Mine geological disaster early warning method based on artificial intelligence

Abstract

The invention relates to the field of mine geological disaster monitoring and early warning, and discloses a mine geological disaster early warning method based on artificial intelligence, and the method comprises the steps: describing the dynamic change of a state variable and a multi-physics coupling effect in a disaster evolution process through constructing a kinetic model; carrying out discrete solution on the model by adopting a numerical method to obtain a disaster evolution path; a disaster instability critical point is accurately identified in combination with a bifurcation theory and Lyapunov stability analysis; designing an optimal early warning strategy by using a reinforcement learning algorithm, and dynamically selecting an alarm signal triggering opportunity; fusing monitoring data and a model in real time, and dynamically updating a disaster state and a risk level; and outputting an alarm signal and providing emergency suggestions. According to the method, real-time response of accurate disaster prediction and early warning is realized, and the mine geological disaster prevention and control capability is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of mine geological disaster monitoring and early warning, and in particular to a mine geological disaster early warning method based on artificial intelligence. Background Art

[0002] Mine geological disasters have always been a major challenge in the field of mine safety production and environmental protection. The complexity of geological disasters is reflected in the coupling of multiple physical fields, dynamic evolution characteristics, and highly nonlinear behavior. These characteristics lead to many technical deficiencies in traditional disaster warning technologies when facing complex mining conditions, making it difficult to effectively meet the needs of real-time, accuracy, and intelligence. The following is an analysis of the deficiencies of existing technologies.

[0003] Existing disaster prediction methods lack a systematic description of multi-physics coupling

[0004] Traditional disaster prediction methods often rely on the analysis of a single physical field, such as a stress field or groundwater seepage field. This method relies too much on specific geological conditions and empirical formulas, and cannot accurately reflect the complexity of the coupling of multiple physical fields during the evolution of disasters. Especially in mining environments, the mutual influence between the stress field and the seepage field is significant, and the analysis results of a single field are difficult to fully characterize the dynamic process of disaster evolution. These deficiencies in existing technologies have led to limited coverage of disaster prediction and are unable to meet the early warning needs of various types of disasters under complex mining conditions.

[0005] Existing critical point identification methods are not accurate enough and have a delayed response

[0006] In disaster warning, the identification of critical points is the key. In existing technologies, many methods use empirical threshold methods or statistical models to determine whether the disaster state has reached instability. However, these methods rely too much on historical data and fixed thresholds, and lack the ability to respond in real time to dynamic changes in disaster systems. In addition, although some methods can capture the evolution trend of the state, there are still large deviations in the accurate identification of critical points, especially in complex nonlinear disaster systems. This lack of accuracy and delayed response directly reduces the credibility of disaster warnings and limits the decision-making efficiency of managers.

[0007] The early warning strategy design is rigid and lacks the ability to dynamically adjust

[0008] Current early warning strategies are usually based on alarm thresholds or empirical parameters set based on fixed rules. This static early warning strategy may have a certain effect in the early stages of a disaster, but when the speed of disaster evolution changes or external environmental conditions (such as rainfall intensity) fluctuate rapidly, the existing strategy cannot be adjusted dynamically, which can easily cause the alarm to be too early or too late. In addition, existing early warning systems are usually unable to balance the lead time and accuracy of alarms, which may affect user trust due to frequent false alarms, and may also miss critical opportunities due to delayed alarms.

[0009] Disaster warning and real-time monitoring data are poorly integrated

[0010] Existing disaster warning technologies are highly dependent on real-time monitoring data, but the dynamic fusion between data and models is still significantly insufficient. Many technologies only use monitoring data for simple model initialization or threshold adjustment, but fail to achieve dynamic parameter calibration and real-time update of disaster status. In addition, since monitoring data is usually heterogeneous and incomplete, existing technologies also lack an effective fusion mechanism in data processing, resulting in the inability to fully reflect the actual application value of monitoring data. This lack of fusion makes it difficult for disaster warning models to adapt to the rapid changes in the mining environment in real time, further affecting the accuracy and timeliness of warnings.

[0011] Therefore, the present invention proposes an artificial intelligence-based mine geological disaster early warning method to address the deficiencies of the prior art. Summary of the invention

[0012] In view of the problems existing in the prior art in mine geological disaster early warning, such as insufficient multi-physical field description, low critical point identification accuracy, rigid early warning strategy, and low degree of data and model fusion, the present invention proposes a mine geological disaster early warning method based on artificial intelligence, which aims to realize dynamic modeling of disaster status, accurate critical point identification, optimal early warning strategy design and real-time triggering of alarm signals, so as to comprehensively improve the intelligence level and real-time response capability of mine geological disaster early warning.

[0013] To achieve the above objectives, the present invention is implemented through the following technical solutions: a mine geological disaster early warning method based on artificial intelligence, comprising the following steps:

[0014] Constructing a dynamic model of mining geological hazards, which describes the state variables of the hazard evolution process and its evolutionary behavior under the influence of external driving forces and spatial diffusion;

[0015] Use numerical methods to discretize and solve the dynamic model to obtain the disaster evolution path;

[0016] The stability of the disaster system is analyzed based on the bifurcation theory, and the critical point of instability of the disaster is identified by calculating the eigenvalue of the Jacobian matrix of the state equation;

[0017] Use reinforcement learning algorithms to design optimal warning strategies and dynamically select the optimal warning time based on the disaster evolution path and critical points;

[0018] Integrate real-time monitoring data with disaster models to dynamically update disaster status and alarm signals;

[0019] Output alarm signals based on the optimal early warning strategy, and provide risk levels and emergency suggestions.

[0020] Preferably, the kinetic model is described by a partial differential equation, which includes a time variation term of the state variable, a spatial diffusion term and an external driving force term, in the form of: the sum of the time derivative of the state variable and the function of the external driving force and the function of the spatial diffusion effect is zero, and the initial conditions and boundary conditions are determined by field monitoring data.

