Goaf collapse risk assessment data fusion system based on big data processing

Abstract

The application discloses a goaf collapse risk assessment data fusion system based on big data processing and concretely relates to the technical field of geological disaster assessment, and comprises three core modules: a multi-source data adaptive weighted fusion module, a unified space-time coordinate system is established, non-raster data is converted into a continuous field through Kriging interpolation, weights are dynamically distributed in combination with information entropy and a correlation coefficient, and enhanced feature fields are generated through self-supervised pre-training; a physical-space-time neural network dynamic prediction module, which is fused with an elastoplastic mechanics constraint loss and multi-task learning, outputs future multi-time step risk probability fields and deformation prediction fields through a space-time convolution-memory network; and a risk field three-dimensional profiling and emergency response module, which clusters three-dimensional body elements of a high-risk area and automatically calculates risk body parameters, generates an emergency scheme in combination with DEM data and an A* algorithm, and improves assessment accuracy and emergency scheme operability through digital twin deduction and evaluation.

Description

Technical Field

[0001] This invention relates to the field of geological hazard assessment technology, and more specifically, to a data fusion system for assessing the risk of subsidence in mining areas based on big data processing. Background Technology

[0002] In current risk assessments of goaf collapse, the fusion processing of multi-source monitoring data has significant limitations. Existing technologies struggle to effectively integrate heterogeneous data from multiple sources, such as InSAR deformation data, microseismic monitoring events, geological attribute parameters (e.g., rock mass strength), and groundwater pressure. Either a lack of a unified spatiotemporal coordinate system leads to inconsistent spatiotemporal benchmarks, or simple interpolation (e.g., linear interpolation) is used for non-grid monitoring point data, failing to accurately convert it into a continuous field. Furthermore, weight allocation is often statically preset, without dynamic adjustment based on data inherent uncertainties (e.g., information entropy) and inter-data synergies (e.g., correlation coefficients), resulting in fusion results that are disconnected from the actual geological scenario and unable to support subsequent accurate risk assessments.

[0003] In the collapse risk prediction stage, existing models generally suffer from the problems of "disconnect between data-driven approaches and physical mechanisms" and "single prediction dimensions." Some models rely solely on data-driven approaches (such as traditional neural networks) without incorporating the deformation coordination laws in elastoplastic mechanics, resulting in prediction results that contradict the physical nature of rock mass instability. Other models focus only on the single objective of risk probability or deformation, ignoring key physical fields such as maximum shear stress (reflecting the shear yielding trend of rock mass) and volumetric strain (related to roof subsidence), failing to comprehensively characterize the collapse evolution process. At the same time, the models lack effective self-supervised pre-training mechanisms, resulting in weak generalization ability in small-sample risk data scenarios and difficulty in adapting to the geological differences of different mining areas.

[0004] Existing emergency response technologies have significant shortcomings in risk body identification and scheme evaluation. On the one hand, the delineation of high-risk areas is mostly limited to two-dimensional planes, without combining three-dimensional geological grids for volumetric clustering. Furthermore, key parameters such as the volume, center of gravity, and equivalent burial depth of risk bodies need to be calculated manually, which is inefficient and prone to errors. On the other hand, emergency plans lack dynamic simulation capabilities after generation. They cannot simulate actual scenarios such as personnel evacuation (e.g., panic, route congestion) and rescue vehicle dispatching based on digital twin environments. They also lack quantitative evaluation indicators such as evacuation efficiency, safety risks, and scheme robustness, resulting in poor operability of the plans and difficulty in coping with complex and sudden collapse disasters. Summary of the Invention

[0005] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide a data fusion system for assessing the risk of goaf collapse based on big data processing.

[0006] To achieve the above objectives, the present invention provides the following technical solution: The data fusion system for risk assessment of goaf collapse based on big data processing includes the following modules: The multi-source data adaptive weighted fusion module is used to establish a unified spatiotemporal coordinate system, rasterize multi-source data of InSAR deformation, microseismic and geological parameters, and convert non-raster data into a continuous field using Kriging interpolation; then, based on information entropy and correlation coefficient, adaptive weights are assigned, and an enhanced feature field is generated through a weighted fusion formula; and the feature quality and model generalization ability are improved through self-supervised pre-training. The physical-spatiotemporal neural network dynamic prediction module is used to organize the fused features into a spatiotemporal cube, and introduces elastoplastic mechanical constraint loss and multi-task learning in the model training; it adopts a spatiotemporal convolution-memory network to output the risk probability field and deformation prediction field for multiple future time steps; The risk field 3D subdivision and emergency response module is used to overlay predicted high-risk areas with 3D geological grids, cluster voxels using the DBSCAN algorithm and calculate risk body parameters; then, it combines surface DEM data to simulate the impact range, uses the A* algorithm to plan evacuation routes, and uses a digital twin environment to deduce emergency plans and quantitatively evaluate them, generating a comprehensive emergency report.

