geometric center of a particular component.

[0017] In order to make the content of the present application easier to be clearly understood, the technical solutions in the embodiments of the present application will be described clearly and completely in conjunction with the drawings in the embodiments of the present application.

[0018] Invention concept: In order to solve the technical problems of insufficient multi-source data fusion in coal mine roadway deformation monitoring, easy interruption or unreliability of laser radar data in harsh environment, and low completeness and reliability of monitoring data, a laser radar three-dimensional dynamic monitoring method is proposed. The method synchronously collects laser radar surface data and optical strain gauge point data, constructs a roadway digital twin model, and uses the Transformer mechanism for multi-source data fusion and data assimilation, combines with Bayesian estimation to dynamically evaluate the reliability of laser radar data, uses the gating mechanism to deduce the full-field deformation when the data is missing or unreliable, and thus generates a three-dimensional dynamic deformation field. This realizes the continuity, completeness and high reliability of deformation monitoring data under any working condition, significantly improves the monitoring accuracy and decision support capability of coal mine safety production.

[0019] A coal mine roadway laser radar three-dimensional dynamic monitoring method, the method comprises:

[0020] Step S10, synchronously acquiring laser radar surface data collected by a laser radar arranged in a coal mine tunnel and optical strain gauge point data collected by a plurality of optical strain gauges arranged at key points of the coal mine tunnel, wherein the laser radar surface data represents full-field deformation information of the coal mine tunnel, and the optical strain gauge point data represents local deformation information of the key points of the coal mine tunnel; this step is the data acquisition basis of the whole monitoring system. The laser radar acquires the spatial coordinates of a large number of points on the tunnel surface by emitting a laser beam and receiving a reflected signal, forms dense point cloud data, i.e. "laser radar surface data", and can comprehensively reflect the deformation of the whole surface of the tunnel. The optical strain gauge is a high-precision local deformation measuring device that monitors the small strain changes at specific positions through fiber Bragg grating technology and forms "optical strain gauge point data". The complementarity of the two data sources lies in that the laser radar provides full-field data that may be disturbed by the environment, and the optical strain gauge provides local but extremely stable and reliable data. In the complex environment of a coal mine, this multi-source data acquisition can effectively overcome the limitations of a single data source, especially in harsh conditions such as dust and water vapor to ensure the continuity of monitoring. Step S20, constructing a tunnel digital twin model based on the geometric parameters and mechanical parameters of the coal mine tunnel, wherein the tunnel digital twin model includes a tunnel geometric model and a tunnel mechanical model, the tunnel geometric model describes the three-dimensional spatial structure of the coal mine tunnel, and the tunnel mechanical model describes the deformation behavior of the coal mine tunnel under stress conditions; the digital twin model is one of the core innovations of the present application, which is not a simple three-dimensional visualization model, but a dynamic simulation system that integrates physical laws. The geometric model accurately reproduces the spatial form of the tunnel through CAD technology, and the mechanical model describes the deformation law of the tunnel under the action of factors such as surrounding rock pressure and underground water based on the principle of rock mass mechanics. The key parameters that need to be considered in constructing this model include the geometric shape of the tunnel (such as cross-sectional size and curvature) and the physical properties of the rock mass (such as elastic modulus and Poisson's ratio). The digital twin model acts like a "virtual laboratory" that can predict the deformation behavior of the tunnel under different working conditions and compare and verify with the actual measured data. Step S30, based on the tunnel digital twin model, using the Transformer mechanism to perform multi-source data fusion and data assimilation processing on the laser radar surface data and the optical strain gauge point data to obtain initial tunnel deformation data, wherein the Transformer mechanism includes converting the laser radar surface data and the optical strain gauge point data into time series data, calculating the correlation weight between the multi-source data through the self-attention mechanism, and performing weighted fusion based on the correlation weight, and the data assimilation processing includes physically constrained optimization of the fused data using the tunnel mechanical model; this step is the core innovation of the present application, which realizes the intelligent fusion of multi-source data. The Transformer mechanism originates from the natural language processing field and is innovatively applied to engineering monitoring data fusion.The principle is to regard data from different sources as "language sequences" and automatically learn the correlation between data points through self-attention mechanism. For example, when the laser radar shows that a certain area has large deformation, the system will automatically focus on the optical strain gauge data near the area and give it a higher weight. Data assimilation is to compare the fusion result with the prediction of the digital twin model and correct possible measurement errors using physical laws. This process can be represented by the following formula:

[0021]

[0022]

[0023] wherein, is the fused deformation value, is the laser radar data, is the optical strain gauge data, is the adaptive weight coefficient, is the weight matrix, and are the feature vectors of laser radar and strain gauge data, respectively.