[0021] Preferably, the disaster evolution path is obtained by a numerical solution method, and the numerical solution method includes:

[0022] The finite element method is used to spatially discretize the dynamic model;

[0023] The time integration algorithm is used to dynamically solve the discretized state equation to obtain the change of the disaster evolution path over time.

[0024] Preferably, the critical point is identified by bifurcation theory, comprising the following steps:

[0025] Linearize the disaster evolution state equation and construct the Jacobian matrix;

[0026] The eigenvalues ​​of the Jacobian matrix are calculated and the instability critical point of the system is identified when the real part of the eigenvalue becomes zero.

[0027] Preferably, the stability of the disaster system is further verified by Lyapunov stability theory, comprising the following steps:

[0028] A Lyapunov function of the disaster system is defined, wherein the Lyapunov function includes a kinetic energy term and an external potential energy term of a system state variable;

[0029] The time derivative of the Lyapunov function is calculated and the critical state of the system is identified when the derivative of the Lyapunov function is zero.

[0030] Preferably, the reinforcement learning algorithm comprises the following steps:

[0031] Define the time series of disaster status as the state input of reinforcement learning, and define whether to issue an alarm signal as the action output;

[0032] Set the reward function and calculate the reward value based on the deviation between the alarm signal and the critical point time and the warning accuracy;

[0033] The reinforcement learning algorithm is used to optimize the strategy network and select the optimal alarm signal output.

[0034] Preferably, the reinforcement learning algorithm adopts the ProximalPolicyOptimization algorithm, and the policy network outputs the probability distribution of the alarm signal by inputting the characteristic vector of the disaster evolution state.

[0035] Preferably, the model is integrated with the real-time monitoring data, comprising the following steps:

[0036] Obtain real-time monitoring data from field sensors, including ground surface displacement, crack growth rate, groundwater level, and rainfall;

[0037] Use real-time data to calibrate the parameters of the dynamic model and dynamically update the disaster evolution path;

[0038] Re-evaluate system stability based on real-time monitoring data and update alarm signals.

[0039] Preferably, the alarm signal output includes risk level assessment information, and the risk level is determined comprehensively by the time distance to the disaster critical point and the stability assessment result of the current state.

[0040] Preferably, the mine geological disaster early warning system based on artificial intelligence includes:

[0041] Data acquisition module, used to collect real-time monitoring data of the mine, including surface displacement, crack width, groundwater level and rainfall;

[0042] Modeling module, used to construct the dynamic model of mining geological hazards, including the time derivative, spatial diffusion and external driving force terms of state variables;

[0043] The calculation module is used to calculate the disaster evolution path using numerical solution methods and identify the critical point of instability based on bifurcation theory and Lyapunov stability theory;

[0044] Reinforcement learning module, used to optimize the timing of alarms based on disaster evolution paths and critical points;

[0045] The alarm module is used to output alarm signals according to the optimal early warning strategy and provide risk levels and emergency suggestions;

[0046] The model update module is used to calibrate model parameters in combination with real-time data and update disaster status and alarm signals.

[0047] The present invention provides an artificial intelligence-based mine geological disaster early warning method, which has the following beneficial effects:

[0048] 1. The present invention adopts mining geological disaster analysis technology based on dynamic modeling and bifurcation theory, which can fully describe the temporal and spatial characteristics of disaster evolution, and achieve the technical effect of accurate prediction of disaster evolution path and efficient identification of critical points. Compared with the technical solutions in the prior art that mainly rely on simple empirical rules or univariate linear prediction methods, the present invention solves the technical problems of poor model universality and inaccurate identification of disaster critical points, especially under complex mining geological conditions (such as multi-physical field coupling, fault structure), and significantly improves the analysis ability. The stability of the disaster state is dynamically monitored through the bifurcation theory. The present invention can also capture the slight changes in the disaster state in real time, providing strong support for early warning of disaster instability.

[0049] 2. The present invention optimizes the dynamic selection of disaster warning opportunities by introducing a deep reinforcement learning algorithm and combining it with a disaster dynamics model, thereby achieving the technical effect of accurate triggering of alarm signals and high response efficiency. Compared with the technical solutions based on fixed alarm thresholds or static rules in the prior art, the present invention solves the problems of delayed warning response or high false alarm rate, especially in scenarios with high complexity of disaster evolution. The reinforcement learning model of the present invention can continuously and adaptively adjust the alarm strategy, making the early warning system highly intelligent and adaptive. At the same time, the present invention achieves a balance between alarm lead time and accuracy through the design of a reward function, avoiding the problem of invalid warnings caused by too early or too late warning times.

[0050] 3. The present invention dynamically integrates real-time monitoring data with the disaster dynamics model, and realizes efficient dynamic tracking of disaster status and accurate output of alarm signals through real-time update of parameter calibration and model solution, achieving the technical effect of highly matching model calculation with actual scene. Compared with the technical solutions based on static data input or single model operation in the prior art, the present invention solves the technical bottlenecks of model update lag and inability of monitoring data to directly drive early warning. In particular, the present invention supports the real-time fusion of multi-source heterogeneous data (such as surface displacement, groundwater level, stress transfer, etc.), can adapt to the complex needs of rapid changes in mining environment, and ensure the real-time, reliability and flexibility of early warning.

[0051] 4. Based on the disaster warning output, the present invention combines the risk level assessment and emergency recommendation generation technology to construct a complete closed-loop solution from disaster monitoring to decision support, achieving the technical effect of providing clear and scientific decision-making basis for managers. Compared with the prior art technical solutions that only provide simple risk prompts or single alarm signals, the present invention solves the problems of insufficient practicality of early warning results and unclear emergency handling guidance. Through the quantitative assessment of risk levels and the combination of spatial distribution visualization technology, the present invention intuitively displays the distribution of mine disaster risks. At the same time, through the emergency recommendation module, accurate treatment strategies can be generated for different disaster types, such as personnel evacuation, equipment shutdown, or engineering reinforcement recommendations, providing comprehensive support for the rapid response and scientific decision-making of mine managers in high-risk situations. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 It is a flowchart of a mine geological disaster warning method based on artificial intelligence;

[0053] Figure 2 It is an architecture diagram of a mine geological disaster warning system based on artificial intelligence. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0054] Next, the technical solutions of the present invention will be clearly and completely described in conjunction with the drawings of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0055] Please refer to the attached Figure 1 , the embodiments of the present invention provide a mine geological disaster warning method based on artificial intelligence. The following will detail each step of the method of the present invention.