[0007] Specifically, the execution process of the multi-source data adaptive weighted fusion module is as follows: Data standardization and spatiotemporal registration: Establishing a unified spatiotemporal coordinate system for InSAR deformation data, microseismic monitoring events, geological attribute parameters, and groundwater pressure. Rasterization is performed; for non-raster data, kriging interpolation is used to convert it into a continuous field.

[0008] in The point to be interpolated Estimated value The weight coefficient for the i-th sampling point is... Let i be the measured value of the i-th sampling point. Let i be the spatial coordinates of the i-th known sampling point. The total number of sampling points participating in the interpolation; Adaptive weight allocation based on information entropy and correlation coefficient: Define the first... Weight coefficients of the data source at time t:

[0009] in: For the first The weight coefficients of the data source at time t, where N is the total number of categories in the multi-source data. For the first Information entropy of class data at time t For the first The correlation coefficient between class j and the data of class j; Feature enhancement and fusion output: A weighted fusion formula is used to generate a fused feature field. : ; in: For spatiotemporal location The fusion feature field, For the Laplace operator, For activation function, For the first Normalized values ​​of data fields The sharpening coefficient is the Laplacian operator. The output is a multidimensional fusion map with enhanced features; A self-supervised pre-training stage is introduced to further improve the quality of feature representations and the model's generalization ability.

[0010] Specifically, the execution steps of the self-supervised pre-training phase are as follows: Pre-training data preparation and masking strategy: collect historical multi-source monitoring data for long time series; For the processed spatiotemporal data cube ,in Represents a spatiotemporal data cube. Given a set of real numbers, T, H, W, and C correspond to the time dimension, height dimension, width dimension, and channel dimension, respectively. Randomly select 15%-25% of the spatial blocks or data on continuous time steps, and set the values ​​to zero or a mask mark to simulate the case of missing data. Pre-trained network architecture: Construct a pre-trained network based on a Transformer encoder or U-Net structure; the input is a masked spatiotemporal data cube, and the output is the predicted value of the masked region; Pre-training loss function and learning: The goal of the pre-training task is to minimize the reconstruction error on the masked region; the loss function uses mean squared error or smoothed L1 loss, and only calculates the difference between the predicted value and the true value at the masked position; the pre-trained network parameters are optimized through backpropagation to capture the complex spatiotemporal dependencies and cross-modal correlations in multi-source data; Transfer learning and fine-tuning: After pre-training, the decoder part of the network is discarded, and the encoder is retained as a feature extractor; Initialize the pre-trained encoder weights into the feature extraction part of the main risk assessment model; The entire network was fine-tuned end-to-end using a labeled risk dataset to adapt it to different risk assessment tasks.

[0011] Specifically, the execution process of the physical-spatiotemporal neural network dynamic prediction module is as follows: Constructing a spatiotemporal sequence input: fusing feature fields The organization is a spacetime cube , where T is the time step, H×W is the spatial grid, and C is the number of feature channels; Introducing a physical constraint loss function: During model training, a conventional mean squared error loss is included, along with a deformation compatibility constraint term based on elastoplastic mechanics theory. ; in For deformation compatibility constraint loss based on elastoplastic mechanics, , For the normal strain in the x and y directions, The shear strain is in the xy plane; The total loss function is: ; For weight hyperparameters, For data-driven mean squared error loss; A multi-task collaborative learning framework is adopted to enhance the model's ability to learn physical mechanisms and improve prediction accuracy; Spatiotemporal convolutional-memory network prediction: employing an encoder-decoder structure. Encoder: Uses 3D convolution to extract spatiotemporal features, followed by GRU units to learn time dependencies; Decoder: Uses deconvolution for upsampling, outputting the risk probability field for the next K time steps. and deformation prediction field ; in For risk probability, Let t be the spatial coordinate variable and t be the current time step. For time intervals, Deformation describes the deformation of the rock mass in the goaf.

