A method and system for dynamic monitoring and early warning of mine environment

Abstract

The present invention relates to the technical field of mine safety management, and discloses a method and system for dynamic monitoring and early warning of mine environment, including the following steps: multi-source data acquisition, collecting dynamic data of mine environment through remote sensing imaging equipment, ground sensors and meteorological stations; data preprocessing, noise cleaning, time alignment and spatial reference coordinate unification of the data; spatiotemporal data fusion, mapping multi-source data to a unified feature space through feature extraction, cross-modal fusion and attention mechanism; intelligent analysis and disaster prediction, predicting surface deformation, landslide risk and tailings dam stability in mining areas based on multi-task joint learning models; risk assessment and early warning response, dividing risk levels according to comprehensive risk scores, and dynamically generating early warning information, evacuation paths and dangerous area isolation boundaries. The present invention solves the problems of difficulty in multi-source data fusion, low disaster prediction accuracy and delayed early warning response in the prior art.

Description

Technical Field

[0001] The present invention relates to the technical field of mine safety management, and in particular to a method and system for dynamic monitoring and early warning of a mine environment. Background Art

[0002] Dynamic monitoring and early warning of mine environment is an important part of mine safety production management, and is of great significance for preventing geological disasters, protecting the lives and property of mining workers, and promoting the sustainable development of mine environment. With the deepening of mining activities, the risk of disasters such as surface deformation, landslides, and tailings dam failures has gradually increased. How to achieve dynamic monitoring and early warning of mine environment through technical means, timely detect risks and take countermeasures has become the focus and difficulty of research in this field.

[0003] At present, dynamic monitoring of mines relies on multi-source data collection, including remote sensing images, ground sensors, meteorological stations and other data sources. These data have significant heterogeneity and differences in spatiotemporal distribution. For example, remote sensing images have high spatial resolution but a long sampling period; ground sensor data have a high sampling frequency but limited coverage; meteorological data have a large regional coverage but insufficient precision. The differences in format, resolution and time scale of these data make it difficult for existing technologies to achieve efficient cross-modal data fusion. In addition, since most data are stored in independent systems, the existing monitoring system lacks a unified data management and collaborative processing mechanism, and the problem of information islands is prominent, which cannot meet the monitoring needs in the complex dynamic environment of mines.

[0004] In terms of disaster prediction, existing technologies are mainly based on single-task models, such as using geomechanical models to predict landslide risks, predicting surface deformation based on time series analysis, or evaluating the stability of tailings dams through physical models. These methods are mostly designed for a single type of disaster, and it is difficult to achieve collaborative prediction of multiple disaster tasks. In addition, existing models have a low utilization rate of multi-source data, often relying on a single data source, and fail to comprehensively utilize the complementary advantages of remote sensing images, sensors, and meteorological data. These deficiencies have led to obvious limitations in the coverage, accuracy, and applicability of existing technologies in disaster prediction, especially in complex mining environments and extreme weather conditions, and the reliability of prediction results is difficult to guarantee.

[0005] In terms of early warning response, existing technologies mostly trigger warnings based on fixed thresholds and rely on static plans for emergency response. This method is difficult to dynamically adapt to changes in environmental conditions, resulting in a strong lag in early warnings and an inability to quickly respond to sudden disasters. In addition, existing early warning systems are usually designed for a single disaster and lack the ability to coordinate the management of multiple disasters. In sudden, chain disasters (such as landslides induced by heavy rains and tailings dam leakage), the response efficiency and accuracy of existing systems are difficult to meet actual needs.

[0006] Therefore, the present invention proposes a method and system for dynamic monitoring and early warning of mine environment to solve the deficiencies of the prior art. Summary of the invention

[0007] In view of the shortcomings of the existing technology, the present invention provides a method and system for dynamic monitoring and early warning of mine environment. Through the efficient fusion of multi-source data, intelligent disaster prediction and dynamic early warning response, a modular and highly adaptable mine dynamic monitoring platform is constructed to achieve accurate monitoring and efficient early warning of geological disasters in the complex dynamic environment of mines.

[0008] To achieve the above objectives, the present invention is implemented through the following technical solutions: A method for dynamic monitoring and early warning of mine environment, comprising the following steps:

[0009] Multi-source data collection: Dynamic data of the mine environment is collected through remote sensing imaging equipment, IoT sensors and meteorological station equipment, including remote sensing imaging data, ground sensor data and meteorological data;

[0010] Data preprocessing: cleaning, time alignment and spatial reference coordinate unification of the data;

[0011] Spatiotemporal data fusion: Map multi-source data into a unified feature space through feature extraction, spatiotemporal alignment, and cross-modal fusion algorithms;

[0012] Intelligent analysis and disaster prediction: Based on a multi-task joint learning model, predict surface deformation, landslide risk and tailings dam stability in mining areas;

[0013] Risk assessment and early warning response: Based on the comprehensive risk score, the risk level is divided and the corresponding early warning measures and emergency response strategies are triggered.