[0056] S1. Construct a dynamic model of mine geological disasters, and the model describes the state variables of the disaster evolution process and their evolution behaviors under the action of external driving forces and spatial diffusion

[0057] In the implementation process of the present invention, step S1 is the basic link of the entire technical solution. This step mainly solves how to accurately describe the evolution process of mine geological disasters through a mathematical model. The establishment of this model not only needs to combine the specific characteristics of the mine environment, but also needs to be able to reflect the multi-physical field coupling mechanism in the disaster evolution process, providing a theoretical basis for subsequent numerical solutions, critical point identification, and optimal early warning strategies. Especially in mine geological disasters, the dynamic interaction between external driving forces (such as rainfall, mining activities) and stress, seepage diffusion is the core influencing factor of disaster evolution. Therefore, this step quantitatively describes these factors through a dynamic model.

[0058] In this embodiment, the dynamic model of mining geological disasters is constructed based on partial differential equations (PDEs), the core of which is to describe the state variables in the disaster evolution process and their dynamic behaviors affected by external driving forces and spatial diffusion mechanisms.

[0059] Generally speaking, the model includes the following three parts: the time evolution term of the state variable, the spatial diffusion term and the external driving force term.

[0060] Specifically, the basic form of the model is as follows:

[0061]

[0062] In the above formula, the definitions of the parameters are as follows:

[0063] u(xt): The state variable of the disaster, which is a function of time t and spatial position x, and can be specifically expressed as the distribution of surface displacement, groundwater pressure or rock stress.

[0064] The time derivative of the state variable reflects the dynamic changes of the state over time.

[0065] The local source term function is used to describe the impact of external driving forces on the evolution of disasters and is often related to external factors such as rainfall intensity and mining load.

[0066] The transfer term describes the diffusion behavior of the disaster state, involving spatial distribution characteristics such as stress transfer and seepage diffusion.

[0067] x∈Ω: The spatial region Ω represents the geological area of ​​the mine.

[0068] t∈[0,T]: time interval, representing the total time range of disaster evolution.

[0069] In some embodiments, the external driving force It is expressed in the following form:

[0070]

[0071] Among them, P(x,t) represents external environmental factors, such as the change of rainfall intensity over time and space;

[0072] Q(ut) represents the local disturbance caused by mining activities, which can be described in linear or nonlinear form.

[0073] As an option, the spatial diffusion term It can be further defined as:

[0074]

[0075] Where D is the diffusion coefficient matrix, which is used to characterize the heterogeneity of different physical fields (such as stress field or seepage field) in the spatial diffusion process; is the spatial gradient of the state variable.

[0076] In general, in order for the model to accurately reflect the actual geological characteristics of the mine, it is also necessary to set initial conditions and boundary conditions:

[0077] The initial conditions are:

[0078] u(x0)=u 0 (x)

[0079] Among them, u0(x) represents the initial state of the disaster, which is usually obtained by inverting the historical data provided by the monitoring system or the on-site geological conditions.

[0080] The boundary conditions are:

[0081] u(xt)=g(x,t),

[0082] Here, g(x,t) represents the known state value on the boundary, such as the surface displacement, groundwater level change, or stress distribution at the boundary point.

[0083] In a possible implementation, the diffusion coefficient matrix D can be dynamically adjusted according to the rock and soil medium properties of the mine. For example, in a multi-layer rock mass structure, different diffusion coefficients are used for different rock layers.

[0084] In this embodiment, in order to consider the dynamic disturbance caused by mining activities, the external driving force term Q(u,t) can also be expressed in a nonlinear form, for example:

[0085] Q(ut)=αu 2 +βsin(ωt)

[0086] Among them, α and β are nonlinear coefficients, reflecting the impact of mining activity intensity on the disaster state; ω is the disturbance frequency, which is used to simulate the periodic vibration of mining equipment.

[0087] The dynamic model of this embodiment also supports the coupled analysis of multiple physical fields. For the stress field σ(x, t) and the seepage field p(x, t), the following state variables can be defined respectively:

[0088]

[0089] Among them, C is the elastic modulus matrix of the rock mass; κ is the permeability coefficient, which represents the flow characteristics of groundwater in different rock and soil media.

[0090] In the multi-physics coupling model, the stress field and the seepage field are dynamically linked through the state variable u(x, t). For example, when the rainfall increases, the change of the seepage field p(x, t) will lead to the redistribution of the stress field σ(x, t), thereby accelerating the instability of the disaster.

[0091] In specific implementation, this dynamic model can be used not only in the mining area of ​​the mine, but also in potential disaster points around the mine, such as slopes, abandoned mines, etc.

[0092] By constructing the above dynamic model, this embodiment can comprehensively describe the dynamic evolution process of mining geological disasters and provide basic theoretical support for subsequent steps. The multi-physical field coupling characteristics and nonlinear perturbation description capabilities of the model enable it to adapt to complex mining environments and diverse disaster types. S2. Discretize and solve the dynamic model using numerical methods to obtain the disaster evolution path

[0093] In the implementation process of the present invention, step S2 aims to numerically solve the mining geological disaster dynamics model constructed in step S1. Due to the nonlinear characteristics of the disaster evolution process and the complexity of multi-physical field coupling, the analytical solution of the dynamic model is usually difficult to achieve, so it is necessary to discretize it using a numerical method in order to calculate and simulate the dynamic evolution path of the disaster state variables. The core of step S2 is to convert the continuous mathematical model into a discrete form through discretization technology, and use an efficient numerical algorithm to obtain the law of change of the disaster evolution state over time and space.