[0012] Specifically, the multi-task collaborative learning framework includes: Define the auxiliary prediction task: Based on the principles of elastoplastic mechanics, two key field quantities are selected as auxiliary learning objectives, including: a) Maximum shear stress field It indicates the shear yielding tendency of the rock mass; b) Volumetric strain field : Reflects the compression or expansion deformation of the rock mass, and is closely related to the subsidence of the roof and the development of delamination in the goaf; Extend the network output layer: Keep the existing encoder-decoder main structure unchanged; modify the final output layer of the decoder, changing it from the original two output channels. Expanded to 4 output channels, corresponding to: Risk probability field; Deformation prediction field; Maximum shear stress field; Volumetric strain field; Constructing a multi-task loss function: Based on the existing total loss function Based on this, a loss term for auxiliary tasks is introduced; The new multi-task loss function is defined as: ; in: The mean squared error loss for risk probability prediction; The mean square error loss for deformation prediction; This is the original physical constraint loss term; This represents the mean square error loss in the prediction of maximum shear stress. This represents the mean square error loss in volumetric strain prediction; , These are newly introduced hyperparameters; Collaborative training and knowledge sharing: Through backpropagation, the loss of all tasks is optimized simultaneously; the underlying encoder parameters of the network are shared by all tasks, learning a general and powerful feature representation that can simultaneously interpret risk, deformation, stress, and strain.

[0013] Specifically, the execution process of the risk field three-dimensional subdivision and emergency response module is as follows: 3D volume element clustering under risk field constraints: This involves identifying predicted high-risk regions, i.e. Overlaying the data with a 3D geological mesh model, the DBSCAN algorithm is used for volumetric clustering, with the distance metric defined as: ; in: Risk weighting coefficients are used to aggregate spatially adjacent cells with similar risk levels into the same risk cell. For body element and The overall distance For spatial Euclidean distance, For body element The risk probability value; Hazard body geometry and stability parameter calculation: For each identified hazardous body, the following is automatically calculated: volume Where m is the total number of voxels constituting the risk body. The volume of a single voxel; bary coordinates ; Let be the coordinate value of the k-th voxel on a certain coordinate axis; Equivalent burial depth and potential radius of influence ; Equivalent burial depth for the risk body; Automatic emergency plan generation: Based on the geometric parameters of the risk body and surface DEM data, a flood filling algorithm is used to simulate the surface impact range, and the A* algorithm is used to plan evacuation routes. ; in The total cost (path priority) for node n. For actual cost, For heuristic functions, Let n be the risk probability at node n. Risk aversion coefficient; The final output includes a comprehensive emergency report containing an impact area map, route planning, and monitoring point deployment scheme. Finally, a real-time simulation and evaluation system for emergency response plans based on digital twins was constructed.

[0014] Specifically, the process of constructing and evaluating emergency response plans based on digital twins in real time is as follows: First, integrate the UAV oblique photogrammetry real-scene 3D model, building BIM model, GIS road and administrative division data and population spatial distribution heat map, and merge to construct a high-fidelity digital twin environment and semantically annotate scene elements; Secondly, for the evacuation personnel intelligent agent, define the rules for path selection, panic value changes and environmental interaction, and for the rescue vehicle intelligent agent, define the road traffic level and target point rules, and build a behavior model; The emergency plan is then injected into the digital twin environment, and the parallel computing engine drives the intelligent agent to execute the plan to achieve real-time simulation of dynamic interaction. Finally, the simulation calculates evacuation efficiency, safety, and robustness indicators. After the simulation, a visual evaluation report is generated, which includes bottleneck diagnosis, risk warning, and optimization suggestions.

[0015] The technical effects and advantages of this invention are as follows: This invention addresses the inconsistency of multi-source data benchmarks by establishing a unified spatiotemporal coordinate system through a multi-source data adaptive weighted fusion module. It employs Kriging interpolation to accurately convert non-raster data into a continuous field, dynamically allocates weights by combining information entropy and correlation coefficients, and improves feature quality and model generalization ability through self-supervised pre-training. The generated enhanced feature field can accurately reflect the geological state of the goaf area, providing high-quality data support for subsequent risk prediction and effectively solving the defects of data fusion disconnect and static weights in existing technologies. The physical-spatiotemporal neural network dynamic prediction module introduces elastoplastic mechanical constraint loss to make the prediction conform to the physical laws of rock mass instability and avoid the results from deviating from the physical nature. Through multi-task learning, it simultaneously predicts risk probability, deformation, maximum shear stress and volumetric strain to comprehensively characterize the collapse evolution process. Combined with a spatiotemporal convolution-memory network, it outputs prediction fields for multiple future time steps, which greatly improves the prediction accuracy and dimensional completeness and solves the problems of missing physical mechanisms and single prediction dimensions in existing models. The risk field 3D subdivision and emergency response module uses an improved DBSCAN algorithm to achieve 3D volume element clustering in high-risk areas, automatically calculating parameters such as the volume and centroid of risk bodies, replacing manual calculations and improving efficiency and accuracy. It combines surface DEM data and the A* algorithm to generate the impact range and evacuation routes, and then uses a digital twin environment to simulate emergency plans and quantitatively evaluate evacuation efficiency, safety and other indicators, generating a report with optimization suggestions, making emergency plans more accurate and implementable, and solving the shortcomings of existing emergency response 2D subdivision and lack of dynamic evaluation of plans. Attached Figure Description