[0024] Step S40, based on the Bayesian estimation theory, judges the reliability of the laser radar surface data, wherein the Bayesian estimation theory takes the optical strain gauge point data as prior information, calculates the posterior probability distribution of the laser radar surface data, and generates the reliability value of the laser radar surface data based on the posterior probability distribution; this step innovatively applies the Bayesian statistical method to sensor data reliability evaluation. The principle is that the optical strain gauge data, due to its fixed installation position and small environmental interference, can be used as a "reliable reference" (prior information); while the laser radar data may be affected by dust, water vapor and other factors to produce errors (likelihood function). Through the Bayesian formula, the reliability of the laser radar data can be calculated after considering the optical strain gauge reference (posterior probability). The specific formula is:

[0025]

[0026] wherein, is the posterior probability distribution of the laser radar data, is the likelihood function, is the prior probability distribution, is the evidence factor.

[0027] The reliability value is usually the inverse of the variance of the posterior distribution, and the smaller the variance, the more reliable the data. The advantage of this method is that it can dynamically evaluate the quality of laser radar data and provide a basis for subsequent data missing processing.

[0028] Step S50, when the laser radar surface data is missing or the reliability value is lower than the preset threshold, based on the optical strain gauge point data and the roadway digital twin model, using a gating mechanism to deduce the full-field deformation of the coal mine roadway, and obtaining deduced roadway deformation data, wherein the gating mechanism includes an input gate, a forget gate and an output gate, which are used to control the information flow of the optical strain gauge point data and the roadway mechanical model in the deduction process, and the hidden state is updated by iteration to simulate the deformation propagation; this step is another core innovation point of the present application, which solves the industry problem of data interruption under extreme conditions. When the laser radar data is unreliable or missing, the system will not simply stop working, but use the point data of the optical strain gauge and the digital twin model to deduce the full-field deformation. The gating mechanism is derived from the gated recurrent unit (GRU) neural network, which contains three key components: the input gate determines the degree of new information addition, the forget gate controls the proportion of historical information retention, and the output gate adjusts the final output. These gating units work together to simulate the spatial propagation law of roadway deformation.

[0029] In this way, the system can reasonably deduce the deformation state of the entire roadway based on limited point data combined with rock mass mechanical laws.

[0030] Step S60, when the laser radar surface data is not missing and the credibility value is not lower than the preset threshold, taking the initial roadway deformation data as output data; this step defines the normal working mode of the system. When the quality and integrity of the laser radar data meet the requirements (i.e. the credibility value is higher than the preset threshold, usually set to 0.6-0.8), the system directly uses the initial roadway deformation data obtained by fusion in step S30 as the final output. This design ensures the use of the highest precision original data under good conditions, avoiding unnecessary information loss caused by data processing. The threshold setting needs to balance sensitivity and stability: a too high threshold will lead to frequent switching to the deduction mode, increasing the computational burden; a too low threshold may miss data quality problems. The present application realizes the scientificity and adaptability of threshold judgment through Bayesian credibility evaluation. Step S70, generating and outputting a three-dimensional dynamic deformation field of the coal mine roadway based on the output data or the deduced roadway deformation data, wherein the three-dimensional dynamic deformation field represents the continuous deformation state of the coal mine roadway over time. This step converts the processed data into an intuitive visual result. The three-dimensional dynamic deformation field not only shows the current deformation state, but also contains the historical deformation trend, forming a continuous monitoring view in space and time. In specific implementation, the system converts discrete point data or deduced data into continuous gridded deformation field through interpolation algorithm, and then applies visualization techniques such as color mapping and vector arrow to intuitively display deformation size and direction. The deformation field is usually updated at a speed of 1-5 frames per second, which ensures real-time performance and avoids data overload. The key innovation is to closely integrate the physical model with visualization, making the deformation field not only beautiful but also consistent with the laws of rock mass mechanics, avoiding unreasonable deformation patterns that may occur in traditional visualization. Further, in step S10, the synchronous acquisition includes:

[0031] Step S101, continuously scanning the coal mine tunnel by the laser radar to generate the laser radar surface data, wherein the laser radar surface data includes point cloud data and its timestamp; this sub-step details the specific process of laser radar data collection. The laser radar emits laser beams by rotating the scanning head, measures the round-trip time of the laser to calculate the distance, combines the angle information to determine the spatial coordinates, and forms the point cloud data. Each point contains (x, y, z) three-dimensional coordinates and reflection intensity information, and the timestamp ensures the timing accuracy of the data. In the coal mine environment, the laser radar is usually installed above the tunnel axis, and the scanning frequency is set to 0.1-5Hz to balance the data quality and system load. The point cloud data density directly affects the deformation monitoring accuracy, and the present application optimizes the scanning parameters to obtain sufficient density of point cloud (usually 100-1000 points per square meter) under the premise of ensuring real-time. Step S102, synchronously collecting strain signals of the key points by the plurality of optical strain gauges to generate the optical strain gauge point data, wherein the optical strain gauge point data includes strain value and its position coordinates and timestamp; this sub-step details the optical strain gauge data collection process. The optical strain gauge is based on the fiber Bragg grating (FBG) technology, when the optical fiber is subjected to strain, the grating period changes, which causes the reflection wavelength to shift, and the strain value can be accurately calculated by measuring the wavelength change. Compared with the traditional electrical strain gauge, the optical strain gauge has the advantages of anti-electromagnetic interference, corrosion resistance, long-distance transmission, etc., and is particularly suitable for coal mine harsh environment. The selection of key points needs to consider the structural weaknesses and stress concentration areas of the tunnel, which usually includes the top center, the middle of both sides of the wall, the bottom center and the corner, etc. Each strain gauge data includes strain value (unit: micro-strain με), accurate installation position coordinates and collection timestamp. Step S103, time alignment processing of the laser radar surface data and the optical strain gauge point data to ensure that the laser radar surface data and the optical strain gauge point data are under the same time reference. This sub-step solves the time synchronization problem of multi-source data. Since the laser radar and the optical strain gauge may use different clock systems and sampling frequencies, direct use will cause time misalignment. Time alignment processing usually adopts two methods: hardware synchronization (triggered by the same clock source) and software synchronization (aligned by interpolation algorithm). The present application preferentially adopts hardware synchronization to ensure high precision, and when hardware synchronization is not feasible, uses algorithms such as cubic spline interpolation for software synchronization to align data of different frequencies to a unified time grid. The time alignment accuracy directly affects the subsequent data fusion effect, and the present application controls the time error within ±10ms to meet the coal mine monitoring requirements. Further, in step S20, the construction of the tunnel digital twin model includes:

[0032] Step S201, obtain the design drawings and geological exploration data of the coal mine tunnel, and extract the geometric parameters, wherein the geometric parameters include the cross-sectional shape, size and spatial coordinates of the tunnel; this sub-step is the basic work of building a geometric model. The design drawings provide the theoretical shape and size of the tunnel, and the geological exploration data describes the distribution of the surrounding rock mass of the tunnel. The geometric parameters not only include the basic cross-sectional shape (such as rectangular, arched, horseshoe-shaped, etc.) and size (width, height, radius of curvature, etc.), but also include the spatial trend (such as inclination, curvature) and accurate coordinate position of the tunnel. These parameters are converted into accurate three-dimensional geometric models by CAD software as the basic framework of digital twinning. The coal mine tunnel is usually not a completely regular geometric body, and there may be construction errors and local deformations, so the modeling of initial deformation is also considered in the present application, so that the geometric model is closer to the actual situation. Step S202, determine the mechanical parameters based on the rock mass mechanical properties of the coal mine tunnel, wherein the mechanical parameters include the elastic modulus, Poisson's ratio and yield strength; this sub-step determines the key parameters of the mechanical model. The rock mass mechanical properties are usually obtained through field sampling and laboratory testing, including elastic modulus (indicating the ability of rock to resist deformation), Poisson's ratio (indicating the ratio of transverse deformation to longitudinal deformation), yield strength (indicating the stress value at which rock begins to deform plastically), etc. These parameters are crucial for accurate simulation of tunnel deformation, for example, the greater the elastic modulus, the smaller the deformation under the same stress; the greater the Poisson's ratio, the more obvious the transverse deformation. The present application also considers the heterogeneity and anisotropy of the rock mass, divides the tunnel area into multiple sub-areas, and assigns different mechanical parameters to each sub-area to improve model accuracy. Step S203, use the geometric parameters and the mechanical parameters to build the tunnel geometric model and the tunnel mechanical model by the finite element method, and integrate to form the tunnel digital twinning model. This sub-step completes the final construction of the digital twinning model. The finite element method discretizes the continuous tunnel structure into a large number of small elements (such as tetrahedrons or hexahedrons), applies mechanical equations on each element, and then solves the deformation behavior of the entire system. The geometric model defines the spatial position and connection relationship of the elements, and the mechanical model provides the material properties and mechanical equations of the elements. After integration, the digital twinning model can simulate the deformation response of the tunnel under various loads (such as surrounding rock pressure, groundwater pressure). The innovation of the present application is that the digital twinning model is closely combined with real-time monitoring data, so that it is not only a static simulation tool, but also can be continuously updated and optimized according to actual monitoring data.