[0094] Generally speaking, this step requires discretization of spatial variables and time variables separately. The spatial discretization usually uses the finite element method (FEM) or the finite difference method (FDM), and the time discretization selects an appropriate time integration algorithm (such as the implicit Euler method or the Runge-Kutta method) according to actual needs.

[0095] In this embodiment, the finite element method is used to perform spatial discretization processing on the dynamic model constructed in step S1. The finite element method converts continuous field variables into finite-dimensional discrete field variables by dividing the calculation domain into a finite number of units, thereby ensuring efficient calculation in complex mine geometry.

[0096] In one possible implementation, the spatial discretization process of the dynamic model is as follows: Assume that the computational domain of the mine is Ω, which is divided into a finite number of units Ω e (e=1,2,…,N), where N is the total number of units. In each unit, the disaster state variable u(x,t) is interpolated by the interpolation function φ i (x) is expressed as:

[0097]

[0098] Among them, φ i (x) is the interpolation function, which represents the distribution form of the state variable on the unit; u i (x) is the value of the state variable at the interpolation node; n is the number of interpolation nodes for each unit.

[0099] Through the above discretization, the spatial terms of the dynamic model Convert to discrete matrix form:

[0100]

[0101] where u e Represents the discrete field variables within the cell.

[0102] As an option, to improve computational efficiency, the transfer term Perform integral division processing and convert it into gradient form:

[0103]

[0104] in, is the gradient of the interpolation function; n is the unit normal vector of the boundary; Γ is the boundary area.

[0105] The boundary integral part of the above formula It represents the contribution of the dynamic model on the mine boundary, which can be determined based on actual monitoring data or boundary conditions.

[0106] After the spatial discretization is completed, the time variable needs to be discretized. In this embodiment, the implicit Euler method is used to discretize the time derivative term. Discretize, the specific formula is as follows:

[0107]

[0108] in:

[0109] u k and u k+1 are the state variables of the kth step and the k+1th step respectively;

[0110] Δt is the time step.

[0111] Combining the dynamic model after spatial discretization, the following matrix equation can be obtained after time discretization:

[0112]

[0113] Where M is the mass matrix, which is obtained by discretizing the time derivative of the state variable; K(u k ) is the stiffness matrix, and the transfer term Discretization gives: f(uk ) is the external force term, which is driven by the external driving force Discretization yields; u k and u k+1 are the discrete state vectors of the current time step and the next time step respectively.

[0114] In a possible implementation, the Newton-Raphson iteration method can be used to solve the above nonlinear equation, and the evolution path of the disaster state over time can be obtained through continuous iteration.

[0115] In some embodiments, in order to improve the efficiency of numerical calculations, the calculation domain of the mine can be partitioned in combination with a regional partitioning strategy. For example, for areas that are severely affected by disasters, a finer unit division can be used; while for peripheral areas that are less affected by disasters, the unit size can be appropriately increased to reduce the amount of calculation.

[0116] This embodiment also supports the selection of multiple time integration algorithms. For example, as an option, if higher time accuracy is required, the Runge-Kutta method can be used to discretize the time variable, which is in the form of:

[0117]

[0118] Where s is the order of the Runge-Kutta method: b i and k i are weights and intermediate increments, which are determined by specific algorithm parameters.

[0119] Through the above-mentioned discretization process of space and time, this embodiment converts the dynamic model of the disaster into a numerically solvable discrete equation group, and dynamically calculates it using a numerical solution method, thereby obtaining the evolution path of the disaster state over time.

[0120] The numerical method in this embodiment is applicable to various complex mining geological structures, including multi-layer rock mass, fault structure, etc., and can dynamically adjust the division of discrete units and the selection of time steps according to the accuracy of monitoring data, thereby ensuring the stability and efficiency of calculation. Ultimately, the results of the discretization solution provide the necessary data support for the subsequent analysis of the disaster evolution path and the identification of critical points.

[0121] S3. Analyze the stability of the disaster system based on bifurcation theory and identify the critical point of instability of the disaster by calculating the eigenvalue of the Jacobian matrix of the state equation

[0122] Step S3 is a further analysis of the above disaster evolution path to identify the critical node of the disaster occurrence - the critical point. The determination of the critical point is crucial for disaster warning because it marks the transition of the disaster system from a stable state to an unstable state. This step is based on the bifurcation theory and Lyapunov stability theory. Through mathematical analysis and numerical calculation, the critical point t of the disaster system is accurately located. c This kind of analysis can provide a scientific basis for early warning strategies and make the triggering of alarm signals more accurate.

[0123] In general, the critical point identification of the disaster system is determined by linear analysis of the dynamic state equation and the law of change of the system energy. The above step S2 has obtained the evolution path u(t) of the disaster state, which provides the required input data for this step.

[0124] In this embodiment, the stability analysis of the disaster system first uses the Jacobian matrix eigenvalue method of bifurcation theory for identification. Specifically, the dynamic state equation is discretized and expressed as follows:

[0125]

[0126] Among them, u is the discrete vector of disaster status; f(u,t) represents the change function of state variables, including the combined effect of external driving force and spatial diffusion.

[0127] In general, the state equation is linearized and the Jacobian matrix J(u) is constructed:

[0128]

[0129] Among them, J(u) represents the local impact of a small disturbance on the state variables in the current state of the system.

[0130] In a possible implementation, the stability of the system can be determined by calculating the eigenvalues ​​of the comparable matrix. When the real part of all eigenvalues ​​is less than zero, the system is in a stable state; when the real part of at least one eigenvalue is greater than or equal to zero, the system enters a bifurcation instability state. c The following conditions are met:

[0131] Re(λ i )=0,λ i ∈eig(J(u))

[0132] Among them, Re(λ i ) is the real part of the eigenvalue; eig(J(u)) represents the eigenvalue set of the Jacobian matrix J(u).