[0016] Figure 1 This is a system block diagram of the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] like Figure 1 As shown, the data fusion system modules for risk assessment of goaf collapse based on big data processing are as follows: The multi-source data adaptive weighted fusion module first establishes a unified spatiotemporal coordinate system, rasterizes multi-source data such as InSAR deformation, microseismic data, and geological parameters, and transforms non-raster data into a continuous field using Kriging interpolation; then, it assigns adaptive weights based on information entropy (measuring uncertainty) and correlation coefficient, generates an enhanced feature field through a weighted fusion formula, and further improves feature quality and model generalization ability through self-supervised pre-training. Data standardization and spatiotemporal registration: Establishing a unified spatiotemporal coordinate system for InSAR deformation data, microseismic monitoring events, and geological attribute parameters (such as rock mass strength). ), groundwater pressure The data is then rasterized. For non-raster data (such as monitoring points), kriging interpolation is used to convert it into a continuous field. ,in The point to be interpolated Estimated value The weight coefficient for the i-th sampling point is... Let i be the measured value of the i-th sampling point. Let i be the spatial coordinates of the i-th known sampling point. The total number of sampling points participating in the interpolation; Adaptive weight allocation based on information entropy and correlation coefficient: Define the first... Weight coefficients of the data source at time t: ,in For the first The weight coefficients of the data source at time t, where N is the total number of categories in the multi-source data. For the first Information entropy of data at time t (a measure of uncertainty) For the first The correlation coefficient between class j and the data in class j. The weights take into account both the uncertainty of the data itself and its correlation with other data. Feature enhancement and fusion output: A weighted fusion formula is used to generate a fused feature field. : ; in For spatiotemporal location The fusion feature field, For the Laplace operator (second spatial derivative, highlighting the gradient change region). The activation function (only positive gradients are retained, and noise is suppressed). For the first Normalized values ​​of data fields represents the Laplacian sharpening coefficient, used to highlight the boundaries of anomalous regions. The output is a multidimensional fused map with enhanced features.

[0019] A self-supervised pre-training stage is introduced to further improve the quality of feature representations and the model's generalization ability; the specific process is as follows: Pre-training data preparation and masking strategy: Collect historical multi-source monitoring data for long time series, which do not require any risk labels; for the processed spatiotemporal data cube... (in Represents a spatiotemporal data cube. Given a set of real numbers (T, H, W, and C corresponding to the time dimension, height dimension, width dimension, and channel dimension, respectively), randomly select a certain proportion (e.g., 15%-25%) of spatial blocks or data from consecutive time steps, and set their values ​​to zero or a special mask marker to simulate the case of missing data. This step constructs the input (masked data) and target (masked real data) of the pre-training task. Pre-trained network architecture: Construct a pre-trained network based on a Transformer encoder or U-Net structure. The input of this network is a masked spatiotemporal data cube, and the output is the predicted value of the masked region. The core of this network is an encoder, which learns the deep contextual feature representation of the entire data cube and uses these representations to reconstruct the masked information. Pre-training loss function and learning: The goal of the pre-training task is to minimize the reconstruction error on the masked region. The loss function typically uses mean squared error (MSE) or smoothed L1 loss, which only calculates the difference between the predicted and true values ​​at the masked location; the pre-trained network parameters are optimized through backpropagation to enable it to fully capture the complex spatiotemporal dependencies and cross-modal correlations in multi-source data; Transfer learning and fine-tuning: After pre-training, the decoder part in the network is discarded, and the encoder is retained as the feature extractor; the pre-trained encoder weights are initialized to the feature extraction part of the main risk assessment model (i.e. the encoder of the "dynamic prediction module of coupling physical mechanism and spatiotemporal neural network"); the entire network (initialized encoder + prediction head) is fine-tuned end-to-end using a labeled risk dataset (usually with a small amount of data) to adapt to the specific risk assessment task.

[0020] The physical-spatiotemporal neural network dynamic prediction module organizes the fused features into a spatiotemporal cube, introduces elastoplastic mechanical constraint loss and multi-task learning (simultaneously predicting risk probability, deformation, maximum shear stress, and volumetric strain) in the model training; and adopts a spatiotemporal convolution-memory network (encoder-decoder structure) to output the risk probability field and deformation prediction field for multiple future time steps. Constructing a spatiotemporal sequence input: fusing feature fields The organization is a spacetime cube , where T is the time step, H×W is the spatial grid, and C is the number of feature channels.

