Multi-source heterogeneous disaster monitoring data fusion and spatio-temporal correlation analysis method and system for coal mine power supply line

Abstract

The application relates to the technical field of coal mine safety monitoring, and discloses a multi-source heterogeneous ground disaster monitoring data fusion and space-time correlation analysis method and system for a coal mine power supply line, which comprises the following steps: collecting multi-source ground disaster data of the coal mine power supply line, and performing standardization processing on the multi-source ground disaster data to obtain a standardized multi-source heterogeneous data set; sequentially performing feature-level fusion and decision-level fusion on the standardized multi-source heterogeneous data set to obtain a real value domain global ground disaster feature tensor; analyzing the global ground disaster feature tensor by using a coal mine ground disaster space-time correlation analysis model constructed based on a quaternion algebra structure, outputting ground disaster prediction results of multiple regions of the coal mine power supply line, and performing ground disaster risk grading early warning based on the ground disaster prediction results, wherein the ground disaster prediction results comprise a ground disaster occurrence probability and a ground surface deformation amount prediction result. The accuracy of ground disaster identification and the advance nature of early warning of the coal mine power supply line are significantly improved.

Description

Technical Field

[0001] This application relates to the field of coal mine safety monitoring technology, and in particular to a method and system for multi-source heterogeneous geological disaster monitoring data fusion and spatiotemporal correlation analysis for coal mine power supply lines. Background Technology

[0002] Power supply lines in coal mines are the lifeline for safe production. These lines often traverse geologically vulnerable areas such as mountainous regions and goaf areas, making them susceptible to geological disasters such as landslides, collapses, debris flows, goaf subsidence, and ground subsidence. Damage to these lines can directly lead to power outages in the mine, potentially causing major safety accidents. Current geological disaster monitoring technology for coal mine power supply lines suffers from the following core technical deficiencies:

[0003] 1. Single monitoring data source and insufficient full coverage capability: Existing solutions mostly rely on single ground-based sensors, which are limited by terrain and environment, resulting in limited coverage and inability to achieve full-area monitoring of the line. Data from single satellites or UAVs suffer from long revisit cycles, susceptibility to weather, and insufficient temporal continuity, making it difficult to adapt to the characteristics of coal mine geological disasters, which are characterized by strong suddenness and rapid evolution.

[0004] 2. Weak ability to fuse multi-source heterogeneous data: Star, air, ground, and well data suffer from different spatiotemporal benchmarks, heterogeneous data structures, and significant differences in spatiotemporal resolution. Traditional fusion methods are mostly simple weighted fusion at the data level or feature dimension splicing, which cannot uncover the deep complementary features of heterogeneous data. They are prone to feature redundancy and low fusion accuracy, and cannot adapt to the geological disaster identification needs of complex coal mine scenarios.

[0005] 3. The spatiotemporal correlation analysis capability suffers from core technical deficiencies, resulting in poor early warning accuracy and predictability:

[0006] 3.1 Existing spatiotemporal graph neural network methods can only linearly fuse multi-source graph structures, and cannot capture the high-order nonlinear coupling effects of explicit and implicit spatial relationships, resulting in distorted spatial dependency characterization.

[0007] 3.2. It is impossible to simultaneously capture the short-term suddenness and long-term gradual change time series dependence of geological disaster evolution, resulting in large cumulative errors in long-term time series predictions and serious delays in early warning;

[0008] 3.3 Existing methods mostly perform time-series analysis on single-point monitoring data, without fully combining the spatial topology of coal mine power supply lines and the spatiotemporal evolution of geological disasters. They cannot achieve integrated analysis of the spatiotemporal transmission path of geological disasters from "point to line to surface", resulting in high false alarm and false alarm rates.

[0009] 4. Inadequate adaptation to the special characteristics of coal mine scenarios and insufficient linkage between surface and underground operations: Existing power transmission line geological disaster monitoring schemes do not fully consider unique geological disaster types such as coal mine goaf collapse and mining-induced landslides / collapses. They only use underground mining data as auxiliary labels and fail to achieve deep linkage spatiotemporal modeling between surface monitoring data and underground mining data, resulting in inaccurate identification of geological disaster causes and a serious lack of risk prediction capabilities.

[0010] 5. Insufficient feasibility of the solution: Existing high-precision spatiotemporal prediction models have a large number of parameters and high computational requirements, making it difficult to adapt to the real-time analysis needs of the edge of coal mine sites; at the same time, the details of the core algorithm are not fully disclosed, making it difficult for those skilled in the art to directly reproduce and implement it.

[0011] Coal gangue is a large-scale solid waste generated during coal mining, washing, and processing. Its stockpile is enormous and its annual increase is continuously rising. Its accumulation not only occupies land resources but also easily causes environmental problems such as dust and leaching pollution. Coal gangue contains rare and dispersed metals such as lithium and gallium, which are core raw materials for strategic emerging industries such as new energy and semiconductors, possessing extremely high recycling value. Therefore, the resource utilization of rare and dispersed metals in coal gangue has become a research hotspot in the field of solid waste disposal and resource regeneration.

[0012] There is still no effective solution to the problem that related technologies cannot simultaneously capture the high-order nonlinear coupling effects of multi-dimensional explicit / implicit spatial correlations and long- and short-term temporal dependencies in the evolution of geological disasters in coal mine power supply lines, resulting in low accuracy of geological disaster identification, delayed early warning, and difficulty in adapting to the mining-induced geological disaster scenarios unique to coal mines. Summary of the Invention

[0013] To address this, this application provides a method and system for fusing and analyzing spatiotemporal correlations of multi-source heterogeneous geological disaster monitoring data for coal mine power supply lines. This method overcomes the problems in related technologies, which cannot simultaneously capture the high-order nonlinear coupling effects of multi-dimensional explicit / implicit spatial correlations and long- and short-term temporal dependencies in the evolution of geological disasters in coal mine power supply lines. These problems result in low accuracy of geological disaster identification, delayed early warning, and difficulty in adapting to the unique mining-induced geological disaster scenarios in coal mines.