[0033] Further, in step S30, the multi-source data fusion and data assimilation processing using the Transformer mechanism includes:

[0034] Step S301, converting the laser radar surface data and the optical strain gauge point data into time sequence data respectively, wherein the time sequence data includes deformation values of multiple time steps; this sub-step completes data preparation before the Transformer processing. The time sequence data organizes the spatially distributed monitoring data in chronological order to form a "data matrix", wherein the rows represent time steps and the columns represent spatial positions. For the laser radar surface data, due to the large and irregular amount of point cloud data, it is necessary to perform gridding processing to discretize the continuous space into fixed grid points; for the optical strain gauge point data, it can be directly arranged in chronological order due to its fixed position. This conversion enables data of different sources and structures to have a unified time sequence representation, facilitating subsequent Transformer processing. The time sequence length is usually set to 10-100 time steps to balance the amount of historical information and computational complexity. Step S302, processing the time sequence data using a Transformer encoder, calculating the correlation weight between the laser radar surface data and the optical strain gauge point data through a self-attention mechanism, and generating fused deformation data based on the correlation weight; this sub-step describes the core processing process of the Transformer in detail. The self-attention mechanism is a key component of the Transformer, which calculates the correlation weight between data points, and the formula is as follows:

[0035]

[0036] wherein, is a query matrix, is a key matrix, is a value matrix, is the dimension of the key vector, and T represents the transpose operation of the matrix.

[0037] In the present application, Q, K, and V are obtained by linear transformation of the input data. The self-attention mechanism can automatically discover the correlation between data points at different positions and different times, for example, when the optical strain gauge in a certain area shows abnormal deformation, the system will automatically increase the weight of the laser radar data near that area. This correlation is not pre-set, but is automatically learned from the training data, and can adapt to different roadway conditions and deformation modes.

[0038] Step S303, performing data assimilation processing on the fused deformation data using the roadway mechanics model, optimizing the initial roadway deformation data by minimizing the error between the model prediction and the fused deformation data. This sub-step completes the data assimilation processing and combines the fusion result with the physical model. Data assimilation is a process of combining observed data with model predictions, and the present application uses a variational assimilation method to optimize the following objective function:

[0039]

[0040] where, is the state vector to be optimized, is the background field (model prediction), is the background error covariance, is the observation operator, is the observation data, is the observation error covariance, T denotes the transpose operation of a matrix.

[0041] By minimizing this objective function, the system finds the optimal solution that both satisfies the physical laws and is close to the actual observation. This processing ensures that the fusion result will not appear "pseudo deformation" that violates the laws of rock mechanics, such as spontaneous large deformation under no external force. Further, in step S40, the judging the reliability of the lidar surface data comprises:

[0042] Step S401, taking the optical strain gauge point data as the prior distribution, a likelihood function of the lidar surface data is constructed; this sub-step establishes the mathematical basis of Bayesian reliability evaluation. The prior distribution represents our initial understanding of the system state, and here the optical strain gauge data is used as the prior because of its high reliability but low spatial resolution. The likelihood function describes the probability of observing the current lidar data under a given true state, which is usually assumed to be a Gaussian distribution:

[0043]

[0044] where, is the lidar observation value, is the true state, is the observation operator, is the observation error standard deviation.

[0045] The variance of the likelihood function reflects the uncertainty of the lidar data, which is affected by environmental conditions (such as dust concentration). Through historical data analysis, the present invention establishes a relationship model between the environmental conditions and , so that the likelihood function can dynamically adapt to different working conditions. Step S402, based on the Bayes formula, the posterior probability distribution of the lidar surface data is calculated; this sub-step applies the Bayes formula to complete the core calculation of the reliability. The Bayes formula combines prior information and observation data to obtain the posterior distribution:

[0046]

[0047] In practical applications, the mean and variance of the posterior distribution are usually of interest. The mean represents the most likely true state, while the variance reflects the uncertainty. The present invention focuses on the posterior variance as a key indicator for reliability assessment. The smaller the posterior variance, the higher the consistency between lidar data and prior information, and the more reliable the data; otherwise, the reliability is low. Step S403, extracting a variance indicator from the posterior probability distribution as the reliability value, wherein the variance indicator represents the uncertainty of the lidar surface data. This sub-step completes the specific calculation of the reliability value. The posterior variance directly reflects the uncertainty of the data, but in order to facilitate use, the present invention converts it into a reliability value in the interval [0, 1]:

[0048]

[0049] wherein, is the posterior variance, is a scaling coefficient for adjusting the sensitivity of the reliability value.