[0021] Introducing a physical constraint loss function: In model training, in addition to the conventional mean square error loss, a deformation compatibility constraint term based on elastoplastic mechanics theory is introduced. ,in For deformation compatibility constraint loss based on elastoplastic mechanics, , The normal strain in the x and y directions (reflecting the expansion and contraction deformation of the rock mass). The shear strain in the xy plane (reflecting the shear deformation of the rock mass); the total loss function is: These are weight hyperparameters used to balance data-driven and physical constraints. This refers to data-driven mean squared error loss (such as risk probability, prediction error of deformation field).

[0022] To enhance the model's ability to learn physical mechanisms and improve prediction accuracy, a multi-task collaborative learning framework is adopted, as follows: Define the auxiliary prediction task: Based on the principles of elastoplastic mechanics, select two key field quantities that are directly related to the physical process of goaf collapse and can be derived from the main output (deformation field) through basic theoretical formulas as auxiliary learning objectives: a) Maximum shear stress field Shear failure is one of the main mechanisms of rock mass instability, and this field directly indicates the shear yielding tendency of the rock mass; b) Volumetric strain field This reflects the compression or expansion deformation of the rock mass and is closely related to the subsidence of the roof and the development of delamination in the goaf. Extend the network output layer: Keep the existing encoder-decoder core structure unchanged. Modify the final output layer of the decoder to have more than two output channels. It has been expanded to have 4 output channels, corresponding to: Risk probability field; Deformation prediction field; Maximum shear stress field; Volumetric strain field; Constructing a multi-task loss function: Based on the existing total loss function Based on this, a loss term for auxiliary tasks is introduced; the new multi-task loss function is defined as: ; in: The mean squared error loss for risk probability prediction; The mean square error loss for deformation prediction; This is the original physical constraint loss term; The mean square error loss for the maximum shear stress prediction (by the network output) (Compared with the theoretical value calculated from the actual deformation field based on constitutive relations) The mean square error loss for volumetric strain prediction (by adjusting the network output) (Compared with theoretical values ​​calculated from the actual deformation field) , This is a newly introduced hyperparameter used to balance the importance of the two auxiliary tasks; Collaborative training and knowledge sharing: Through backpropagation, the loss of all tasks is optimized simultaneously. The network's underlying encoder parameters are shared by all tasks, forcing it to learn a general and powerful feature representation that can simultaneously explain risk, deformation, stress, and strain. This shared representation naturally incorporates physical laws, enabling the model to make more reasonable predictions even when faced with scenarios not fully covered by training data.

[0023] Spatiotemporal convolutional-memory network prediction: employing an encoder-decoder structure. Encoder: Uses 3D convolution to extract spatiotemporal features, followed by GRU units to learn temporal dependencies; Decoder: Uses deconvolution for upsampling, outputting the risk probability field for the next K time steps. and deformation prediction field .in For risk probability, Let t be the spatial coordinate variable and t be the current time step. For time intervals, Deformation describes the deformation of the rock mass in the goaf.

[0024] The risk field 3D subdivision and emergency response module first overlays the predicted high-risk area with the 3D geological grid, uses the improved DBSCAN algorithm to cluster the elements and calculate the risk body parameters (volume, center of gravity, etc.); then it combines the surface DEM data to simulate the impact range, uses the A* algorithm to plan the evacuation route, and uses the digital twin environment to deduce the emergency plan and quantitatively evaluate it, generating a comprehensive emergency report; 3D volume element clustering under risk field constraints: predicting high-risk regions ( The data is overlaid with a 3D geological mesh model, and a modified DBSCAN algorithm is used for volumetric clustering. The distance metric is defined as follows: ; in Risk weighting coefficients are used to aggregate spatially adjacent cells with similar risk levels into the same risk cell. For body element and The overall distance It represents the spatial Euclidean distance (the L2 norm of the three-dimensional coordinate difference). For body element The risk probability value; Hazard body geometry and stability parameter calculation: For each identified hazardous body, the following is automatically calculated: volume Where m is the total number of voxels constituting the risk body. The total volume of the risk body is obtained by summing the volumes of all voxels, where voxels are the volumes of a single voxel.