[0014] To achieve the above objectives, according to a first aspect of the embodiments of this application, a method for fusing and spatiotemporal correlation analysis of multi-source heterogeneous geological disaster monitoring data for coal mine power supply lines is provided, comprising: collecting multi-source geological disaster data of coal mine power supply lines, and standardizing the multi-source geological disaster data to obtain a standardized multi-source heterogeneous dataset, wherein the multi-source geological disaster data includes: spaceborne remote sensing data, airborne aerial survey data, ground tower monitoring data, and underground mining data; sequentially performing feature-level fusion and decision-level fusion on the standardized multi-source heterogeneous dataset to obtain a real-valued global geological disaster feature tensor; and constructing a coal mine geological disaster spatiotemporal correlation based on a quaternion algebraic structure. The analysis model analyzes the global geological disaster feature tensor and outputs geological disaster prediction results for multiple areas of the coal mine power supply line. The output layer of the coal mine geological disaster spatiotemporal correlation analysis model is divided into two branches: the main branch uses the Sigmoid activation function to output the probability value of geological disaster occurrence for each tower node at a specified future time; the auxiliary regression branch uses the linear activation function to output the predicted value of cumulative surface deformation at a specified future time. Based on the multiple geological disaster prediction results, geological disaster risk classification and early warning are performed. The multiple geological disaster prediction results correspond one-to-one with the multiple areas. The geological disaster prediction results include: probability of geological disaster occurrence and predicted surface deformation results.

[0015] In an optional embodiment, the global geological disaster feature tensor is analyzed using a coal mine geological disaster spatiotemporal correlation analysis model constructed based on a quaternion algebra structure. This includes: encoding the global geological disaster feature tensor into a quaternion feature tensor using a cross-time feature aggregation strategy, wherein each element of the quaternion feature tensor consists of a real part and three imaginary parts. The real part represents the geological disaster baseline feature at the current time step, and the three imaginary parts are the associated temporal sequence features with preset cross-time step intervals; constructing a quaternion graph adjacency matrix corresponding to the coal mine power supply line, wherein the quaternion graph adjacency matrix uses the tower nodes of the coal mine power supply line as graph nodes, and the structure of the quaternion graph adjacency matrix corresponds one-to-one with the quaternion feature tensor, with the real part and the quaternion graph adjacency matrix being adjacent to each other. The mining-deformation implicit adjacency matrix is ​​constructed, with the three imaginary parts corresponding to the spatial topological explicit adjacency matrix, the temporal semantic explicit adjacency matrix, and the geological attribute explicit adjacency matrix, respectively. A quaternion spatiotemporal graph neural network is constructed, wherein the quaternion spatiotemporal graph neural network includes alternating stacked one-dimensional quaternion convolution modules and quaternion graph convolution modules. The coal mine geological disaster spatiotemporal correlation analysis model includes the quaternion spatiotemporal graph neural network. The one-dimensional quaternion convolution module performs one-dimensional convolution operation based on the quaternion Hamiltonian product to extract long- and short-term temporal correlation features in the quaternion feature tensor. The quaternion graph convolution module performs graph convolution operation based on the quaternion Hamiltonian product to capture the high-order nonlinear coupling effect of multi-dimensional spatial correlation in the quaternion graph adjacency matrix.

[0016] In an optional embodiment, the global geological disaster feature tensor is encoded into a quaternion feature tensor using a cross-time feature aggregation strategy, including: the construction formula of the quaternion feature tensor is: ,in, Let be the quaternion characteristic matrix at time step t-T+n. For the real part, , , Let i, j, and k be the imaginary units of the quaternion, satisfying the following condition: The feature mapping rule between the real part and the three imaginary parts is as follows: , , , , Let be the geological disaster baseline characteristics at time step t-T+n. For the first The associated temporal characteristics of time steps For the first The associated temporal characteristics of time steps For the first The associated temporal characteristics of time steps This is a preset time step.

[0017] In an optional embodiment, the preset time step The prediction scenarios employ a multi-scale combination setting, including: for 24-hour short-term disaster early warning scenarios, A value of 1 is used to adapt to minute-level data from ground sensors; this is for scenarios involving 7-day mid-term trend prediction. A value of 3 is used to adapt to hourly downhole mining data; for 30-day long-term evolution prediction scenarios. A value of 7 is used to adapt to weekly InSAR satellite data; this is suitable for general scenarios across the entire data lifecycle. A multi-scale combination of [1,3,7] is adopted.

[0018] In an optional embodiment, constructing the quaternion graph adjacency matrix corresponding to the coal mine power supply line includes: using the formula Construct the adjacency matrix of the quaternion graph, where, The quaternion graph adjacency matrix is... This refers to the sampling-deformation implicit adjacency matrix. Let be the explicit adjacency matrix of the spatial topology. This is the explicit adjacency matrix of the temporal semantics. Let i, j, k be the quaternion imaginary unit, satisfying the following conditions: .

[0019] In an alternative embodiment, the dynamic-deformation implicit adjacency matrix Constructed using an adaptive graph learning strategy, including: a randomly initialized learnable node embedding matrix. Where N is the number of tower nodes and d is the embedding dimension; the feature matrix is ​​obtained through nonlinear mapping. , ,in, To control the hyperparameter of the activation function saturation rate, The weight parameters are learnable; the implicit incidence matrix is ​​initially calculated. ; using Top-K mask matrix Sparsity processing is performed, retaining the top K edges with the strongest association strength for each node, to obtain the final mining-deformation implicit adjacency matrix. This mining-deformation implicit adjacency matrix is ​​obtained through a joint loss function. Achieve end-to-end optimization The mean absolute error loss for deformation prediction. The F-norm regularization term is used to avoid... Overfitting Regularization terms are used to embed nodes to avoid overfitting of node embeddings. This is the regularization weight coefficient.

[0020] In an optional embodiment, the method further includes: the spatial topological explicit adjacency matrix. The temporal semantic explicit adjacency matrix is ​​constructed based on the topological distance of power supply lines between tower nodes and the spatial Euclidean distance of geological units; Based on the dynamic time warping algorithm, the similarity of deformation time series, rainfall time series, and mining time series of different tower nodes or geological units is constructed. The multiple regions include the different tower nodes or geological units. The geological attribute explicit adjacency matrix... The similarity is constructed based on geological environmental attributes, which include: rock and soil type, terrain slope, degree of impact of mining subsidence, and vegetation coverage.

[0021] In an optional embodiment, the convolution operation formula of the one-dimensional quaternion convolution module includes: ,in, Let be the quaternion feature tensor input to the I-th layer. The quaternion convolution kernel of layer I, For the quaternion Hamiltonian product, The kernel size is the convolution kernel size. Let f be the number of input feature channels, r be the spatial location index, s be the output channel index, t be the time step index, and k be the kernel size index. For the first The quaternion feature tensor output by the one-dimensional quaternion convolution module is in Positional elements, used to characterize Location-based fusion of long- and short-term temporal dependence features of coal mine geological disasters; For the first Layer quaternion convolution kernel in The element at the specified position; For the first The quaternion feature tensor of the layer input is in The element at a given position.