[0050] This conversion ensures that the larger the variance, the lower the reliability, and the relationship is monotonic within a reasonable range. A reliability value close to 1 indicates that the data is very reliable, and a value close to 0 indicates that the data is almost not reliable. The system automatically decides whether to switch to the inference mode according to the reliability value, realizing intelligent management of the monitoring process. Further, in step S50, the using the gating mechanism to infer the full-field deformation of the coal mine roadway comprises:

[0051] Step S501, when the lidar surface data is missing, extracting the deformation of the key points from the optical strain gauge point data as input; this sub-step defines the trigger condition and initial input of the inference mode. When the lidar fails or is completely disabled due to extreme environment, the system only relies on the point data of the optical strain gauge to start the inference. The deformation of the key points serves as the "seed data" for inference, and the system assumes that the measurements of these points are accurate and reliable. Since the number of optical strain gauges is limited (usually 5-10), and the roadway surface contains thousands of points, this inference from sparse points to the full field is extremely challenging. The innovation of the present invention lies in combining physical models and data-driven methods, which not only constrain the inference process with rock mass mechanics laws, but also learn the actual deformation pattern through the gating mechanism, avoiding the problem of inaccurate parameters that may exist in pure physical models. Step S502, using a gated recurrent unit network with the deformation of the key points as the initial input, and through the input gate controlling the addition of new information, the forget gate controlling the retention of historical information, and the output gate controlling the output of hidden states, to perform full-field deformation inference; this sub-step describes the working principle of the gating mechanism in detail. The gated recurrent unit (GRU) is a special recurrent neural network, and its gating mechanism includes:

[0052] Input gate: determines the degree of influence of new observation data on the current state;

[0053] forget gate: control the retention ratio of historical information;

[0054] output gate: adjust the deformation field of final output;

[0055] These gating units are implemented by sigmoid functions, and the output values are between [0, 1], representing the "opening degree". For example, when the system detects that the deformation mode changes suddenly, the input gate will open more, allowing new information to quickly affect the deduction result; when the deformation mode is stable, the forget gate will retain more historical information, making the deduction result smoother. This mechanism enables the system to adapt to different deformation dynamics, neither being too sensitive to noise nor missing real deformation trends.

[0056] Step S503, based on the roadway mechanical model, the physical constraint correction of the deduction result is carried out, and the hidden state of the gating cycle unit network is updated through iteration, and the deduction roadway deformation data is output. This sub-step completes the physical rationality correction of the deduction result. Pure data-driven deduction may produce results that violate physical laws, such as spontaneous large deformation under no external force. The present application compares the deduction result with the prediction of the roadway mechanical model, calculates the residual error:

[0057]

[0058] In the formula, is the deduction result, is the mechanical model prediction, is the observation operator.

[0059] Then use gradient descent method to adjust the GRU network parameters to minimize the residual error. This process ensures that the deduction result not only conforms to the actual observation trend, but also meets the basic laws of rock mass mechanics, avoiding unreasonable "pseudo-deformation". Iterative updating of the hidden state enables the system to continuously learn the deformation mode, improving the accuracy of long-term deduction. Further, in step S503, the physical constraint correction includes:

[0060] Step S5031, calculate the residual error of the deduction result and the prediction value of the roadway mechanical model at the key points; this sub-step defines the specific calculation method of the physical constraint correction. The residual error calculation focuses on the key points where the optical strain gauges are installed, because these points have both deduction results and actual observation values, which can be directly compared. The residual error formula is:

[0061]

[0062] In the formula, is the residual error of the i-th key point, is the deduction result, is the mechanical model prediction, For the mechanical model prediction value.