[0014] Preferably, the multi-source data collection includes:

[0015] Obtain optical and SAR image data of the mining area through remote sensing imaging equipment, with a resolution of 10 meters and a sampling period of daily or weekly levels;

[0016] The geological dynamic change data are collected through inclinometers, fracture extension meters and gas sensors, with a sampling frequency of seconds and a data format of time series;

[0017] The rainfall, wind speed, temperature and humidity data are collected through the meteorological station, and the sampling time interval is hourly.

[0018] Preferably, the data preprocessing includes

[0019] Adaptive wavelet denoising method is used to clean the noise of sensor data, and its expression is:

[0020]

[0021] in:

[0022] S(t): original sensor signal at sampling time t; ψ k : kth order component of wavelet basis function; λ k : noise threshold; I(·): indicator function, which takes the value of 1 when the condition is met, otherwise it takes the value of 0; S′(t): signal after noise removal;

[0023] The time series data is time aligned by linear interpolation method, and the expression is:

[0024]

[0025] in:

[0026] t: Alignment target time; t i , t i+1 : adjacent original sampling time; S(t i ), S(t i+1 ): the corresponding original data value; S′(t): the interpolated value after alignment;

[0027] The spatial reference system of remote sensing images and sensor data is unified, and the shadow transformation formula used is:

[0028]

[0029] in:

[0030] x, y: original space coordinates; x′, y′: unified space coordinates after transformation, T: projection transformation matrix.

[0031] Preferably, the feature extraction in the spatiotemporal data fusion includes:

[0032] Use multi-scale convolutional network to extract remote sensing image spatial features, the expression is:

[0033] F R =ReLU(W c *R+b)

[0034] in:

[0035] R: input remote sensing image data; W c : convolution kernel; b: bias; ReLU(·): activation function, defined as max(0,x); F R : Extracted spatial features;

[0036] Use a bidirectional long short-term memory network to extract temporal dynamic features, the expression is:

[0037] h t =BiLSTM(St ,W t )

[0038] in:

[0039] S t : sensor data at time t; W t : meteorological data at time t; h t : Dynamic features extracted at the corresponding moment.

[0040] Preferably, the cross-modal fusion of the spatiotemporal data fusion is achieved by weighting the attention mechanism, and the formula is:

[0041] F=α R F R +α S F S +α W F W

[0042] in:

[0043] F R : Remote sensing characteristics; F S : Sensor characteristics; F W : Meteorological characteristics;

[0044] α R ,α S ,α W : The dynamic weight of the corresponding feature is calculated by the following formula:

[0045]

[0046] in:

[0047] W k : feature fusion weight matrix; F k : The characteristics of the corresponding mode.

[0048] Preferably, the intelligent analysis and disaster prediction includes using the Transformer model to predict surface deformation, and its core formula is:

[0049] y t = Transformer(F t )

[0050] Among them, F t For fusion features;

[0051] The graph neural network is used to model the spatial characteristics of the landslide area, and its expression is:

[0052] H (l+1) =σ(D -1 AH (l) W(l) )

[0053] in:

[0054] H (l) : The lth layer node features; A: adjacency matrix; D: degree matrix; W (l) : weight matrix; σ(·): activation function.

[0055] Preferably, the multi-task joint learning model is adjusted by the following optimization objectives

[0056] L total =λ1L 形变 +λ2L 滑坡 +λ3L 坝稳定

[0057] in:

[0058] L 形变 , L 滑坡 , L 坝稳定 : are the loss functions of deformation prediction, landslide risk and tailings dam stability respectively; λ1, λ2, λ3: the weights of the corresponding tasks.

[0059] Preferably, the risk assessment calculates a comprehensive risk score by the following formula.

[0060]

[0061] in:

[0062] R k : risk score of the kth subtask; ω k : The corresponding weights are adjusted based on the dynamic optimization strategy.

[0063] 9. A method for dynamic monitoring and early warning of a mine environment according to claim 1, characterized in that the early warning response comprises:

[0064] Through the path planning formula:

[0065]

[0066] where d i The distance between path nodes is used to plan the optimal evacuation path;

[0067] The high-risk area isolation boundary is dynamically generated using the following formula:

[0068] B={x,y|R(x,y)≥τ 隔离}

[0069] in:

[0070] x, y: spatial coordinates within the danger zone;

[0071] R(x,y): risk score of location (x,y);

[0072] τ 隔离 : Risk score threshold of the dangerous area, used to determine whether isolation is required.

[0073] Preferably, a mine environment dynamic monitoring and early warning system comprises:

[0074] Data acquisition module:

[0075] Obtain optical images and SAR images through remote sensing imaging equipment;

[0076] Collect ground sensor data through inclinometers, gas sensors, and crack extension meters;

[0077] Obtain rainfall, wind speed, temperature and humidity information through weather stations;

[0078] Data preprocessing module:

[0079] Perform noise cleaning of sensor data, remote sensing image restoration, time alignment and spatial coordinate unification;

[0080] Space-time fusion module:

[0081] Includes multi-scale convolutional networks, bidirectional long short-term memory networks, and attention mechanisms for extracting and fusing remote sensing, sensor, and meteorological features;

[0082] Intelligent analysis module:

[0083] Integrate Transformer models, graph neural networks, and physical models for surface deformation prediction, landslide risk assessment, and tailings dam stability analysis;

[0084] Risk assessment and early warning response module:

[0085] Trigger warnings through risk score calculation and grading;

[0086] Plan evacuation routes, dynamically define isolation boundaries for high-risk areas, and automatically issue early warning information.