[0022] In an optional embodiment, the graph convolution operation formula of the quaternion graph convolution module includes: ,in, For the first Quaternion features output by the layer quaternion graph convolution module; The quaternion features are the input of the I-th layer. For learnable quaternion weight parameters, It is a quaternion graph adjacency matrix. It is the ReLU activation function. It is the Hamiltonian product of quaternions.

[0023] According to a second aspect of the embodiments of this application, a system for fusing and spatiotemporal correlation analysis of multi-source heterogeneous geological disaster monitoring data for coal mine power supply lines is also provided, comprising: a data acquisition module for acquiring multi-source geological disaster data of coal mine power supply lines and standardizing the multi-source geological disaster data to obtain a standardized multi-source heterogeneous dataset, wherein the multi-source geological disaster data includes: spaceborne remote sensing data, airborne aerial survey data, ground tower monitoring data, and underground mining data; a feature fusion module for sequentially performing feature-level fusion and decision-level fusion on the standardized multi-source heterogeneous dataset to obtain a real-valued global geological disaster feature tensor; and a geological disaster prediction module for constructing a quaternion algebraic structure. The obtained spatiotemporal correlation analysis model for coal mine geological disasters analyzes the global geological disaster feature tensor and outputs geological disaster prediction results for multiple areas of the coal mine power supply line. The output layer of the spatiotemporal correlation analysis model for coal mine geological disasters is divided into two branches: the main branch uses the Sigmoid activation function to output the probability value of geological disaster occurrence for each tower node at a specified future time; the auxiliary regression branch uses the linear activation function to output the predicted value of cumulative surface deformation at a specified future time. Based on the multiple geological disaster prediction results, geological disaster risk classification and early warning are carried out. The multiple geological disaster prediction results correspond one-to-one with the multiple areas. The geological disaster prediction results include: probability of geological disaster occurrence and predicted surface deformation results.

[0024] According to a third aspect of the embodiments of this application, an electronic device is also provided, including a memory and a processor, wherein the memory stores a computer program, and the processor is configured to execute, through the computer program, the method for multi-source heterogeneous geological disaster monitoring data fusion and spatiotemporal correlation analysis for coal mine power supply lines as described in the first aspect.

[0025] According to a fourth aspect of the embodiments of this application, a computer program product is also provided, including a computer program that, when executed by a processor, implements the method for multi-source heterogeneous geological disaster monitoring data fusion and spatiotemporal correlation analysis for coal mine power supply lines as described in the first aspect.

[0026] This application collects four types of geological hazard data: spaceborne remote sensing, airborne aerial surveying, ground tower monitoring, and underground mining activity data for coal mine power supply lines. After spatiotemporal benchmark unification and standardization, a standardized multi-source heterogeneous dataset is constructed. Feature-level fusion and decision-level fusion are sequentially performed on the standardized multi-source heterogeneous dataset to generate a real-valued global geological hazard feature tensor. This global geological hazard feature tensor is input into a spatiotemporal correlation analysis model based on a quaternion algebra structure. Quaternion encoding is used to simultaneously capture the nonlinear coupling effects of long-term and short-term temporal dependencies and multi-dimensional spatial correlations of geological hazards, outputting the probability of geological hazard occurrence and surface deformation for each region. The method predicts the magnitude of geological disasters and uses this to complete the geological disaster risk classification and early warning. This solution introduces a quaternion algebraic structure to realize high-order nonlinear fusion and spatiotemporal correlation modeling of multi-source heterogeneous data, which significantly improves the accuracy of geological disaster identification and the advance warning of coal mine power supply lines. It is particularly suitable for coal mine-specific scenarios such as mining-induced geological disasters. This solves the problem in related technologies that cannot simultaneously capture the high-order nonlinear coupling effect of multi-dimensional explicit / implicit spatial correlation and long-short-term time series dependence in the evolution of geological disasters in coal mine power supply lines, resulting in low geological disaster identification accuracy, delayed early warning and difficulty in adapting to mining-induced geological disaster scenarios unique to coal mines.

[0027] It should be understood that the description in the Summary Section is not intended to limit the key or essential features of the embodiments of this application, nor is it intended to restrict the scope of this application. Other features of this application will become readily apparent from the following description. Attached Figure Description

[0028] The above and other features, advantages, and aspects of the embodiments of this application will become more apparent from the accompanying drawings and the following detailed description. The drawings are provided for a better understanding of the invention and are not intended to limit the scope of the application. In the drawings, the same or similar reference numerals denote the same or similar elements, wherein:

[0029] Figure 1 This is a flowchart of an optional method for multi-source heterogeneous geological disaster monitoring data fusion and spatiotemporal correlation analysis for coal mine power supply lines, according to an embodiment of this application.

[0030] Figure 2 This is a flowchart illustrating an optional method for multi-source heterogeneous geological disaster monitoring data fusion and spatiotemporal correlation analysis for coal mine power supply lines, according to an embodiment of this application.

[0031] Figure 3 This is a schematic diagram illustrating the overall geological disaster risk of a coal mine power supply line, as an embodiment of this application.

[0032] Figure 4 This is a structural block diagram of an optional multi-source heterogeneous geological disaster monitoring data fusion and spatiotemporal correlation analysis system for coal mine power supply lines, according to an embodiment of this application. Detailed Implementation

[0033] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0034] Furthermore, the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. Additionally, the character " / " in this article generally indicates that the preceding and following related objects have an "or" relationship.

[0035] To address the technical problems existing in related technologies, this embodiment provides a method for fusing and spatiotemporal correlation analysis of multi-source heterogeneous geological disaster monitoring data for coal mine power supply lines. Figure 1 This is a flowchart of a method for fusing and spatiotemporal correlation analysis of multi-source heterogeneous geological disaster monitoring data for coal mine power supply lines, according to an embodiment of this application. The process includes the following steps:

[0036] Step S102: Collect multi-source geological disaster data of coal mine power supply lines and standardize the multi-source geological disaster data to obtain a standardized multi-source heterogeneous dataset. The multi-source geological disaster data includes: spaceborne remote sensing data, airborne aerial survey data, ground tower monitoring data and underground mining data.

[0037] Step S104: Perform feature-level fusion and decision-level fusion sequentially on the standardized multi-source heterogeneous dataset to obtain the global geological disaster feature tensor in the real-valued domain.