[0063] The residual reflects the inconsistency between the deduction result and the physical model, and is the basis for correction. The selection of key points is crucial and should cover the main deformation modes of the roadway, such as roof subsidence and sidewall bulging. The invention also considers the spatial correlation of the residual, that is, the residuals of adjacent key points often have similarity, which helps to improve the accuracy of correction. Step S5032, adjusting the parameters of the gated recurrent unit network to minimize the residual using the gradient descent method; this sub-step describes the optimization method of parameter adjustment. The gradient descent method adjusts the parameters in the opposite direction of the gradient by calculating the gradient of the loss function on the network parameters, gradually reducing the residual. The loss function is defined as:

[0064]

[0065] In the formula, is the loss function, is the number of key points, is the residual, is the network parameter, is the regularization coefficient.

[0066] Regularization term Prevent overfitting caused by excessive parameter adjustment. The choice of learning rate is crucial: too large may cause the optimization process to oscillate, and too small will slow down the convergence speed. The invention adopts an adaptive learning rate strategy, dynamically adjusting the learning rate according to the residual size, ensuring that the optimization process is both fast and stable. Step S5033, repeat steps S5031 and S5032 until the residual is less than the tolerance value. This sub-step completes the iterative process of physical constraint correction. Iterative correction ensures high consistency between the deduction result and the physical model, and the tolerance value is usually set to 5%-10% of the average deformation of the key points. The iterative process may require 3-10 times, depending on the degree of difference between the initial deduction and the physical model. To prevent infinite loops, the system also sets a maximum number of iterations (usually 20). This iterative correction mechanism enables the system to find the best balance point between physical laws and actual observations, neither over-relying on possibly inaccurate physical models nor completely ignoring the constraints of physical laws. Further, in step S70, the generation and output of the three-dimensional dynamic deformation field includes:

[0067] Step S701, converting the output data or the deduced roadway deformation data into gridded deformation data, wherein the gridded deformation data includes deformation vectors and time series of each grid point of the coal mine roadway; this sub-step completes the conversion from discrete data to continuous field. Gridding is the process of interpolating point data (whether it is a laser radar point cloud or a deduced result) onto a regular grid, and common methods include inverse distance weighting, Kriging interpolation, etc. The present application uses an adaptive grid based on the geometry of the roadway, and denser grids are used in key areas (such as the top and corners), and sparser grids are used in flat areas, which ensures the accuracy of key areas and controls the amount of calculation. The deformation vector of each grid point contains three directional displacement components (u, v, w) and time series information, forming a four-dimensional data set (x, y, z, t). Step S702, based on the gridded deformation data, generating a graphical representation of the three-dimensional dynamic deformation field through a visualization algorithm; this sub-step realizes the intuitive display of deformation data. Visualization algorithms include:

[0068] Deformation cloud chart: use color mapping to represent deformation size, usually red for large deformation and blue for small deformation

[0069] Deformation vector: use arrows to represent deformation direction and size

[0070] Deformation animation: show the evolution of deformation over time

[0071] Sectional view: show the deformation distribution on a specific cross-section

[0072] The innovation of the present application lies in the close combination of physical deformation and roadway structure, for example, the original profile of the roadway is retained as a reference in the cloud chart, making the deformation degree obvious at a glance. The system also provides multiple perspective switching and local magnification functions to facilitate technicians to focus on key areas. Step S703, outputting the three-dimensional dynamic deformation field to the monitoring system for real-time display and early warning. This sub-step completes the final application of the monitoring result. The output interface usually uses standard communication protocols (such as MQTT, OPCUA) to ensure compatibility with existing monitoring systems. The system not only provides visual display, but also integrates intelligent early warning function: when the deformation rate exceeds the threshold or the deformation mode is abnormal, different levels of early warning (such as yellow warning, red warning) are automatically triggered. The early warning threshold is set based on historical data and rock mass mechanics analysis, taking into account the specific conditions and safety margin of the roadway. In addition, the system also provides data export function, which is convenient for subsequent analysis and archiving. Further, step S80 is included;

[0073] Step S80, periodically updating the roadway digital twin model, wherein the updating is based on the historical three-dimensional dynamic deformation field, using a Transformer mechanism for feature extraction, and through a gating mechanism to control the updating rhythm of the model parameters. This step is an extension of the invention, realizing the adaptive evolution of the digital twin model. In traditional monitoring systems, the digital twin model is usually static and cannot adapt to long-term changes in roadway conditions (such as rock mass creep and support structure aging). The present invention updates the model parameters periodically, enabling the digital twin model to "learn" the actual behavior of the roadway and improve the accuracy of long-term monitoring. The updating process includes two key links: first, using the Transformer mechanism to extract time series features from historical deformation data to identify long-term trends in deformation patterns; then, through the gating mechanism to control the updating rhythm of the model parameters, avoiding overfitting to short-term fluctuations. This "learning-updating" mechanism enables the system to have long-term adaptability and cope with slow changes in roadway conditions.