[0087] The present invention provides a method and system for dynamic monitoring and early warning of mine environment, which has the following beneficial effects:

[0088] 1. The present invention innovatively adopts multi-source heterogeneous data fusion technology based on multimodal deep learning, combined with multi-scale convolutional neural network (MSCNN), bidirectional long short-term memory network (BiLSTM) and attention mechanism, to extract features and dynamically fuse remote sensing image data, ground sensor data and meteorological data, and solve the inconsistency of data in time sampling frequency, spatial resolution and format through spatiotemporal alignment algorithm, and construct a unified feature space. This technical solution realizes the efficient fusion of dynamic consistency of multi-source heterogeneous data, and significantly improves the accuracy and coordination of monitoring data. Compared with the existing technology that only relies on a single data source (such as remote sensing image or sensor data), resulting in limited monitoring range and insufficient ability to capture dynamic changes, the present invention solves the problem of difficulty in fusion of multi-source heterogeneous data and low data utilization. Especially in complex mining environments, the present invention not only significantly improves the monitoring accuracy through multi-modal data collaborative processing, but also broadens the monitoring range, providing a complete solution for dynamic monitoring of complex mining scenes.

[0089] 2. The present invention adopts a multi-task joint learning model, combining Transformer, graph neural network (GNN) and hybrid physical model, and performs collaborative modeling for three disaster prediction tasks: surface deformation, landslide risk and tailings dam stability. Transformer captures the long-term dependencies of time series data, GNN models the spatial correlation between sensor nodes, and combines physical models such as finite element analysis to improve the interpretability of predictions, significantly improving the accuracy and adaptability of disaster predictions. Compared with the single-task model in the prior art, which can usually only predict a single disaster, and the problem of reduced prediction accuracy due to insufficient model adaptability in multi-disaster scenarios, the present invention realizes collaborative prediction of multiple disasters through a multi-task learning framework, solving the technical bottleneck of narrow coverage and insufficient prediction accuracy of traditional models. At the same time, the hybrid physical model scheme of the present invention shows stronger robustness and generalization ability in complex environments, providing more accurate support for mine disaster risk management.

[0090] 3. The present invention proposes an early warning mechanism based on dynamic comprehensive risk scoring, dynamically adjusts the weights of multimodal features through the attention mechanism, and combines real-time risk level division and dynamic dangerous area isolation algorithm to achieve dynamic and automated management from risk assessment to early warning response. The early warning response module generates the isolation boundary of the dangerous area in real time according to the comprehensive risk score, optimizes the evacuation path, and pushes early warning information through multiple channels to ensure the timely protection of personnel and equipment. Compared with the early warning method that relies on fixed thresholds and static models in the prior art, it is difficult to adapt to the dynamic environmental changes of mines, and the early warning has strong lag and insufficient response capabilities. The present invention solves the shortcomings of traditional methods in dynamic adaptability and response efficiency by adjusting risk scores and early warning strategies in real time. Especially in sudden disasters (such as landslides induced by heavy rains or tailings dam leakage), the present invention can respond quickly, significantly improve the timeliness and accuracy of early warnings, and reduce the loss of personnel and property caused by disasters.

[0091] 4. The present invention adopts a modular design, and divides the system into five modules: data acquisition, preprocessing, spatiotemporal fusion, intelligent analysis, and early warning response. It supports multi-source data access, adapts to different mining environments and disaster scenarios, and significantly improves the applicability and flexibility of the system. At the same time, the modular architecture of the present invention supports the seamless integration of a variety of acquisition equipment (such as remote sensing satellites, drones, ground sensors) and analysis models (such as deep learning models and physical models) to meet the diverse needs of dynamic monitoring of mines. Compared with the existing technology, the monitoring system has a single function, poor scalability, and can only be applied to specific mine types or specific disasters. The restrictive problem, the present invention solves the technical bottleneck of insufficient system versatility and adaptability through modular design, and significantly improves the system's application capabilities in different mine types and different disaster scenarios. At the same time, the modular design of the present invention can be flexibly expanded or adjusted according to specific needs, reducing the development and deployment costs of the system, and providing a solution with high economic benefits and high practicality for mine safety monitoring. BRIEF DESCRIPTION OF THE DRAWINGS

[0092] Figure 1 This is a flow chart of the method for dynamic monitoring and early warning of mine environment. DETAILED DESCRIPTION

[0093] The technical solution of the present invention will be described clearly and completely below in conjunction with the accompanying drawings of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0094] Please refer to the attached Figure 1 :

[0095] The embodiment of the present invention provides a method for dynamic monitoring and early warning of a mine environment, comprising the following steps:

[0096] Multi-source data collection: Dynamic data of the mine environment is collected through remote sensing imaging equipment, IoT sensors and meteorological station equipment, including remote sensing imaging data, ground sensor data and meteorological data;

[0097] In this embodiment, the specific technical implementation of multi-source data collection is as follows:

[0098] Remote sensing image data acquisition

[0099] The surface morphology and dynamic change data of a large area of ​​the mining area are obtained through satellite remote sensing (such as Sentinel-1, Landsat) and drone equipment.