[0038] Step S106: The spatiotemporal correlation analysis model for coal mine geological disasters, constructed based on a quaternion algebraic structure, is used to analyze the global geological disaster feature tensor and output the geological disaster prediction results for multiple areas of the coal mine power supply line. The output layer of the spatiotemporal correlation analysis model for coal mine geological disasters is divided into two branches: the main branch uses the Sigmoid activation function to output the probability value of geological disaster occurrence for each tower node at a specified future time; the auxiliary regression branch uses the linear activation function to output the predicted value of cumulative surface deformation at a specified future time. Based on the multiple geological disaster prediction results, a geological disaster risk classification and early warning is performed. The multiple geological disaster prediction results correspond one-to-one with the multiple areas. The geological disaster prediction results include: the probability of geological disaster occurrence and the predicted result of surface deformation.

[0039] This embodiment collects four types of geological disaster data: spaceborne remote sensing, airborne aerial surveying, ground tower monitoring, and underground mining activity data for coal mine power supply lines. After spatiotemporal benchmark unification and standardization, a standardized multi-source heterogeneous dataset is constructed. Feature-level fusion and decision-level fusion are then performed sequentially on the standardized multi-source heterogeneous dataset to generate a real-valued global geological disaster feature tensor. This global geological disaster feature tensor is input into a spatiotemporal correlation analysis model based on a quaternion algebra structure. Quaternion encoding is used to simultaneously capture the nonlinear coupling effect of long-term and short-term temporal dependencies and multi-dimensional spatial correlations of geological disasters, outputting the probability of geological disaster occurrence and surface topography for each region. The variable prediction results are used to complete the geological disaster risk classification and early warning. This solution introduces a quaternion algebra structure to realize high-order nonlinear fusion and spatiotemporal correlation modeling of multi-source heterogeneous data, which significantly improves the accuracy of geological disaster identification and the advance of early warning for coal mine power supply lines. It is particularly suitable for coal mine-specific scenarios such as mining-induced geological disasters. This solves the problem in related technologies that cannot simultaneously capture the high-order nonlinear coupling effect of multi-dimensional explicit / implicit spatial correlation and long-short-term time series dependence in the evolution of geological disasters in coal mine power supply lines, resulting in low geological disaster identification accuracy, delayed early warning, and difficulty in adapting to mining-induced geological disaster scenarios unique to coal mines.

[0040] Optionally, the global geological disaster feature tensor is analyzed using a coal mine geological disaster spatiotemporal correlation analysis model constructed based on a quaternion algebra structure. This includes: encoding the global geological disaster feature tensor into a quaternion feature tensor using a cross-time feature aggregation strategy, wherein each element of the quaternion feature tensor consists of one real part and three imaginary parts. The real part represents the geological disaster baseline feature at the current time step, and the three imaginary parts are the associated temporal sequence features at preset time step intervals; constructing a quaternion graph adjacency matrix corresponding to the coal mine power supply line, wherein the quaternion graph adjacency matrix uses the tower nodes of the coal mine power supply line as graph nodes, and the structure of the quaternion graph adjacency matrix corresponds one-to-one with the quaternion feature tensor, with the real part corresponding to the quaternion graph adjacency matrix. The mining-deformation implicit adjacency matrix is ​​constructed, with the three imaginary parts corresponding to the spatial topological explicit adjacency matrix, the temporal semantic explicit adjacency matrix, and the geological attribute explicit adjacency matrix, respectively. A quaternion spatiotemporal graph neural network is constructed, wherein the quaternion spatiotemporal graph neural network includes alternating stacked one-dimensional quaternion convolution modules and quaternion graph convolution modules. The coal mine geological disaster spatiotemporal correlation analysis model includes the quaternion spatiotemporal graph neural network. The one-dimensional quaternion convolution module performs one-dimensional convolution operation based on the quaternion Hamiltonian product to extract the long-short-term temporal correlation features in the quaternion feature tensor. The quaternion graph convolution module performs graph convolution operation based on the quaternion Hamiltonian product to capture the high-order nonlinear coupling effect of multi-dimensional spatial correlation in the quaternion graph adjacency matrix.

[0041] This embodiment discloses the specific construction method of the quaternion spatiotemporal correlation analysis model. The core of the model lies in using the quaternion algebraic structure to achieve efficient encoding and nonlinear modeling of the spatiotemporal characteristics of coal mine geological disasters.

[0042] 1. Quaternion Feature Tensor Construction: A cross-time feature aggregation strategy is adopted to convert the real-domain global geological disaster feature tensor into a quaternion-domain representation. Specifically, each quaternion element contains one real part and three imaginary parts: the real part carries the geological disaster baseline features at the current time step; the first imaginary part carries the historical features at an interval of Tm; the second imaginary part carries the historical features at an interval of 2Tm; and the third imaginary part carries the historical features at an interval of 3Tm. Taking a 24-hour short-term early warning scenario as an example, Tm is set to 1, corresponding to minute-level data from ground sensors. A single quaternion element can simultaneously encode the feature information of the current minute and the previous 1 minute, 3 minutes, and 7 minutes, realizing short-term capture of sudden geological disasters. Taking a 30-day long-term prediction scenario as an example, Tm is set to 7, corresponding to weekly InSAR satellite data. A single quaternion element can simultaneously encode the feature information of the current week and the previous 1 week, 3 weeks, and 7 weeks, realizing long-term trend prediction of slowly changing geological disasters such as mining subsidence.

[0043] 2. Quaternion Graph Adjacency Matrix Construction: A quaternion graph adjacency matrix is ​​constructed using coal mine power line towers as graph nodes. The matrix structure maintains an algebraic correspondence with the quaternion feature tensor. The real part adopts a mining-deformation implicit adjacency matrix, which uses an adaptive graph learning strategy to mine the potential nonlinear correlation between underground mining disturbances and surface deformation. This matrix is ​​designed for the specific scenario of coal mines, achieving deep linkage between underground and surface data. The first imaginary part adopts a spatial topological explicit adjacency matrix, constructed based on the topological distance between power lines and the spatial Euclidean distance of geological units, characterizing geographical proximity relationships. The second imaginary part adopts a temporal semantic explicit adjacency matrix, constructed based on the similarity of different tower deformation time series, rainfall time series, and mining time series calculated using a dynamic time warping algorithm, capturing semantic relationships that are spatially distant but have highly similar evolutionary patterns. The third imaginary part adopts a geological attribute explicit adjacency matrix, constructed based on the similarity of geological environmental attributes such as soil and rock type, terrain slope, and degree of impact of goaf, characterizing attribute relationships of units with the same geological vulnerability.