[0025] bary coordinates ; Let k be the coordinate value of the k-th voxel on a certain coordinate axis (such as the x-axis); Equivalent burial depth and potential radius of influence ; The equivalent burial depth of the risk body (the equivalent vertical depth from the surface to the risk body). Automatic emergency plan generation: Based on the geometric parameters of the risk body and surface DEM data, a flood filling algorithm is used to simulate the surface impact range, and the A* algorithm is used to plan evacuation routes. ,in The total cost (path priority) for node n. For actual cost, For heuristic functions, Let n be the risk probability at node n. This represents the risk aversion coefficient. The final output includes a comprehensive emergency report comprising an impact area map, route planning, and monitoring point deployment scheme.

[0026] Finally, a real-time simulation and evaluation system for emergency response plans based on digital twins is constructed, as follows: Construction of a high-fidelity digital twin environment: Data fusion and scene generation: Integrating and fusing higher-precision data sources, including: 1) High-precision surface models generated based on UAV oblique photogrammetry real-scene 3D modeling; 2) BIM models or detailed 3D models of key buildings (such as houses, factories, and bridges); 3) Road network and administrative division data imported from GIS systems; 4) Simulated population spatial distribution heat maps generated based on population census and mobile phone signaling data; Semantic annotation of the environment: This involves semantically labeling elements in a 3D scene, such as the grade, width, and load-bearing capacity of roads; the structural type and capacity of buildings; and open areas and obstacles. This provides a basis for the agent's decision-making. Intelligent agent behavior modeling and rule definition: Evacuation personnel intelligent agent: Define behavioral rules for it, such as: Path selection: Prioritize the optimal path planned by the system, but when the path congestion exceeds the threshold, it has the ability to autonomously find alternative paths based on the real-time environment.

[0027] Panic Psychology: Introducing a "panic value" parameter, which increases as the surrounding risk level P_risk rises. Agents with high panic values ​​may be more inclined to conform, deviate from predetermined routes, or change their movement speed; Environmental Interaction: Agents can identify and avoid obstacles and interact with other agents at intersections based on simple traffic rules (such as yielding); Rescue Vehicle Agent: Defining its behavioral rules, such as only being able to travel on roads of a specific level and having a designated relief supply distribution point as the destination.

[0028] Emergency plan injection and real-time simulation: The emergency plans generated by the upstream module (including evacuation route sets, control area boundaries, monitoring point locations, and rescue force deployment points) are taken as input, parsed, and injected into the digital twin environment; the parallel computing engine is activated, driving tens of thousands to millions of intelligent agents to simultaneously execute the plan. The simulation process can be accelerated (for example, simulating the evacuation process over the next few hours requires only a few minutes of computation time); The simulation process is dynamic and interactive: commanders can view the evacuation situation in real time and modify parameters midway (such as temporarily closing a path) to observe the system's response; multi-dimensional performance index quantitative evaluation: during the simulation, the system calculates and records a series of key performance indicators (KPIs) in real time, including but not limited to: Evacuation efficiency indicators include: total evacuation completion time, the number of remaining unevacuated personnel over time, and throughput and congestion index for each route; safety indicators include: cumulative overall risk exposure of personnel (the product of the time each agent spends in the risk area and the local risk value), and the maximum number of people exposed to risk at any given moment; and robustness indicators include: simulating the introduction of random disturbances (such as a small road being suddenly interrupted due to a secondary disaster) to test the probability of plan failure or the extent of performance degradation. These indicators comprehensively and quantitatively measure the merits of the plan.

[0029] Generate an evaluation and optimization recommendation report: After the simulation, the system automatically generates a visual evaluation report. The report includes not only the KPI data but also: Bottleneck diagnosis: Clearly identify potential bottlenecks in the plan (e.g., "Severe congestion occurred at a certain intersection 15 minutes after the start of evacuation, resulting in a 30% decrease in overall efficiency"); Risk warning: High-risk areas are indicated (e.g., "The risk of a certain evacuation route increases significantly in the later stages, and priority should be given to evacuation"); Optimization suggestions: Provide automated suggestions based on simulation results (e.g., "It is recommended to adjust the evacuation priority of route A to be higher than that of route B" or "It is recommended to increase the number of command personnel at location X to manage traffic").

[0030] The above formulas are all dimensionless calculations. Dimensionless calculations can be performed using various methods such as standardization, which will not be elaborated here. The formulas are derived from software simulations based on a large amount of collected data, and the preset parameters in the formulas can be set by those skilled in the art according to the actual situation.

[0031] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more sets of available media. The available medium can be a magnetic medium (e.g., floppy disk, ATA hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium. The semiconductor medium can be a solid-state ATA hard disk.