[0100] Data format: Image data is in raster format with a fixed spatial resolution (e.g. 10 meters) and sampling period (e.g. daily or weekly).

[0101] Technical implementation:

[0102] Satellite remote sensing images use optical and synthetic aperture radar (SAR) technologies to obtain information on surface vegetation destruction, terrain elevation changes, and deformation, respectively;

[0103] The drone uses LiDAR (laser radar) technology to construct a three-dimensional point cloud model, providing high-precision three-dimensional data supplements for the areas around the tailings dam and areas prone to landslides.

[0104] Application scenarios of remote sensing images:

[0105] Capture the overall surface deformation of the landslide area;

[0106] Provide topographic changes of the tailings dam and surrounding areas.

[0107] Ground sensor data collection:

[0108] A ground sensor network is deployed in key areas of the mining area (such as high-risk areas for landslides and around tailings dams) to collect dynamic environmental data in real time.

[0109] Data Type:

[0110] Inclinometer: monitors changes in surface and slope inclination;

[0111] Crack expansion meter: collects the crack expansion rate of rock or dam body;

[0112] Gas detection sensor: measure the concentration of harmful gases in mining areas (such as methane);

[0113] Seismic sensors: Detect microseismic activity in the area.

[0114] Data format: The data collected by the sensor is in time series format with a high sampling frequency (such as seconds).

[0115] The inclinometer data measures the inclination change in real time through the built-in accelerometer and gyroscope. The data format is S 侦 (t) = (θ x (t),θ y (t)), where θ x ,θ y are the inclination angles in the x and y directions respectively; the crack extension instrument records the change in crack width by laser ranging or strain measurement:

[0116] S 裂 (t) = d(t)

[0117] in:

[0118] d(t) is the crack width in millimeters.

[0119] Meteorological data collection:

[0120] The meteorological station collects environmental parameters around the mining area, including rainfall, wind speed, temperature and humidity, to analyze the dynamic impact of the external environment on the mining area.

[0121] Data Type:

[0122] Rainfall: W 降 (t), in mm / h;

[0123] Wind speed: W 风 (t)=(v x (t),v y (t)), unit is meter / second; temperature and humidity: W 温 (t),W 显 (t), the units are degrees Celsius and percentage respectively.

[0124] Technical implementation:

[0125] The weather station collects data in real time through barometers, rain gauges, and temperature and humidity sensors. The rainfall intensity is calculated using the following formula

[0126]

[0127] in:

[0128] Δt is the sampling time interval.

[0129] Data collection network architecture

[0130] Network transmission method:

[0131] Use low-power wide area networks (LoRa or NB-IoT) for sensor data transmission to ensure low power consumption and long-distance coverage;

[0132] 5G technology is used to transmit remote sensing images and drone image data to meet high bandwidth requirements.

[0133] Edge computing assists:

[0134] Edge computing units are deployed at the collection terminals to perform preliminary data processing (such as format conversion and noise reduction), reduce the computing pressure of the central system and improve real-time performance.

[0135] Data storage and management:

[0136] Use distributed storage systems (such as HDFS) to manage multi-source data and support classified storage of different data formats (such as images, time series data, and structured data).

[0137] Data sampling formula:

[0138] For sensor data, the sampling point t i The value of is determined by the following formula:

[0139] S(t i )=f 传感器 (x,y,t i )

[0140] where f 传感器 is the sensor output function, (x, y) is the coordinates of the sensor location.

[0141] Preliminary structuring of multimodal data:

[0142] Data from different data sources are initially structured by adding timestamp T and geographic location (x, y):

[0143]

[0144] in:

[0145] N is the number of data collection nodes.

[0146] This step achieves comprehensive coverage of the macro and micro dynamic information of the mining area through multi-source collaborative collection of remote sensing images, sensors and meteorological data. The real-time and reliability of data collection are guaranteed through the optimized design of the collection network and the introduction of edge computing, which lays the foundation for subsequent data preprocessing and spatiotemporal data fusion.

[0147] Data preprocessing: cleaning, time alignment and spatial reference coordinate unification of the data;

[0148] In this embodiment, the specific technology of data preprocessing is implemented as follows:

[0149] Data cleaning:

[0150] Sensor data noise removal:

[0151] During the collection process, sensor data may be interfered by environmental factors (such as temperature and humidity changes) or equipment failures, generating high-frequency noise. This embodiment uses an adaptive wavelet denoising method to process sensor data, and the formula is as follows:

[0152]

[0153] in:

[0154] S(t): original sensor signal at sampling time t; ψ k : kth order component of wavelet basis function; λ k : noise threshold; I(·): indicator function, which takes the value of 1 when the condition is met, otherwise it takes the value of 0; S′(t): signal after noise removal.