[0044] 3. Quaternion Spatiotemporal Graph Neural Network Construction: The network adopts an architecture that alternates between one-dimensional quaternion convolution modules and quaternion graph convolution modules. The one-dimensional quaternion convolution module performs convolution operations based on the quaternion Hamiltonian product. Through the Hamiltonian product operation between the quaternion convolution kernel and the input feature tensor, it simultaneously extracts the short-term sudden temporal dependencies and long-term slowly varying temporal dependencies of geological disaster evolution. Gated linear units perform nonlinear filtering on the convolution output to retain temporal features strongly correlated with geological disaster evolution. The quaternion graph convolution module performs graph convolution operations based on the quaternion Hamiltonian product. Taking the quaternion graph adjacency matrix and the output of the one-dimensional quaternion convolution module as input, it directly captures the high-order nonlinear coupling effects between four types of spatial relationships: implicit correlation between mining and deformation, spatial topological correlation, temporal semantic correlation, and geological attribute correlation, through the triple Hamiltonian product operation of the adjacency matrix, node features, and learnable weight parameters. This accurately characterizes the complex physical processes in which multiple factors mutually restrict and stimulate each other. The network uses residual connections to avoid gradient vanishing, and the output layer adopts a dual-branch structure to output the probability of geological disaster occurrence and the predicted values ​​of surface deformation, respectively.

[0045] This embodiment introduces a quaternion algebra structure to achieve a unified representation of multi-dimensional spatiotemporal features with a compact encoding method of 1 real part plus 3 imaginary parts. By leveraging the nonlinear interaction characteristics of the Hamiltonian product, it breaks through the limitations of traditional linear fusion. With only one-quarter of the number of parameters in a real-valued network of the same dimension, it significantly improves the modeling accuracy and computational efficiency of spatiotemporal correlation of geological disasters, effectively adapting to the real-time analysis needs of the edge of coal mine sites.

[0046] Optionally, a cross-time feature aggregation strategy is used to encode the global geological disaster feature tensor into a quaternion feature tensor, including: the construction formula of the quaternion feature tensor is: ,in, Let be the quaternion characteristic matrix at time step (t-T+n). For the real part, , , Let i, j, and k be the imaginary units of the quaternion, satisfying the following condition: The feature mapping rule between the real part and the three imaginary parts is as follows: , , , , The geological disaster baseline characteristics at time step (t-T+n) are as follows. For the ( The associated temporal characteristics of time steps, For the ( The associated temporal characteristics of time steps, For the ( The associated temporal characteristics of time steps, This is a preset time step.

[0047] This embodiment discloses a mathematical implementation method for a cross-time feature aggregation strategy, which encodes the real-valued global geological disaster feature tensor into a quaternion feature tensor using a specific formula.

[0048] The encoding formula is In this formula The quaternion characteristic matrix representing the (t-T+n)th time step is composed of a linear combination of the real part and three imaginary parts. The real part corresponds to the geological disaster baseline feature at the (t-T+n)th time step. , , The three imaginary parts correspond to historical correlation features with different time steps. i, j, and k are quaternion imaginary units that satisfy the Hamiltonian algebra rules. This rule determines that quaternion multiplication is noncommutative, providing an algebraic basis for subsequent high-order nonlinear coupling modeling.

[0049] The feature mapping rules clarify the data sources for the real part and the three imaginary parts. Real part Directly take the features of the current time step First imaginary part Take interval T m Features of time step Second imaginary part Features at 2Tm time steps The third imaginary part Take an interval of 3T m Features of time step T mTo preset the time step parameters, zero-padding is performed on feature values ​​that exceed the time series range.

[0050] Taking landslide early warning scenarios as an example, ground sensors collect minute-level data, T m If the value is 1, and the current time step is the 100th minute, then a single quaternion element encodes the real-valued feature of the 100th minute as the real part, the feature of the 101st minute as the first imaginary part, the feature of the 103rd minute as the second imaginary part, and the feature of the 107th minute as the third imaginary part, thus capturing sudden temporal changes between adjacent minute-level time steps. Taking the long-term prediction scenario of goaf collapse as an example, InSAR satellites provide weekly data, T m If the value is 7 and the current time step is week 4, then the single quaternion element encodes the real feature of week 4 as the real part, the feature of week 5 as the first imaginary part, the feature of week 6 as the second imaginary part, and the feature of week 7 as the third imaginary part, so as to simultaneously preserve the short-term fluctuations at the weekly level and the long-term gradual trend at the monthly level.

[0051] This embodiment leverages the algebraic structure of quaternions to integrate historical features scattered across different time steps into a unified mathematical object in a compact form of 1 real part plus 3 imaginary parts. This avoids the information structure destruction caused by traditional vector concatenation, enabling a single data unit to simultaneously possess the ability to represent the current state and remember multi-scale historical associations, thus laying a data foundation for the efficient nonlinear operations of subsequent spatiotemporal graph neural networks.

[0052] Optionally, the preset time step The prediction scenarios employ a multi-scale combination setting, including: for 24-hour short-term disaster early warning scenarios, A value of 1 is used to adapt to minute-level data from ground sensors; this is for scenarios involving 7-day mid-term trend prediction. A value of 3 is used to adapt to hourly downhole mining data; for 30-day long-term evolution prediction scenarios. A value of 7 is used to adapt to weekly InSAR satellite data; this is suitable for general scenarios across the entire data lifecycle. A multi-scale combination of [1,3,7] is adopted.

[0053] This embodiment describes a preset time step. The multi-scale adaptive configuration strategy enables quaternion feature encoding to accurately match the temporal resolution requirements of different prediction scenarios.

[0054] The parameter determines the time interval between the imaginary and real parts of the quaternion's features, and its value is directly related to the data acquisition cycle and evolution characteristics of geological disaster monitoring data. For 24-hour short-term disaster early warning scenarios, A value of 1 matches the minute-level data acquisition frequency of ground sensors. The three imaginary parts of the quaternion element correspond to historical features from the previous 1 minute, 3 minutes, and 7 minutes, respectively, enabling the capture of minute-level sudden temporal changes indicating landslides or micro-seismic events triggered by heavy rainfall. For 7-day medium-term trend prediction scenarios, A value of 3 is chosen to match the hourly acquisition frequency of downhole mining data. The three imaginary parts of the quaternion element correspond to the historical characteristics of the previous 3 hours, 9 hours, and 21 hours, respectively, enabling simultaneous capture of short-term intraday changes and medium-term trends on a 3-day scale, thus adapting to the gradual process of deformation evolution induced by mining disturbances. For a 30-day long-term evolution prediction scenario... A value of 7 is chosen to match the weekly revisit cycle of InSAR satellites. The three imaginary parts of the quaternion element correspond to the historical characteristics of the previous 1, 3, and 7 weeks, respectively, enabling the simultaneous capture of short-term weekly fluctuations and long-term gradual monthly trends, thus adapting to the slow accumulation process of goaf subsidence and ground settlement. For a universal scenario covering the entire lifecycle, By employing a multi-scale combination of [1,3,7], a single quaternion matrix can simultaneously cover full-scale time series information from minute to week, achieving unified modeling across the entire domain and throughout the entire cycle.