[0074] Step S80, in step S80, the periodic updating includes:

[0075] Step S801, collecting the three-dimensional dynamic deformation field of the historical period as a training data set;

[0076] This sub-step defines the data basis for model updating. The training data set contains sufficient historical deformation data (usually 3-6 months), covering various working conditions and deformation patterns. Data preprocessing includes outlier rejection, normalization processing, etc. to ensure the quality of the training data. The time span of historical data is critical: too short to capture long-term trends, too long may contain outdated deformation patterns. The present invention uses a sliding time window strategy, prioritizing recent data while retaining some historical data to maintain the continuity of long-term trends.

[0077] Step S802, using a Transformer encoder to extract time series features from the training data set to obtain a feature vector;

[0078] This sub-step details the feature extraction process. The Transformer encoder extracts key time series features such as deformation trends, periodic changes, and abrupt points from historical deformation data through multiple layers of self-attention mechanisms. Compared with traditional feature extraction methods, Transformer can capture long-distance dependencies, such as identifying complex patterns like "top deformation acceleration usually occurs 24 hours after the side wall deformation is obvious". The feature vector is a compressed representation of the original data, retaining the most important information while removing noise and redundancy.

[0079] Step S803, inputting the feature vector into a gated recurrent unit network, outputting the corrected mechanical parameters through a gating mechanism, and updating the roadway digital twin model.

[0080] The sub-step completes the adaptive update of the model parameters. The gating mechanism functions as a "regulating valve" to control the amplitude and rhythm of the parameter update. For example, when the feature vector shows that the deformation trend is stable, the gating mechanism limits the amplitude of the parameter update to avoid overfitting to short-term fluctuations; when the feature vector shows a significant trend change, the gating mechanism allows a larger update to enable the model to quickly adapt to the new situation. The specific update formula is:

[0081]

[0082] wherein, is the updated parameter, is the original parameter, is the gating signal (0≤g≤1), is the update amount based on the feature vector .

[0083] The above only describes the preferred embodiments of the patent and is not intended to limit the patent. Any modifications, equivalent replacements, and improvements made within the spirit and principle of the patent should be included in the protection scope of the patent.

Claims

1. A three-dimensional dynamic monitoring method for coal mine roadways using lidar, characterized in that, The method includes: Step S10, simultaneously acquiring lidar surface data collected by lidar deployed in the coal mine roadway and optical strain gauge point data collected by multiple optical strain gauges deployed at key points in the coal mine roadway; Step S20, constructing a digital twin model of the roadway based on the geometric and mechanical parameters of the coal mine roadway, the digital twin model including a roadway geometric model and a roadway mechanical model; Step S30, based on the roadway digital twin model, using the Transformer mechanism to perform multi-source data fusion and data assimilation processing on the lidar surface data and the optical strain gauge point data to obtain initial roadway deformation data; Step S40, based on Bayesian estimation theory, determining... Step S50: When the lidar surface data is missing or the confidence value is lower than a preset threshold, based on the optical strain gauge point data and the roadway digital twin model, a gating mechanism is used to deduce the full-field deformation of the coal mine roadway to obtain deduce roadway deformation data; Step S60: When the lidar surface data is not missing and the confidence value is not lower than the preset threshold, the initial roadway deformation data is used as output data; Step S70: Based on the output data or the deduce roadway deformation data, a three-dimensional dynamic deformation field of the coal mine roadway is generated and output, wherein the three-dimensional dynamic deformation field represents the continuous deformation state of the coal mine roadway in time sequence.

2. The three-dimensional dynamic monitoring method for coal mine roadways using lidar according to claim 1, characterized in that, In step S10, the synchronous acquisition includes: step S101, continuously scanning the coal mine roadway using the lidar to generate lidar surface data, wherein the lidar surface data includes point cloud data and its timestamp; step S102, synchronously acquiring strain signals of the key points using the multiple optical strain gauges to generate optical strain gauge point data, wherein the optical strain gauge point data includes strain values, their position coordinates, and timestamps; step S103, performing time alignment processing on the lidar surface data and the optical strain gauge point data to ensure that the lidar surface data and the optical strain gauge point data are under the same time reference.