[0155] Liansensing image restoration:

[0156] The pixel missing areas in the remote sensing image due to cloud cover or equipment failure are supplemented by the image interpolation-based repair method. The formula is:

[0157]

[0158] in:

[0159] R′(x,y): pixel value after restoration; The neighborhood of a pixel (x,y); interpolation weights are dynamically assigned based on the similarity of pixels in the neighborhood.

[0160] Time Alignment:

[0161] Multi-source data have different sampling frequencies and inconsistent time steps, so they need to be time-aligned to form a unified time series.

[0162] Time interpolation:

[0163] For sensor and meteorological data, linear interpolation is used to align the data to a uniform time step (such as hourly level). The interpolation formula is:

[0164]

[0165] in:

[0166] t: Alignment target time; t i , t i+1 : adjacent original sampling time; S(ti ), S(t i+1 ): the corresponding original data value; S ′ (t): interpolated value after alignment.

[0167] Remote sensing image time stamping:

[0168] The sampling period of remote sensing images is relatively long. To facilitate the subsequent spatiotemporal data fusion, the image data is labeled and associated with the sensor and meteorological data according to the acquisition time.

[0169] Spatial reference coordinate system:

[0170] Multi-source data may use different coordinate systems or projection methods when collected, and need to be uniformly mapped to the reference geographic coordinate system of the mining area (such as WGS-84).

[0171] Coordinate transformation formula: Remote sensing images and sensor data are unified in spatial coordinates through the projection transformation formula:

[0172]

[0173] in:

[0174] x, y: original space coordinates; x ′ ,y ′ : unified spatial coordinates after transformation, T: projective transformation matrix, calculated by geographic calibration method.

[0175] Remote sensing image projection adjustment:

[0176] Bilinear interpolation is used to adjust the pixel values ​​of remote sensing images:

[0177]

[0178] in:

[0179] R′(x′,y′): pixel value in the target coordinate system; w i,j : Interpolation weight, calculated according to the spatial relationship between the original pixel and the target pixel.

[0180] Data structuring:

[0181] Time and space annotation:

[0182] Add a unified timestamp T and spatial coordinates (x, y) to multi-source data to generate a structured data format:

[0183]

[0184] in:

[0185] D(t): the structured data set at time t; Si : The value of the i-th data point; x i ,y i : corresponding geographic coordinates; T i : timestamp.

[0186] Multi-source data classification storage:

[0187] Data is stored in a distributed file system (such as HDFS) by type (such as images, time rooms, and meteorological data) to facilitate efficient calls for subsequent spatiotemporal data fusion.

[0188] Through data cleaning, the noise of sensor data is effectively removed; through time alignment and spatial coordinate unification, the time steps and geographic reference systems of different data sources are coordinated. Finally, the structured multi-source data is stored in a unified format, providing high-quality input for subsequent spatiotemporal data fusion.

[0189] This step solves the consistency issues of multi-source data in terms of quality, time and space, laying a solid foundation for data fusion and analysis in subsequent steps.

[0190] Spatiotemporal data fusion: Map multi-source data into a unified feature space through feature extraction, spatiotemporal alignment, and cross-modal fusion algorithms;

[0191] In this embodiment, the specific technology of spatiotemporal data fusion is implemented as follows:

[0192] Feature extraction:

[0193] Remote sensing image feature extraction:

[0194] Remote sensing images contain spatial distribution information such as the surface morphology and deformation of the mining area. Spatial features are extracted through multi-dimensional convolutional neural network (MSCNN):

[0195] F R =ReLU(W c *R+b)

[0196] in:

[0197] R: input remote sensing image data; W c : convolution kernel; b: bias; ReLU(·): activation function, defined as max(0,x); F R : Extracted spatial features.

[0198] Ground sensor temporal feature extraction:

[0199] The ground sensor data is in the form of time series, capturing the dynamic changes in the mining area. Its temporal features are extracted through a bidirectional long short-term memory network (BiLSTM):

[0200] ht =BiLSTM(S t ,W t )

[0201] in:

[0202] S t : sensor data at time t; W t : meteorological data at time t; h t : Dynamic features extracted at the corresponding moment.

[0203] Meteorological feature extraction:

[0204] Meteorological data includes the dynamic changes of the external environment on the mining area. By calculating the statistical characteristics and dynamic change rate, meteorological

[0205] feature:

[0206]

[0207] Where: W 降 , W 风 , W 温 , W 湿 : They are rainfall star, wind speed, temperature and humidity respectively.

[0208] Indicates the dynamic rate of change.

[0209] Temporal feature fusion:

[0210] The time characteristics of the sensor and gas machine data are weighted and combined to construct the dynamic time series characteristics of the mining area:

[0211] F T =α T h t +(1-α T )F W

[0212] in:

[0213] F T : The time feature after fusion; α T : Dynamic dual factor, calculated based on time correlation.