[0133] As an option, in order to improve the efficiency and accuracy of the calculation, the QR decomposition method can be used to decompose the Jacobian matrix eigenvalues. In the specific calculation process, the numerical method is used to interpolate the evolution of u(t) in time to obtain the state variable value at any time, thereby realizing the dynamic monitoring of the eigenvalue.

[0134] In order to further verify the stability of the system, this embodiment also adopts the Lyapunov stability theory, and determines the state of the system by defining the energy function V(u,t) of the disaster system. Specifically, the Lyapunov function is as follows:

[0135]

[0136] in, represents the kinetic energy of the system state variables; U(u,t) represents the potential energy of the system, which is related to the external driving force.

[0137] As a possible implementation, the definition of the external potential energy U(u,t) can be based on the external driving force in the dynamic model Perform calculations. For example:

[0138]

[0139] By calculating the time derivative of the Lyapunov function Determine the energy change trend of the system:

[0140]

[0141] when , the system reaches a critical point.

[0142] In one embodiment, if the Lyapunov function of the disaster system is at a certain time t c The derivative of is zero, and the energy of the system continues to increase thereafter, then this time point can be judged as the critical point of instability of the disaster.

[0143] In some embodiments, in order to improve the robustness of critical point identification, the energy method and frequency domain analysis can be combined for comprehensive judgment. For example, by analyzing the frequency spectrum characteristics of the state variable u(t), the sudden change point of the characteristic frequency can be identified as an auxiliary criterion for the critical point.

[0144] This embodiment also supports classification analysis of bifurcation patterns for different types of disasters. For example, for mine landslide disasters, the bifurcation point may appear as a saddle-node bifurcation or a Hopf bifurcation; while for debris flow disasters, the bifurcation point may appear as a supercritical bifurcation. Specifically, the characteristic value variation law and Lyapunov function characteristics of different bifurcation modes are different. In practical applications, the accuracy of critical point identification can be further improved based on monitoring data and model calculation results, combined with the specific classification method of bifurcation theory.

[0145] Through the combined analysis of the above bifurcation theory and Lyapunov stability theory, this embodiment achieves accurate identification of the critical point of the disaster. This step provides the necessary input data for the subsequent optimal early warning strategy design and ensures that the early warning signal can be triggered in time. At the same time, since this embodiment supports the combined application of multiple mathematical analysis methods, it can adapt to the complex needs of different mining geological environments and disaster types.

[0146] S4. Use reinforcement learning algorithms to design optimal warning strategies and dynamically select optimal warning times based on disaster evolution paths and critical points

[0147] Step S4 is based on the disaster evolution path calculated in step S2 and the disaster instability critical point t identified in step S3. c , optimize the optimal early warning strategy through reinforcement learning algorithm. The core goal of this step is to select the optimal alarm time t according to the dynamic disaster evolution state and the predicted instability moment. p , so as to give early warning as close to the critical point as possible and maximize the actual benefits of early warning. The reinforcement learning algorithm in this step can learn the complex evolution law of disasters and dynamically adjust the timing of alarms.

[0148] Generally speaking, disaster warning strategies need to balance the following two aspects: First, the advance time of the alarm, that is, the warning time t p With critical point t c The second is the accuracy of the warning, that is, whether the alarm is triggered near the critical point. This step achieves the above goals by defining a reasonable reward function and combining it with a deep reinforcement learning model for strategy optimization.

[0149] In this embodiment, a disaster warning strategy is designed using a deep reinforcement learning (DRL) algorithm. The basic framework of reinforcement learning consists of an environment, state, action, and reward function, which is specifically applied to the present invention and is defined as follows:

[0150] environment:

[0151] Dynamic model of the disaster system, including its evolution path u(t) and critical point of instability t cThe environment is responsible for receiving the current state input and returning the result of the action.

[0152] Status t :

[0153] Current status of the disaster t , which is composed of the disaster evolution path obtained by numerical solution and real-time monitoring data, mainly including: current state variables u(t), such as surface displacement, groundwater pressure, stress intensity, etc.;

[0154] Current time t;

[0155] Critical point prediction value t c

[0156] As an option, the state input can be represented by a feature vector of the state variable u(t). For example, key features can be extracted by a dimensionality reduction algorithm (such as PCA principal component analysis or autoencoder) to reduce computational complexity.

[0157] Action a t :

[0158] Whether to send out an alarm signal at the current time t. Specifically defined as a binary action:

[0159] a t =1: send out an alarm signal;

[0160] a t =0: No alarm signal is issued.

[0161] Reward function R t :

[0162] The reward function is used to measure the contribution of the current action to the warning target and is defined as follows:

[0163] R t =-α(t p -t c ) 2 +β·A

[0164] Among them, t p Indicates the warning timing, that is, the time when the alarm is triggered; t c It represents the critical point of disaster instability; α is the weight coefficient of time deviation, which is used to control the influence of alarm advance; β is the weight coefficient of accuracy; A is the accuracy evaluation value of alarm.

[0165] As a possible implementation method, the alarm accuracy A can be calculated according to the following formula:

[0166]

[0167] Where γ is the smoothing factor, which controls the speed of accuracy decay.

[0168] In one possible implementation, the Proximal Policy Optimization (PPO) algorithm is used to optimize the warning strategy. PPO is a deep reinforcement learning algorithm that is optimized through the policy gradient method and can effectively solve the complex problem of continuous states and discrete actions in disaster warning scenarios.

[0169] Specifically, the reinforcement learning policy network π θ (a t |s t ) is defined as a deep neural network, inputting the current state s t , output alarm action a t The training goal of the policy network is to maximize the cumulative reward function G:

[0170]

[0171] Among them, γ is a discount factor used to balance short-term and long-term rewards.

[0172] In this embodiment, the structure of the policy network may adopt a multi-layer perceptron (MLP), which specifically includes:

[0173] Input layer: receiving disaster status features s t ;

[0174] Hidden layer: A fully connected layer with ReLU activation function is used to extract deep features of the disaster status;

[0175] Output layer: Use Softmax function to output alarm action a t probability.