[0055] Taking the comprehensive monitoring of a 110kV power supply line in a coal mine as an example, the line is 25 kilometers long and traverses three goaf areas and several steep mountain slopes, posing risks of landslides and goaf collapses. The monitoring system deployed along this line simultaneously includes minute-level data from ground sensors, hourly data from underground mining activities, and weekly data from InSAR satellites, employing a multi-scale combination of [1,3,7]. After parameter settings, a single quaternion feature tensor can simultaneously encode minute-level rainfall mutation information required for landslide early warning in the next 24 hours, hour-level mining disturbance information required for risk prediction of mined-out affected sections in the next 7 days, and weekly deformation trend information required for overall line settlement prediction in the next 30 days, avoiding the redundant overhead of building multiple independent models for different prediction tasks.

[0056] This embodiment is illustrated by... The scenario-based adaptive configuration of parameters enables the same quaternion coding framework to flexibly adapt to the full spectrum of coal mine geological disasters, from sudden to gradual changes. While maintaining the uniformity of the model structure, it achieves accurate predictions at multiple time scales, significantly improving the engineering practicality and deployment convenience of the method.

[0057] In an exemplary embodiment, constructing the quaternion graph adjacency matrix corresponding to the coal mine power supply line includes: using the formula Construct the adjacency matrix of the quaternion graph, where, The quaternion graph adjacency matrix is... This refers to the sampling-deformation implicit adjacency matrix. Let be the explicit adjacency matrix of the spatial topology. This is the explicit adjacency matrix of the temporal semantics. Let i, j, k be the quaternion imaginary unit, satisfying the following conditions: .

[0058] This embodiment describes a mathematical construction method for quaternion graph adjacency matrices, which encodes the multi-dimensional spatial associations of coal mine power supply lines into a graph structure representation of the quaternion field through a specific formula.

[0059] The formula is as follows In this formula This represents a quaternion graph adjacency matrix, with coal mine power supply line towers as graph nodes, and the matrix dimension matching the number of towers. The component is designed for coal mine-deformation implicit adjacency matrix, which serves as the real part of the quaternion. It uses an adaptive graph learning strategy to mine the potential nonlinear correlation between underground mining disturbances and surface deformation. This component is designed for coal mine-specific scenarios and enables in-depth linkage modeling of monitoring data from both underground and surface sources. The spatial topological explicit adjacency matrix, with the first imaginary part coefficient, is constructed based on the topological distance between power supply lines between towers and the spatial Euclidean distance of geological units, characterizing the geographical spatial proximity relationship. As a temporal semantic explicit adjacency matrix, and as the second imaginary part coefficient, it is constructed based on the dynamic time warping algorithm to calculate the similarity of different tower deformation time series, rainfall time series, and mining time series, capturing semantic associations that are spatially distant but have highly similar evolutionary patterns. This is an explicit adjacency matrix of geological attributes, with the third imaginary part coefficients constructed based on the similarity of geological environmental attributes such as soil and rock type, terrain slope, degree of impact from mining subsidence, and vegetation cover, characterizing the attribute associations of units with the same geological vulnerability. i, j, and k are quaternion imaginary units, satisfying Hamiltonian algebra rules. This rule causes a nonlinear interaction effect in the quaternion multiplication operation of the four correlation matrices.

[0060] Taking the K18-K22km section of a 110kV power supply line in a coal mine as an example, this section directly crosses the goaf of the first mining face, with mining-induced surface subsidence being the main risk. The real part of the adjacency matrix of the quaternion graph for this section... The weighting has been increased to 60%, with a focus on the spatiotemporal transmission patterns of underground mining progress and surface subsidence; the first imaginary part. The weights are adjusted accordingly to focus on capturing the transmission path of collapse risk between adjacent towers; the second imaginary part The high temporal similarity between this segment and historical subsidence areas was identified; the third imaginary part The geological attribute correlation between this segment and other affected segments along the route is marked. After the four matrices are integrated by quaternion addition, they are coupled with each other in the subsequent Hamiltonian product operation to accurately depict how mining disturbance is transmitted to distant ends through similar geological attributes, how it is amplified in spatial topological adjacent areas, and how it forms a resonance effect with segments with similar temporal evolution patterns.

[0061] This embodiment integrates four types of heterogeneous spatial relationships into a unified mathematical object through a quaternion algebra structure, avoiding the loss of high-order coupling information caused by traditional multi-graph linear weighting. This allows subsequent graph neural networks to directly capture the mutual modulation and excitation effects of multi-dimensional relationships through Hamiltonian product operations, significantly improving the physical realism and prediction accuracy of the spatial dependence characterization of coal mine geological disasters.

[0062] Optionally, the sampling-deformation implicit adjacency matrix The adaptive graph learning strategy is constructed, including: initializing a randomly learnable node embedding matrix. This is used to map tower nodes to a high-dimensional feature space, where N is the number of tower nodes and d is the embedding dimension; the feature matrix is ​​obtained through nonlinear mapping. , ,in, To control the hyperparameter of the activation function saturation rate, The weight parameters are learnable; the implicit incidence matrix is ​​initially calculated. ; using Top-K mask matrix Sparsity processing is performed, retaining the top K edges with the strongest association strength for each node, to obtain the final mining-deformation implicit adjacency matrix. This mining-deformation implicit adjacency matrix is ​​obtained through a joint loss function. Achieve end-to-end optimization The mean absolute error loss for deformation prediction. The F-norm regularization term is used to avoid... Overfitting Regularization terms are used to embed nodes to avoid overfitting of node embeddings. This is the regularization weight coefficient. The feature matrix after node embedding nonlinear mapping is the basis for subsequent implicit correlation matrix calculation.

[0063] This embodiment discloses an adaptive graph learning method for constructing the implicit adjacency matrix of mining-deformation. It automatically mines the potential nonlinear correlation between downhole mining data and surface deformation data through an end-to-end optimization strategy, without relying on manual prior definition of correlation rules.