[0214] Spatial feature fusion:

[0215] The spatial features extracted from remote sensing images are further integrated with the temporal features to form the final spatiotemporal features.

[0216] F=α R F R +β T F T

[0217] Where: F: fused spatiotemporal features; α R , β T : are the fusion weights of spatial and temporal features respectively.

[0218] Cross-modal fusion mechanism:

[0219] Attention mechanism optimizes fusion weights

[0220] During the fusion process, the weight α R ,β T Dynamically adjust to enhance data coordination. Use the attention mechanism to calculate weights:

[0221]

[0222] in:

[0223] W k : feature weight matrix; F k : input features; α k : The dynamic weight of the k-th feature.

[0224] Generation of feature space:

[0225] Structured feature expression:

[0226] Finally, the fused spatiotemporal features F are mapped to a unified high-dimensional feature space to form a structured representation for subsequent intelligent analysis and prediction.

[0227] The final formula in feature space:

[0228]

[0229] in:

[0230] F k Includes synaesthesia images, sensor time characteristics and meteorological data characteristics.

[0231] Mapping of multimodal data to fusion features: Data of different modalities are mapped to the same feature dimension through a feature extraction network to ensure data consistency and fusion.

[0232] This embodiment uses multimodal feature extraction and fusion technology to achieve consistent expression of multi-source data in the temporal and spatial dimensions, forming comprehensive representation features and providing high-quality auxiliary input data for intelligent analysis. At the same time, through the dynamic expansion rate: appropriate weight mechanism, the coordination and robustness of the data are effectively enhanced.

[0233] Intelligent analysis and disaster prediction: Based on a multi-task joint learning model, predict surface deformation, landslide risk and tailings dam stability in mining areas;

[0234] In this embodiment, the specific technical implementation of intelligent analysis and disaster prediction is as follows:

[0235] Multi-task joint learning architecture:

[0236] Model design:

[0237] This embodiment adopts a multi-task learning architecture, takes surface deformation prediction of the mining area, landslide risk assessment and tailings dam stability analysis as three subtasks, and constructs a unified multi-task prediction model.

[0238] L total =λ1L 形变 +λ2L 滑坡 +λ3L 坝稳定

[0239] in:

[0240] L 形变 , L 滑坡 , L 坝稳定 : are the loss functions for deformation prediction, landslide risk and tailings dam stability respectively; λ1, λ2, λ3: correspond to the weights of the tasks, which are dynamically adjusted to optimize the balance between tasks.

[0241] Surface deformation prediction:

[0242] Model structure:

[0243] The surface deformation prediction adopts a Transformer-based time series prediction model to capture the long-term dependence of surface deformation.

[0244] y t = Transformer(F t )

[0245] in:

[0246] F t : fused spatiotemporal features at time t; y t : Deformation prediction result at time t.

[0247] Input and Output:

[0248] Input: Fusion features including sensor tilt changes and remote sensing image deformation features

[0249] Output: Predicted surface deformation rate.

[0250] Landslide Risk Assessment

[0251] Model structure:

[0252] The landslide risk assessment uses a spatial analysis model based on a graph neural network (GNN) to perform spatial correlation modeling on the sensor network in the landslide area:

[0253] H (l+1) =σ(D -1 AH (l) W (l) )

[0254] in:

[0255] A: The adjacency matrix of the graph, representing the spatial connections between sensor nodes; D: The degree matrix of the graph; H (l) : sensor features of the first layer; W (l) : weight matrix of the lth layer; σ(·): activation function.

[0256] Risk score calculation:

[0257] The landslide risk value is defined as:

[0258]

[0259] Where: H i : Characteristics of the i-th sensor node; N: Total number of sensor nodes.

[0260] Tailings dam stability analysis

[0261] Mixed Model Design:

[0262] The implicit qualitative analysis of tailings dams combines physical models with data-driven models. The physical model is based on finite element analysis, and the deep learning model learns the dam instability characteristics in historical data through neural networks. The prediction formula for the combination of the two is:

[0263] R 坝 =α·R 物理 +(1-α)·R 深度学习

[0264] in:

[0265] R 物理 : Dam stability score calculated by the physical model; R 深度学习 : Stability score of data-driven model prediction; α: Weight factor, dynamically adjusted according to data confidence.

[0266] Dynamic weight optimization:

[0267] Dynamically adjust task weights:

[0268] Use reinforcement learning to optimize the weight parameters λ1, λ2, λ3 of each subtask in multi-task learning:

[0269]

[0270] in:

[0271] η: learning rate; λ: k : The weight of task k.

[0272] Comprehensive evaluation of disaster prediction results:

[0273] Comprehensive scoring formula:

[0274] The prediction results of each subtask are comprehensively evaluated to generate an overall disaster risk score:

[0275]

[0276] in:

[0277] R k : risk score of the kth subtask; ω k : The dynamic weight of the corresponding task.

[0278] Joint learning framework; by dynamically adjusting the weights of subtasks, the model's adaptability in different disaster scenarios is optimized.