[0032] It should be understood that in the various embodiments of this application, the order of the above-mentioned processes does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0033] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0034] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0035] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0036] In addition, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0037] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable ATA hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0038] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A data fusion system for assessing the risk of subsidence in goaf areas based on big data processing, characterized in that, Includes the following modules: The multi-source data adaptive weighted fusion module is used to establish a unified spatiotemporal coordinate system, rasterize multi-source data of InSAR deformation, microseismic and geological parameters, and convert non-raster data into a continuous field using Kriging interpolation; then, based on information entropy and correlation coefficient, adaptive weights are assigned, and an enhanced feature field is generated through a weighted fusion formula; and the feature quality and model generalization ability are improved through self-supervised pre-training. The physical-spatiotemporal neural network dynamic prediction module is used to organize the fused features into a spatiotemporal cube, and introduces elastoplastic mechanical constraint loss and multi-task learning in the model training; it adopts a spatiotemporal convolution-memory network to output the risk probability field and deformation prediction field for multiple future time steps; The risk field 3D subdivision and emergency response module is used to overlay predicted high-risk areas with 3D geological grids, cluster voxels using the DBSCAN algorithm and calculate risk body parameters; then, it combines surface DEM data to simulate the impact range, uses the A* algorithm to plan evacuation routes, and uses the digital twin environment to deduce emergency plans and quantitatively evaluate them, generating a comprehensive emergency report. The execution process of the multi-source data adaptive weighted fusion module is as follows: Data standardization and spatiotemporal registration: Establishing a unified spatiotemporal coordinate system for InSAR deformation data, microseismic monitoring events, geological attribute parameters, and groundwater pressure. Rasterization is performed; for non-raster data, kriging interpolation is used to convert it into a continuous field. ； in The point to be interpolated Estimated value The weight coefficient for the i-th sampling point is... Let i be the measured value of the i-th sampling point. Let i be the spatial coordinates of the i-th known sampling point. The total number of sampling points participating in the interpolation; Adaptive weight allocation based on information entropy and correlation coefficient: Define the first... Weight coefficients of the data source at time t: ； in: For the first The weight coefficients of the data source at time t, where N is the total number of categories in the multi-source data. For the first Information entropy of class data at time t For the first The correlation coefficient between class j and the data of class j; Feature enhancement and fusion output: A weighted fusion formula is used to generate a fused feature field. : ； in: For spatiotemporal location The fusion feature field, For the Laplace operator, For activation function, For the first Normalized values ​​of data fields The sharpening coefficient is the Laplacian operator. The output is a multidimensional fusion map with enhanced features; A self-supervised pre-training stage is introduced to further improve the quality of feature representations and the model's generalization ability.

2. The data fusion system for assessing the risk of subsidence in goaf areas based on big data processing as described in claim 1, characterized in that, The execution steps of the self-supervised pre-training phase are as follows: Pre-training data preparation and masking strategy: collect historical multi-source monitoring data for long time series; For the processed spatiotemporal data cube ,in Represents a spatiotemporal data cube. Given a set of real numbers, T, H, W, and C correspond to the time dimension, height dimension, width dimension, and channel dimension, respectively. Randomly select 15%-25% of the spatial blocks or data on continuous time steps, and set the values ​​to zero or a mask mark to simulate the case of missing data. Pre-trained network architecture: Construct a pre-trained network based on a Transformer encoder or U-Net structure; the input is a masked spatiotemporal data cube, and the output is the predicted value of the masked region; Pre-training loss function and learning: The goal of the pre-training task is to minimize the reconstruction error on the masked region; the loss function uses mean squared error or smoothed L1 loss, and only calculates the difference between the predicted value and the true value at the masked position; the pre-trained network parameters are optimized through backpropagation to capture the complex spatiotemporal dependencies and cross-modal correlations in multi-source data; Transfer learning and fine-tuning: After pre-training, the decoder part of the network is discarded, and the encoder is retained as a feature extractor; Initialize the pre-trained encoder weights into the feature extraction part of the main risk assessment model; The entire network was fine-tuned end-to-end using a labeled risk dataset to adapt it to different risk assessment tasks.