[0176] As an option, in order to further improve the performance of the policy network, a long short-term memory network (LSTM) can be introduced to process the time series data of disaster evolution. For example:

[0177] As an option, in order to further improve the performance of the policy network, a long short-term memory network (LSTM) can be introduced to process the time series data of disaster evolution.

[0178] Input the time series of disaster evolution path {u(t 0 ),u(t 1 ),…,u(t k )}

[0179] LSTM is used to extract sequence features and capture the dynamic evolution of disaster status.

[0180] In some embodiments, in order to speed up the training process of reinforcement learning, policy pre-training can be combined with simulation data. Specifically, the disaster evolution path and critical points under different parameters are generated through the disaster dynamics model as offline training data of the environment. In addition, real-time monitoring data can be used to fine-tune the policy network online, thereby improving the system's adaptability to actual scenarios.

[0181] In order to verify the effectiveness of the early warning strategy, this embodiment designs the following verification process:

[0182] The early warning strategy is evaluated offline through disaster simulation data, and the alarm time t is calculated. p Deviation distribution and cumulative reward value of ;

[0183] Simulate mining geological disasters in a laboratory environment and verify the triggering timing of alarm signals through real-time monitoring data from sensors;

[0184] Compare the warning effects under different weight parameters α and β, and select the optimal parameter combination.

[0185] Through the reinforcement learning algorithm design of this step, this embodiment can dynamically select the optimal warning time t p , and effectively balance the advance time and accuracy of the alarm, providing a scientific decision-making basis for disaster risk prevention and control. Because the reinforcement learning model can continuously learn from the environment, its adaptability and robustness are significantly better than traditional rule-based early warning strategies.

[0186] S5. Integrate real-time monitoring data with disaster models to dynamically update disaster status and alarm signals

[0187] Step S5 aims to achieve dynamic integration of real-time monitoring data and disaster dynamics models, ensuring that the model can timely adjust the prediction results and alarm signals of disaster status according to changes in the actual mining environment. Through real-time feedback of monitoring data, this step effectively solves the problem of inconsistency between the model and the actual scene in the disaster prediction process, and further improves the accuracy and reliability of disaster warning. Step S5 is closely related to steps S1 to S4. The realization of the aforementioned constructed dynamic model, numerical solution method and optimal warning strategy all rely on the real-time data support provided by this step.

[0188] Generally speaking, the monitoring data of mining geological disasters include multiple sources, such as surface displacement, groundwater pressure, crack expansion rate, rainfall, etc. By dynamically integrating the above data and models, online updates of the evolution status of disasters and real-time optimization of alarm signals can be achieved.

[0189] In this embodiment, real-time monitoring data is collected through a multi-source sensor network, including but not limited to surface displacement sensors, groundwater level monitors, stress sensors, and rain gauges. In order to achieve dynamic fusion of data, this embodiment adopts the following process:

[0190] Data collection and preprocessing:

[0191] Real-time monitoring data 1 (t),d 2 (t),…,d n (t)} is obtained by different sensors at time t, where n is the number of sensors. The monitoring data of each sensor is usually heterogeneous and noisy, so the following preprocessing operations are required:

[0192] Data interpolation: For sensor data with inconsistent sampling frequencies, the time axis can be unified through linear interpolation or spline interpolation methods.

[0193] Data filtering: A low-pass filter is used to remove high-frequency noise and retain the main features of the disaster signal.

[0194] Data normalization: Normalize data of different dimensions to the interval [0,1] for subsequent model calculations.

[0195] As an option, data with missing values ​​can be filled in through historical data interpolation or prediction methods based on machine learning. For example, the autoregressive moving average model (ARIMA) based on time series is used to predict missing data points to ensure the continuity of monitoring data.

[0196] Dynamic calibration of model parameters:

[0197] In this embodiment, the core parameters of the kinetic model (such as the external driving force term The diffusion coefficient matrix D) needs to be dynamically updated according to the real-time monitoring data. Specifically, the least squares method (LeastSquaresMethod) or Bayesian parameter estimation method is used to invert the model parameters. A possible implementation formula is given below:

[0198]

[0199] in, is the objective function, which represents the error between the actual monitoring data and the model prediction value; d i (t) is the monitoring data of the i-th sensor; is the predicted value of the model based on parameter Θ; Θ is the set of model parameters to be inverted.

[0200] By optimizing the above objective function and dynamically calibrating the model parameters, the kinetic model can reflect the latest changes in the monitoring data.

[0201] Real-time updates on the disaster status:

[0202] After the model parameters are calibrated, the high-dispersion equations of the dynamic model are re-solved to update the disaster state variable u(t). Specifically, the time step Δt is determined by the sampling frequency of the real-time monitoring data. For example, for monitoring data sampled per second, the time step can be set to Δt = 1 second. The model is time-integrated in combination with the implicit Euler method or the Runge-Kutta method to obtain the updated disaster state.

[0203] As a possible implementation method, in order to improve computing efficiency, multi-threaded computing or GPU acceleration technology can be introduced into the updating process of state variables, especially in the case of large-scale monitoring data.

[0204] Dynamic optimization of alarm signals:

[0205] Real-time updated disaster status u(t) and monitoring data {d 1 (t),d 2 (t),…,d n (t)} is used as the input of the reinforcement learning strategy network to dynamically adjust the alarm action a t Specifically, the policy network recalculates the probability distribution π of the alarm signal based on the latest features of the disaster state. θ (a t )|s t ), select the optimal alarm time t p .

[0206] In one possible implementation, a sliding window technique can be introduced to comprehensively consider the state change trend of the last m time steps. For example, at the current time t, the data in the sliding window is:

[0207] {u(t-m+1),u(t-m+2),…,u(t)}

[0208] By analyzing the changes in disaster status within the sliding window, the triggering logic of the alarm signal can be further optimized.