[0279] Spatiotemporal feature mapping: Each subtask uses a specific model structure (such as Transformer and GNN) to fully explore spatiotemporal features and improve the accuracy of disaster prediction.

[0280] Physics and data fusion: The combination of physical models and deep learning in tailings dam analysis makes up for the limitations of a single method.

[0281] This embodiment uses joint learning and deep models to comprehensively predict various disaster risks in mining areas, ensuring the breadth and accuracy of the prediction results, and providing reliable basic data support for the next step of risk assessment and early warning response.

[0282] Risk assessment and early warning response: Based on the comprehensive risk score, the risk level is divided and the corresponding early warning measures and emergency response strategies are triggered.

[0283] In this embodiment, the specific technical implementation of risk assessment and early warning response is as follows:

[0284] Overall risk score:

[0285] Rating calculation:

[0286] The comprehensive risk score is a weighted sum of the prediction results of multiple disaster tasks, and the formula is as follows:

[0287]

[0288] in:

[0289] R 综合 : Comprehensive risk score; R k: The single risk score of the kth hazard task (such as landslide, deformation or tailings dam stability); ω k : Dynamically adjusted weights that reflect the importance of each task.

[0290] Dynamic weight optimization:

[0291] Optimize weights ω using attention mechanism k :

[0292]

[0293] in:

[0294] W k : The weight matrix of task k; R k : Risk score vector for the task.

[0295] Risk level classification:

[0296] Classification rules:

[0297] According to the comprehensive risk score, the mining area risks are divided into different levels. The formula is as follows:

[0298]

[0299] in:

[0300] L: risk level (1 to 4, the higher the value, the higher the risk); τ1, τ2, τ3: pre-set risk score thresholds.

[0301] Dynamic Threshold Adjustment:

[0302] Based on historical data and current environmental changes, thresholds are dynamically adjusted to accommodate different mining scenarios and seasonal factors.

[0303] Dynamic early warning trigger mechanism:

[0304] Information Release:

[0305] The corresponding early warning measures are triggered according to the risk level, and the early warning information is released through the following channels:

[0306] SMS and APP push: Send real-time warning information to mine managers and front-line workers;

[0307] Sound and light alarm: Sound and light signals are emitted from the on-site early warning equipment deployed in the mining area.

[0308] Hazardous area isolation:

[0309] Based on the comprehensive risk score, the boundary of the dangerous area is delineated, and the formula is as follows:

[0310] B={(x,y)|R(x,y)≥τ隔离}

[0311] in:

[0312] B: the boundary set of the dangerous area; R(x,y): the risk score of the spatial location (x,y); τ 隔离： The risk score threshold of the dangerous area is used to determine whether it needs to be isolated.

[0313] Dynamic emergency response:

[0314] Personnel evacuation route planning:

[0315] Generate the optimal evacuation path based on the dangerous area and personnel distribution, and the optimization goal is to minimize the evacuation time:

[0316]

[0317] in:

[0318] P: evacuation path; d i : The distance between path nodes.

[0319] Resource scheduling optimization:

[0320] Allocate emergency resources (such as rescue equipment and personnel) according to the risk level, giving priority to high-risk areas.

[0321] Automatic generation of response strategies

[0322] Generate rules:

[0323] Combining risk scores and levels, the system generates emergency response strategies, such as:

[0324] Low risk (L=1): Routine inspection, no emergency measures required;

[0325] Medium risk (L=2, L=3): restrict entry into the danger zone and initiate early warning broadcast;

[0326] High risk (L=4): Full evacuation and mobilization of emergency resources to designated areas.

[0327] Comprehensive score optimization formula:

[0328]

[0329] The comprehensiveness and real-time performance of multi-hazard risk scoring are improved through dynamic weight optimization.

[0330] Formula for dynamic delineation of isolation boundaries:

[0331] B={(x,y)|R(x,y)≥τ 隔离}

[0332] Automatic demarcation of high-risk areas has been achieved.

[0333] This embodiment realizes closed-loop management from risk quantification to specific response strategies through comprehensive risk scoring, dynamic level division and response mechanism, providing efficient technical support for timely warning and rapid response to mine disasters.

[0334] 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 method for dynamic monitoring and early warning of mine environment, characterized in that: The following steps are involved: Multi-source data collection: Dynamic data of the mine environment is collected through remote sensing imaging equipment, IoT sensors and meteorological station equipment, including remote sensing imaging data, ground sensor data and meteorological data; Data preprocessing: cleaning, time alignment and spatial reference coordinate unification of the data; Spatiotemporal data fusion: Map multi-source data into a unified feature space through feature extraction, spatiotemporal alignment, and cross-modal fusion algorithms; Intelligent analysis and disaster prediction: Based on a multi-task joint learning model, predict surface deformation, landslide risk and tailings dam stability in mining areas; Risk assessment and early warning response: Based on the comprehensive risk score, the risk level is divided and the corresponding early warning measures and emergency response strategies are triggered.