3. The three-dimensional dynamic monitoring method for coal mine roadways using lidar according to claim 1, characterized in that, In step S20, constructing the digital twin model of the mine roadway includes: step S201, acquiring the design drawings and geological exploration data of the coal mine roadway, and extracting the geometric parameters, wherein the geometric parameters include the roadway cross-sectional shape, size, and spatial coordinates; step S202, determining the mechanical parameters based on the rock mass mechanical properties of the coal mine roadway, wherein the mechanical parameters include the elastic modulus, Poisson's ratio, and yield strength; step S203, using the geometric parameters and the mechanical parameters, constructing the roadway geometric model and the roadway mechanical model through the finite element method, and integrating them to form the digital twin model of the roadway.

4. The three-dimensional dynamic monitoring method for coal mine roadways using lidar according to claim 1, characterized in that, In step S30, the multi-source data fusion and assimilation processing using the Transformer mechanism includes: Step S301, converting the lidar surface data and the optical strain gauge point data into time-series data, wherein the time-series data includes deformation values ​​at multiple time steps; Step S302, processing the time-series data using a Transformer encoder, calculating the correlation weight between the lidar surface data and the optical strain gauge point data through a self-attention mechanism, and performing a weighted summation based on the correlation weight to generate fused deformation data; Step S303, assimilating the fused deformation data using the tunnel mechanics model, optimizing the initial tunnel deformation data by minimizing the error between the model prediction and the fused deformation data.

5. The three-dimensional dynamic monitoring method for coal mine roadways using lidar according to claim 1, characterized in that, In step S40, determining the credibility of the lidar surface data includes: step S401, constructing a likelihood function for the lidar surface data using the optical strain gauge point data as a prior distribution; step S402, calculating the posterior probability distribution of the lidar surface data based on Bayes' theorem; and step S403, extracting a variance index from the posterior probability distribution as the credibility value, wherein the variance index represents the uncertainty of the lidar surface data.

6. The three-dimensional dynamic monitoring method for coal mine roadways using lidar according to claim 1, characterized in that, In step S50, the process of using a gating mechanism to deduce the full-field deformation of the coal mine roadway includes: Step S501, when the lidar surface data is missing, extracting the deformation amount of the key point from the optical strain gauge point data as input; Step S502, using a gated cyclic unit network, taking the deformation amount of the key point as the initial input, and controlling the addition of new information through the input gate, the retention of historical information through the forget gate, and the output of the hidden state through the output gate, to deduce the full-field deformation; Step S503, performing physical constraint correction on the deduction results based on the roadway mechanical model, and outputting the deduced roadway deformation data by iteratively updating the hidden state of the gated cyclic unit network.

7. The three-dimensional dynamic monitoring method for coal mine roadways using lidar according to claim 6, characterized in that, In step S503, the physical constraint correction includes: step S5031, calculating the residual between the deduction result and the predicted value of the tunnel mechanics model at the key point; step S5032, adjusting the parameters of the gated recurrent unit network using the gradient descent method to minimize the residual; step S5033, repeating steps S5031 and S5032 until the residual is less than the tolerance value.

8. The three-dimensional dynamic monitoring method for coal mine roadways using lidar according to claim 1, characterized in that, In step S70, generating and outputting the three-dimensional dynamic deformation field includes: step S701, converting the output data or the inferred roadway deformation data into gridded deformation data, wherein the gridded deformation data includes the deformation vector and time series of each grid point of the coal mine roadway; step S702, generating a graphical representation of the three-dimensional dynamic deformation field based on the gridded deformation data using a visualization algorithm; and step S703, outputting the three-dimensional dynamic deformation field to a monitoring system for real-time display and early warning.

9. The three-dimensional dynamic monitoring method for coal mine roadways using lidar according to claim 1, characterized in that, The method further includes: step S80, periodically updating the digital twin model of the tunnel, wherein the update is based on the historical three-dimensional dynamic deformation field, feature extraction is performed using the Transformer mechanism, and the update rhythm of the model parameters is controlled by a gating mechanism.

10. The three-dimensional dynamic monitoring method for coal mine roadways using lidar according to claim 9, characterized in that, In step S80, the periodic update includes: step S801, collecting the three-dimensional dynamic deformation field from historical periods as a training dataset; step S802, using a Transformer encoder to extract temporal features from the training dataset to obtain feature vectors; step S803, inputting the feature vectors into a gated recurrent unit network, outputting the corrected mechanical parameters through a gating mechanism, and updating the tunnel digital twin model.

Images