[0209] In this embodiment, the dynamic optimization of the alarm signal also includes real-time evaluation of the risk level. The risk level R is based on the critical point t c The time difference between the predicted value and the current time t is calculated. The specific formula is as follows:

[0210]

[0211] Among them, λ is an adjustment parameter used to control the time sensitivity of the risk level.

[0212] The output of risk level R is not only used to trigger the alarm signal, but also can provide managers with an intuitive assessment of the severity of the disaster.

[0213] Integration of models and data:

[0214] Generally, the integration of models and monitoring data is achieved through the Internet of Things platform. Specifically, sensor data is transmitted to the data center in real time through LoRa or 5G networks. The data center runs the disaster model through high-performance computing nodes and sends alarm signals to mobile terminals or management platforms.

[0215] As an option, edge computing technology can be further introduced to move part of the model calculation to the sensor node. For example, preliminary data processing and model parameter estimation can be performed on the sensor side, and only key results can be uploaded to the data center to reduce data transmission delays and bandwidth usage.

[0216] Through the implementation of this step, this embodiment can dynamically update the disaster dynamics model in combination with real-time monitoring data to ensure the real-time and reliability of the prediction results. At the same time, the dynamic optimization of the alarm signal combined with the reinforcement learning strategy enables the system to quickly respond to changes in the mine environment and provide accurate support for disaster risk prevention and control. The implementation of this step provides the entire disaster warning system with powerful real-time adaptability and efficient computing performance.

[0217] S6. Output alarm signals according to the optimal early warning strategy, and provide risk levels and emergency suggestions.

[0218] Step S6 is the final link of the entire disaster warning method. This step relies on the early warning strategy of the dynamic model, numerical solution, critical point identification and reinforcement learning optimization constructed in the previous steps, combines real-time monitoring data and the latest model calculation results, outputs disaster alarm signals, and evaluates the current risk level, thereby providing scientific emergency response suggestions for mine managers. The implementation of this step ensures a complete closed loop of the disaster warning system from dynamic modeling to alarm output.

[0219] In general, the triggering of the alarm signal needs to comprehensively consider the current disaster status, the predicted critical point time t c At the same time, the generation of emergency recommendations needs to combine historical disaster handling experience and current monitoring data to provide managers with actionable guidance.

[0220] In this embodiment, the generation of the alarm signal is based on the reinforcement learning strategy network π θ (a t |s t ) output. Specifically, the policy network is based on the current disaster status s t Calculate alarm action a tThe probability distribution of t =1, the alarm signal is triggered. The triggering logic of the alarm signal is as follows:

[0221] Alarm signal triggering conditions:

[0222] P(a t =1|s t )≥P threshold

[0223] Among them, P(a t =1|s t ) is the probability of triggering an alarm signal, which is calculated by the strategy network; P threshold It is the alarm trigger threshold, which is adjusted according to the specific disaster scenario.

[0224] As an option, in order to avoid frequent triggering of alarm signals, an alarm suppression mechanism can be introduced. When the alarm signal is triggered continuously, the system will determine whether to maintain the alarm state based on the dynamic changes of the monitoring data. For example, if the change amplitude of the monitoring data is lower than the preset threshold, the alarm signal will be automatically cancelled.

[0225] Risk level assessment:

[0226] In this embodiment, the risk level R is based on the current time t, the critical point time t c And the changing trend of the state variable u(t) is calculated. The formula for calculating the risk level is as follows:

[0227]

[0228] Among them, α and β are weight coefficients, which control the influence of time sensitivity and state sensitivity respectively; λ is the time scale factor, which is used to adjust the decay speed of risk level over time; u c It is the critical value of the disaster state, such as the critical threshold of stress intensity.

[0229] Generally, the risk level R is divided into multiple levels, such as low risk, medium risk, high risk and extremely high risk. The risk level classification criteria can be adjusted through experimental data or expert experience.

[0230] As a possible implementation method, the risk level can be further combined with the geographic information system (GIS) for spatial distribution visualization. For example, by drawing a heat map of the risk level of the mining area, the risk distribution of different areas can be intuitively displayed to help managers quickly identify key prevention and control areas.

[0231] Emergency Recommendation Generation:

[0232] The generation of emergency suggestions is based on the risk level R and the current disaster status. Specifically, this embodiment generates emergency suggestions through the following method:

[0233] When R is at high risk or extremely high risk, the system will prioritize generating emergency evacuation recommendations, such as immediately evacuating mine workers and shutting down equipment or facilities that may induce disasters.

[0234] When R is at medium risk, the system will recommend increasing the monitoring frequency and arrange inspectors to check the conditions of key areas.

[0235] When R is at low risk, the system will recommend regular checkups and record monitoring data for subsequent analysis.

[0236] In one possible implementation, the generation of emergency recommendations can also be combined with disaster types and geological characteristics. For example, for slope landslide disasters, the system can recommend the deployment of support equipment in high-risk areas; and for collapse disasters caused by groundwater seepage, the system will prioritize drainage treatment measures.

[0237] Propagation and feedback of alarm signals:

[0238] In this embodiment, the alarm signal is transmitted to relevant managers through multiple channels, including but not limited to text messages, emails, dedicated early warning terminals or mobile applications. At the same time, the system also supports a feedback mechanism, and managers can confirm the receipt of the alarm signal through the feedback terminal and adjust the system's alarm strategy according to actual conditions. For example, when the on-site inspection personnel confirm that the risk has been eliminated, the transmission of the alarm signal can be suspended through the feedback terminal.

[0239] In some embodiments, the triggering of alarm signals and the generation of emergency recommendations can also be integrated with an external decision support system (DSS). For example, by combining a disaster economic loss model, the costs and effects of different emergency response measures can be evaluated to provide managers with a variety of decision-making options.

[0240] As an extension, in order to further improve the reliability of the alarm signal, this embodiment supports multimodal data fusion. For example, sensor data, remote sensing data and historical disaster data are combined to dynamically adjust the trigger threshold P of the alarm signal. threshold and risk level classification standards.