2. A method for dynamic monitoring and early warning of mine environment according to claim 1, characterized in that: The multi-source data collection includes: Obtain optical and SAR image data of the mining area through remote sensing imaging equipment, with a resolution of 10 meters and a sampling period of daily or weekly levels; The geological dynamic change data are collected through inclinometers, fracture extension meters and gas sensors, with a sampling frequency of seconds and a data format of time series; The rainfall, wind speed, temperature and humidity data are collected through the meteorological station, and the sampling time interval is hourly.

3. A method for dynamic monitoring and early warning of mine environment according to claim 1, characterized in that: The data preprocessing includes Adaptive wavelet denoising method is used to clean the noise of sensor data, and its expression is: in: S(t): original sensor signal at sampling time t; ψ k : kth order component of wavelet basis function; λ k : noise threshold; I(·): indicator function, which takes the value of 1 when the condition is met, otherwise it takes the value of 0; S′(t): signal after noise removal; The time series data is time aligned by linear interpolation method, and the expression is: in: t: Alignment target time; t i , t i+1 : adjacent original sampling time; S(t i ), S(t i+1 ): the corresponding original data value; S′(t): the interpolated value after alignment; The spatial reference system of remote sensing images and sensor data is unified, and the shadow transformation formula used is: in: x, y: original space coordinates; x′, y′: unified space coordinates after transformation, T: projection transformation matrix.

4. A method for dynamic monitoring and early warning of mine environment according to claim 1, characterized in that: The feature extraction in the spatiotemporal data fusion includes: Use multi-scale convolutional network to extract remote sensing image spatial features, the expression is: F R =ReLU(W c *R+b) in: R: input remote sensing image data; W c : convolution kernel; b: bias; ReLU(·): activation function, defined as max(0,x); F R : Extracted spatial features; Use a bidirectional long short-term memory network to extract temporal dynamic features, the expression is: h t =BiLSTM(S t ,W t ) in: S t : sensor data at time t; W t : meteorological data at time t; h t : Dynamic features extracted at the corresponding moment.

5. The method for dynamic monitoring and early warning of mine environment according to claim 1, characterized in that: The cross-modal fusion of the spatiotemporal data fusion is achieved through weighted attention mechanism, and its formula is: F=a R F R +a S F S +a W F W in: F R : Remote sensing characteristics; F S : Sensor characteristics; F W : Meteorological characteristics; α R , α S , α W : The dynamic weight of the corresponding feature is calculated by the following formula: in: W k : feature fusion weight matrix; F k : The characteristics of the corresponding mode.

6. A method for dynamic monitoring and early warning of mine environment according to claim 1, characterized in that: The intelligent analysis and disaster prediction includes using the Transformer model to predict surface deformation, and its core formula is: y t =Transformer(F t ) where F t For fusion features; The graph neural network is used to model the spatial characteristics of the landslide area, and its expression is: H (l+1) =σ(D -1 AH (l) W (l) ) in: H (l) : The lth layer node features; A: adjacency matrix; D: degree matrix; W (l) : weight matrix; σ(·): activation function.

7. A method for dynamic monitoring and early warning of mine environment according to claim 1, characterized in that: The multi-task joint learning model is adjusted by the following optimization objectives L total =λ1L 形变 +λ2L 滑坡 +λ3L 坝稳定 in: L 形变 , L 滑坡 , L 坝稳定 : are the loss functions of deformation prediction, landslide risk and tailings dam stability respectively; λ1, λ2, λ3: the weights of the corresponding tasks.

8. The method for dynamic monitoring and early warning of mine environment according to claim 1, characterized in that: The risk assessment calculates a comprehensive risk score using the following formula. in: R k : risk score of the kth subtask; ω k : The corresponding weights are adjusted based on the dynamic optimization strategy.

9. A method for dynamic monitoring and early warning of mine environment according to claim 1, characterized in that: The early warning response includes: Through the path planning formula: where d i The distance between path nodes is used to plan the optimal evacuation path; The high-risk area isolation boundary is dynamically generated using the following formula: B={x,y|R(x,y)≥τ 隔离 } in: x, y: spatial coordinates within the danger zone; R(x,y): risk score of location (x,y); τ 隔离 : Risk score threshold of the dangerous area, used to determine whether isolation is required. A mine environment dynamic monitoring and early warning system, used to execute the method described in any one of claims 1 to 9, characterized in that: include: Data acquisition module: Obtain optical images and SAR images through remote sensing imaging equipment; Collect ground sensor data through inclinometers, gas sensors, and crack extension meters; Obtain rainfall, wind speed, temperature and humidity information through weather stations; Data preprocessing module: Perform noise cleaning of sensor data, remote sensing image restoration, time alignment and spatial coordinate unification; Space-time fusion module: Includes multi-scale convolutional networks, bidirectional long short-term memory networks, and attention mechanisms for extracting and fusing remote sensing, sensor, and meteorological features; Intelligent analysis module: Integrate Transformer models, graph neural networks, and physical models for surface deformation prediction, landslide risk assessment, and tailings dam stability analysis; Risk assessment and early warning response module: Trigger warnings through risk score calculation and grading; Plan evacuation routes, dynamically define isolation boundaries for high-risk areas, and automatically issue early warning information.

Images