[0064] The construction process comprises five core steps. The initialization step sets up two random, learnable node embedding matrices. Where N is the total number of tower nodes, and d is the embedding dimension, which is set to 40 by default. This dimension balances feature representation capability and computational cost. The nonlinear mapping step maps the node embeddings to a high-dimensional feature space using a hyperbolic tangent activation function and learnable weight parameters, resulting in... , Where α is a hyperparameter controlling the saturation rate of the activation function, with a default value of 0.1. The nonlinear mapping, which uses a learnable weight parameter matrix, enhances feature representation and enables subsequent association mining to overcome linear limitations. The implicit association calculation stage captures the asymmetric transmission relationship between mining and deformation through antisymmetric matrix constraints, as shown in the formula. ,in, and The difference operation characterizes the unidirectional influence of disturbances on deformation, while the ReLU activation function filters out negative values ​​and retains valid correlations. In the sparsity processing stage, a Top-K mask matrix is ​​used to retain only the top K edges with the strongest correlation strength for each node (default K=20). This operation filters out noise interference and improves matrix computation efficiency. The end-to-end optimization stage employs a joint loss function. ,in, The mean absolute error loss in deformation prediction is used as the optimization objective. , 10 respectively -4 With 10 -5 The regularization weight coefficients, and the two types of F-norm regularization terms respectively constrain the size of the adjacency matrix and the smoothness of node embedding to avoid overfitting.

[0065] This embodiment overcomes the limitations of traditional methods that rely on manually defined association rules by using an adaptive graph learning mechanism. It enables the automatic discovery and dynamic updating of complex transmission patterns of mining-deformation, providing core data support for the accurate tracing of the causes and risk prediction of coal mine-specific geological disasters, and significantly improving the model's ability to identify coal mine-specific risks such as goaf collapse.

[0066] Optionally, the method further includes: the spatial topological explicit adjacency matrix. The temporal semantic explicit adjacency matrix is ​​constructed based on the topological distance of power supply lines between tower nodes and the spatial Euclidean distance of geological units; Based on the dynamic time warping algorithm, the similarity of deformation time series, rainfall time series, and mining time series of different tower nodes or geological units is constructed. The multiple regions include the different tower nodes or geological units. The geological attribute explicit adjacency matrix... The similarity is constructed based on geological environmental attributes, which include: rock and soil type, terrain slope, degree of impact of mining subsidence, and vegetation coverage.

[0067] This embodiment discloses a specific method for constructing three explicit adjacency matrices in a quaternion graph adjacency matrix, which respectively characterize the explicit correlation rules of geological disaster evolution of coal mine power supply lines from three dimensions: spatial topology, temporal semantics, and geological attributes.

[0068] The spatial topological explicit adjacency matrix is ​​constructed based on the topological distance of power lines between tower nodes and the spatial Euclidean distance of geological units. The topological distance is measured along the power line direction, reflecting the physical channels of electrical connections and geological disaster transmission between towers. The Euclidean distance is calculated based on the straight-line distance of latitude and longitude coordinates, characterizing geographic spatial proximity. The fusion of the two distances allows the matrix to simultaneously consider the line direction and spatial proximity, accurately capturing the direct transmission characteristics of geological disasters between adjacent towers.

[0069] The temporal semantic explicit adjacency matrix is ​​constructed based on the similarity of deformation time series, rainfall time series, and mining time series among different tower nodes or geological units calculated using a dynamic time warping algorithm. The dynamic time warping algorithm calculates the morphological similarity of time series sequences of different lengths through nonlinear bending alignment, unaffected by differences in acquisition frequency or phase shift. Deformation time series similarity identifies tower groups that are spatially distant but exhibit highly similar deformation patterns. Rainfall time series similarity identifies tower clusters controlled by the same meteorological system. Mining time series similarity identifies sets of towers affected by the same mining face. This matrix enables the model to detect spatially separated but synchronously evolving precursors to geological disaster clusters.

[0070] The explicit adjacency matrix of geological attributes is constructed based on the similarity of rock and soil types, topographic slope, degree of impact from mining subsidence, and vegetation cover. Similar rock and soil types characterize the attribute associations of geological units with the same vulnerability, such as the clustering pattern of landslides at the mudstone-sandstone interface. Similar topographic slopes identify groups of vulnerable slopes. The degree of impact from mining subsidence directly indicates the risk level of tower disturbance caused by mining. Similar vegetation cover reflects soil and water conservation capacity and slope stability. This matrix addresses the problem of characterizing the clustering patterns of similar geological disasters in similar geological environments.

[0071] This embodiment constructs explicit associations from three dimensions: spatial proximity, temporal evolution, and geological environmental attributes through three explicit adjacency matrices. These matrices, together with the mining-deformation implicit adjacency matrix, form a multidimensional association system that is both explicit and implicit. After unified encoding with quaternions, high-order nonlinear coupling is achieved through Hamiltonian product, which accurately characterizes the complex interactive effects of multidimensional spatial associations of coal mine geological disasters.

[0072] Optionally, the convolution operation formula of the one-dimensional quaternion convolution module includes: ,in, Let be the quaternion feature tensor input to the I-th layer. The quaternion convolution kernel of layer I, For the quaternion Hamiltonian product, The kernel size is the convolution kernel size. Let f be the number of input feature channels, r be the spatial location index, s be the output channel index, t be the time step index, and k be the kernel size index. For the first The quaternion feature tensor output by the one-dimensional quaternion convolution module is in Positional elements, used to characterize Location-based fusion of long-term and short-term temporal dependence features of coal mine geological disasters.

[0073] This embodiment discloses the mathematical operation mechanism of a one-dimensional quaternion convolution module, which realizes the nonlinear extraction of geological disaster temporal features and the synchronous capture of long and short-term dependencies through the quaternion Hamiltonian product.

[0074] The calculation formula is: In this formula Indicates the first The quaternion feature tensor output by the one-dimensional quaternion convolution module is in The output feature element of the location is a fusion of long-term and short-term temporal dependence information of coal mine geological disasters. The quaternion feature tensor, which is the input of the I-th layer, carries multi-scale temporal information encoded by a cross-temporal feature aggregation strategy. The kernel is a quaternion convolution kernel for layer I. The kernel parameters are optimized through training to adapt to the characteristics of geological disaster evolution. This is a quaternion Hamiltonian product, which enables a non-linear interaction between the convolution kernel and the input features, unlike the linear weighted summation of traditional real-valued convolutions. The kernel size determines the length of the time-aware window. Let f be the number of input feature channels, r be the spatial location index, s be the output channel index, t be the time step index, and k be the kernel size index. A double summation operation completes the accumulation and aggregation of the channel dimension and the kernel size dimension.

[0075] The Hamiltonian product rule gives quaternion convolution a unique advantage. Quaternion multiplication satisfies... Furthermore, it is non-commutative; the real part and the three imaginary parts are coupled together during multiplication to produce cross terms, and a single convolution operation simultaneously integrates the features of the current time step and the interval. 2 3 It captures the historical characteristics of time and the higher-order nonlinear relationships between them. Traditional one-dimensional convolution can only capture linear temporal dependencies within a local short time window, while quaternion convolution expands the range of perception and enhances the ability to express nonlinearities through algebraic structure properties.