3. The data fusion system for assessing the risk of subsidence in goaf areas based on big data processing as described in claim 1, characterized in that, The execution process of the physical-spatiotemporal neural network dynamic prediction module is as follows: Constructing a spatiotemporal sequence input: fusing feature fields The organization is a spacetime cube , where T is the time step, H×W is the spatial grid, and C is the number of feature channels; Introducing a physical constraint loss function: During model training, a conventional mean squared error loss is included, along with a deformation compatibility constraint term based on elastoplastic mechanics theory. ； in For deformation compatibility constraint loss based on elastoplastic mechanics, , For the normal strain in the x and y directions, The shear strain is in the xy plane; The total loss function is: ； For weight hyperparameters, For data-driven mean squared error loss; A multi-task collaborative learning framework is adopted to enhance the model's ability to learn physical mechanisms and improve prediction accuracy; Spatiotemporal convolutional-memory network prediction: employing an encoder-decoder structure. Encoder: Uses 3D convolution to extract spatiotemporal features, followed by GRU units to learn time dependencies; Decoder: Uses deconvolution for upsampling, outputting the risk probability field for the next K time steps. and deformation prediction field ; in For risk probability, Let t be the spatial coordinate variable and t be the current time step. For time intervals, Deformation describes the deformation of the rock mass in the goaf.

4. The data fusion system for assessing the risk of subsidence in goaf areas based on big data processing according to claim 3, characterized in that, The multi-task collaborative learning framework includes: Define the auxiliary prediction task: Based on the principles of elastoplastic mechanics, two key field quantities are selected as auxiliary learning objectives, including: a) Maximum shear stress field It indicates the shear yielding tendency of the rock mass; b) Volumetric strain field : Reflects the compression or expansion deformation of the rock mass, and is closely related to the subsidence of the roof and the development of delamination in the goaf; Extend the network output layer: Keep the existing encoder-decoder main structure unchanged; modify the final output layer of the decoder, changing it from the original two output channels. Expanded to 4 output channels, corresponding to: Risk probability field; Deformation prediction field; Maximum shear stress field; Volumetric strain field; Constructing a multi-task loss function: Based on the existing total loss function Based on this, a loss term for auxiliary tasks is introduced; The new multi-task loss function is defined as: ； in: The mean squared error loss for risk probability prediction; The mean square error loss for deformation prediction; This is the original physical constraint loss term; This represents the mean square error loss in the prediction of maximum shear stress. This represents the mean square error loss in volumetric strain prediction; , These are newly introduced hyperparameters; Collaborative training and knowledge sharing: Through backpropagation, the loss of all tasks is optimized simultaneously; the underlying encoder parameters of the network are shared by all tasks, learning a general and powerful feature representation that can simultaneously interpret risk, deformation, stress, and strain.

5. The data fusion system for assessing the risk of subsidence in goaf areas based on big data processing according to claim 1, characterized in that, The execution process of the risk field three-dimensional subdivision and emergency response module is as follows: 3D volume element clustering under risk field constraints: This involves identifying predicted high-risk regions, i.e. Overlaying the data with a 3D geological mesh model, the DBSCAN algorithm is used for volumetric clustering, with the distance metric defined as: ； in: Risk weighting coefficients are used to aggregate spatially adjacent cells with similar risk levels into the same risk cell. For body element and The overall distance For spatial Euclidean distance, For body element The risk probability value; Hazard body geometry and stability parameter calculation: For each identified hazardous body, the following is automatically calculated: volume Where m is the total number of voxels constituting the risk body. The volume of a single voxel; bary coordinates ; Let be the coordinate value of the k-th voxel on a certain coordinate axis; Equivalent burial depth and potential radius of influence ; Equivalent burial depth for the risk body; Automatic emergency plan generation: Based on the geometric parameters of the risk body and surface DEM data, a flood filling algorithm is used to simulate the surface impact range, and the A* algorithm is used to plan evacuation routes. ； in The total cost of node n For actual cost, For heuristic functions, Let n be the risk probability at node n. Risk aversion coefficient; The final output includes a comprehensive emergency report containing an impact area map, route planning, and monitoring point deployment scheme. Finally, a real-time simulation and evaluation system for emergency response plans based on digital twins was constructed.

6. The data fusion system for assessing the risk of subsidence in goaf areas based on big data processing according to claim 5, characterized in that, The process of constructing and evaluating emergency response plans based on digital twins in real time is as follows: First, integrate the UAV oblique photogrammetry real-scene 3D model, building BIM model, GIS road and administrative division data and population spatial distribution heat map, and merge to construct a high-fidelity digital twin environment and semantically annotate scene elements; Secondly, for the evacuation personnel intelligent agent, define the rules for path selection, panic value changes and environmental interaction, and for the rescue vehicle intelligent agent, define the road traffic level and target point rules, and build a behavior model; The emergency plan is then injected into the digital twin environment, and the parallel computing engine drives the intelligent agent to execute the plan to achieve real-time simulation of dynamic interaction. Finally, the simulation calculates evacuation efficiency, safety, and robustness indicators. After the simulation, a visual evaluation report is generated, which includes bottleneck diagnosis, risk warning, and optimization suggestions.

Images