[0241] Through the implementation of this step, the present invention realizes a closed loop of the entire process from disaster modeling to alarm signal output. The dynamic triggering of alarm signals and the real-time assessment of risk levels can significantly improve the accuracy of early warnings, and the generation of emergency recommendations provides a scientific basis for decision-making for mine managers. This step is closely linked to the previous steps, forming a complete mine geological disaster early warning technology system.

[0242] Please refer to the attached Figure 2The present invention also provides an artificial intelligence-based mining geological disaster early warning system. The specific implementation methods of each step are described below in conjunction with the workflow of the system of the present invention.

[0243] Data acquisition module, used to collect real-time monitoring data of the mine, including surface displacement, crack width, groundwater level and rainfall;

[0244] Modeling module, used to construct the dynamic model of mining geological hazards, including the time derivative, spatial diffusion and external driving force terms of state variables;

[0245] The calculation module is used to calculate the disaster evolution path using numerical solution methods and identify the critical point of instability based on bifurcation theory and Lyapunov stability theory;

[0246] Reinforcement learning module, used to optimize the timing of alarms based on disaster evolution paths and critical points;

[0247] The alarm module is used to output alarm signals according to the optimal early warning strategy and provide risk levels and emergency suggestions;

[0248] The model update module is used to calibrate model parameters in combination with real-time data and update disaster status and alarm signals.

[0249] The system of this embodiment can be used to execute the above method embodiments, and its principles and technical effects are similar, which will not be repeated here.

[0250] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A mine geological disaster early warning method based on artificial intelligence, characterized in that: The following steps are involved: Constructing a dynamic model of mining geological hazards, which describes the state variables of the hazard evolution process and its evolutionary behavior under the influence of external driving forces and spatial diffusion; Use numerical methods to discretize and solve the dynamic model to obtain the disaster evolution path; The stability of the disaster system is analyzed based on the bifurcation theory, and the critical point of instability of the disaster is identified by calculating the eigenvalue of the Jacobian matrix of the state equation; Use reinforcement learning algorithms to design optimal warning strategies and dynamically select the optimal warning time based on the disaster evolution path and critical points; Integrate real-time monitoring data with disaster models to dynamically update disaster status and alarm signals; Output alarm signals based on the optimal early warning strategy, and provide risk levels and emergency suggestions.

2. The method for early warning of mining geological disasters based on artificial intelligence according to claim 1 is characterized in that: The dynamic model is described by partial differential equations, which include time variation terms of state variables, spatial diffusion terms and external driving force terms, in the form of: the sum of the time derivative of the state variable and the function of the external driving force and the function of the spatial diffusion effect is zero, and the initial conditions and boundary conditions are determined by field monitoring data.

3. The mine geological disaster early warning method based on artificial intelligence according to claim 1 is characterized in that: The disaster evolution path is obtained by a numerical solution method, and the numerical solution method includes: The finite element method is used to spatially discretize the dynamic model; The time integration algorithm is used to dynamically solve the discretized state equation to obtain the change of the disaster evolution path over time.

4. The mine geological disaster early warning method based on artificial intelligence according to claim 1 is characterized in that: The critical point is identified by bifurcation theory, which includes the following steps: Linearize the disaster evolution state equation and construct the Jacobian matrix; The eigenvalues ​​of the Jacobian matrix are calculated and the instability critical point of the system is identified when the real part of the eigenvalue becomes zero.

5. The method for early warning of mining geological disasters based on artificial intelligence according to claim 1 is characterized in that: The stability of the disaster system is further verified by Lyapunov stability theory, including the following steps: A Lyapunov function of the disaster system is defined, wherein the Lyapunov function includes a kinetic energy term and an external potential energy term of a system state variable; The time derivative of the Lyapunov function is calculated and the critical state of the system is identified when the derivative of the Lyapunov function is zero.

6. The mine geological disaster early warning method based on artificial intelligence according to claim 1 is characterized in that: The reinforcement learning algorithm comprises the following steps: Define the time series of disaster status as the state input of reinforcement learning, and define whether to issue an alarm signal as the action output; Set the reward function and calculate the reward value based on the deviation between the alarm signal and the critical point time and the warning accuracy; The reinforcement learning algorithm is used to optimize the strategy network and select the optimal alarm signal output.

7. The mine geological disaster early warning method based on artificial intelligence according to claim 1 is characterized in that: The reinforcement learning algorithm adopts the Proximal Policy Optimization algorithm. The policy network outputs the probability distribution of the alarm signal by inputting the characteristic vector of the disaster evolution state.

8. The method for early warning of mining geological disasters based on artificial intelligence according to claim 1 is characterized in that: The model is integrated with real-time monitoring data, including the following steps: Obtain real-time monitoring data from field sensors, including ground surface displacement, crack growth rate, groundwater level, and rainfall; Use real-time data to calibrate the parameters of the dynamic model and dynamically update the disaster evolution path; Re-evaluate system stability based on real-time monitoring data and update alarm signals.

9. The method for early warning of mining geological disasters based on artificial intelligence according to claim 1, characterized in that: The alarm signal output includes risk level assessment information, and the risk level is determined by comprehensively evaluating the time distance to the disaster critical point and the stability of the current state.

10. A mine geological disaster early warning system based on artificial intelligence, applied to the method according to any one of claims 1 to 9, characterized in that: include: Data acquisition module, used to collect real-time monitoring data of the mine, including surface displacement, crack width, groundwater level and rainfall; Modeling module, used to construct the dynamic model of mining geological hazards, including the time derivative, spatial diffusion and external driving force terms of state variables; The calculation module is used to calculate the disaster evolution path using numerical solution methods and identify the critical point of instability based on bifurcation theory and Lyapunov stability theory; Reinforcement learning module, used to optimize the timing of alarms based on disaster evolution paths and critical points; The alarm module is used to output alarm signals according to the optimal early warning strategy and provide risk levels and emergency suggestions; The model update module is used to calibrate model parameters in combination with real-time data and update disaster status and alarm signals.

Images