[0076] This embodiment replaces the linear operation of traditional real-valued convolution with quaternion Hamiltonian product, and realizes high-order nonlinear fusion of multi-scale temporal information with compact algebraic structure. It completes the long-short-term dependencies that traditional methods would need to stack multiple layers to capture at a single level, significantly improving the efficiency and accuracy of temporal feature extraction, and providing core computational support for the synchronous characterization of the dual evolutionary characteristics of suddenness and slow change of geological disasters.

[0077] Optionally, the graph convolution operation formula of the quaternion graph convolution module includes: ,in, The quaternion features are the input of the I-th layer. For learnable quaternion weight parameters, It is a quaternion graph adjacency matrix. It is the ReLU activation function. It is the Hamiltonian product of quaternions.

[0078] This embodiment discloses the mathematical operation mechanism of the quaternion graph convolution module, which realizes high-order nonlinear coupling modeling of multi-dimensional spatial correlation through triple quaternion Hamiltonian product.

[0079] The calculation formula is In this formula The output features of the (l+1)th layer quaternion graph convolution module at time step t are represented, which integrates the high-order nonlinear dependencies of multi-dimensional spatial correlation of coal mine geological disasters. The quaternion features input to the l-th layer carry temporal correlation information extracted by the one-dimensional quaternion convolution module. It is a quaternion graph adjacency matrix that integrates implicit associations of mining-deformation and three types of explicit associations: spatial topology, temporal semantics, and geological attributes. The quaternion weight parameters learned in the l-th layer are optimized through training to meet the spatial feature extraction requirements. The ReLU activation function performs a non-linear interaction between the adjacency matrix, node features, and weight parameters. It is a quaternion Hamiltonian product that performs a nonlinear interaction between the adjacency matrix, node features, and weight parameters.

[0080] In an optional embodiment, this application proposes a method for fusing and spatiotemporal correlation analysis of multi-source heterogeneous geological disaster monitoring data for coal mine power supply lines. The core idea of ​​this method is as follows: construct an integrated multi-source data acquisition system of "satellite-space-ground-well", and achieve standardized adaptation of multi-source data through unified spatiotemporal benchmarks and heterogeneous data normalization processing; achieve two-level deep fusion of feature level and decision level based on an improved multimodal deep learning model to eliminate data redundancy and mine complementary features; construct a spatiotemporal correlation analysis model specific to coal mine scenarios based on spatiotemporal graph convolutional neural network, and achieve unified nonlinear modeling of multi-source heterogeneous features and multi-dimensional spatial correlation through quaternion encoding, simultaneously capturing the long-term and short-term temporal dependencies of geological disasters, and finally realizing the full-domain perception, cause tracing, evolution prediction and graded early warning of geological disaster risks.

[0081] Specifically, the steps of this method are as follows: Figure 2 As shown, the specific steps include S1-S5:

[0082] Step S1: Construct a multi-source heterogeneous geological disaster monitoring data acquisition system encompassing space, ground, and wells to complete the aggregation of data across the entire region.

[0083] For the entire power supply line route in coal mines, a multi-source data acquisition system covering the entire spatial domain, the entire temporal cycle, and all causal dimensions is constructed. A time-series data lake is built based on a Hadoop+Spark distributed big data architecture to achieve unified storage and management of multi-source heterogeneous data. Specific acquisition dimensions include:

[0084] 1. Satellite Data: Including InSAR satellite surface deformation time series data, high-resolution optical satellite imagery, and meteorological satellite rainfall / temperature / wind speed / humidity time series data, to achieve long-term time series data collection of surface deformation, land cover changes, and meteorological factors across the entire line area;

[0085] 2. Airborne data (airborne): This includes high-precision DSM / DEM data, LiDAR laser point cloud data, and infrared thermal imaging data generated by UAVs carrying oblique photography. For key areas such as mined-out areas, steep slopes, and geologically vulnerable sections, it enables the detailed acquisition of centimeter-level three-dimensional surface morphology, tower foundation deformation, and ground feature cracks.

[0086] 3. Ground monitoring data (ground): including real-time time series data of GNSS displacement monitoring stations, tower tilt sensors, foundation crack gauges, soil moisture sensors, rain gauges, and microseismic monitoring instruments deployed on power line towers and along the line, to achieve high-precision dynamic monitoring of the tower body and surrounding ground surface at the second to minute level;

[0087] 4. Underground mining data (specific to coal mine scenarios): including coal mine underground working face mining progress, three-dimensional distribution of goaf, mine pressure monitoring data, roof displacement data, and advance support pressure data, to achieve joint collection of underground and surface data on core geological disaster causes.

[0088] Step S2: Spatiotemporal benchmark unification and standardization preprocessing of multi-source heterogeneous data to eliminate heterogeneous differences.

[0089] This step lays the foundation for multi-source data fusion and consists of two core components: unifying spatiotemporal benchmarks and standardizing heterogeneous data preprocessing. It also addresses the challenge of synchronizing and aligning data from different temporal resolutions.

[0090] S2.1 Spatiotemporal Reference Unification:

[0091] Spatial dimension: All multi-source data are uniformly mapped to a unified benchmark. Joint geometric calibration is performed on InSAR data, UAV point cloud and GNSS data. The corresponding point matching algorithm is used to eliminate spatial registration error and control the spatial registration accuracy to the centimeter level.

[0092] Time dimension: The timestamps of all data are unified to the UTC time standard. To address the time resolution differences between downhole mining data (hourly), satellite InSAR data (daily / weekly), and ground sensor data (second / minute), a scheme combining time window-based Kriging interpolation and GAN generative adversarial network completion is adopted. The Flink distributed stream processing engine is used to achieve real-time timestamp alignment and time axis normalization of multi-source data. The time alignment error is controlled within 1 minute, solving the synchronization problem between satellite-borne daily data and downhole second-level data.

[0093] S2.2 Heterogeneous Data Standardization Preprocessing and Normalization:

[0094] For heterogeneous data with different structures, preprocessing is performed separately, and then the data is uniformly converted into a standardized feature format that can be adapted to the model, thus solving the problem of data heterogeneity:

[0095] 1. Structured time series data (ground sensor, meteorological, and downhole mining data): Wavelet transform is used to remove noise, outliers are removed by the 3σ criterion, missing values ​​are filled in based on GAN generative adversarial network, and then the data is mapped to the [0,1] interval by min-max standardization to eliminate dimensional differences;

[0096] 2. Unstructured spatial data (satellite optical imagery, UAV imagery, infrared thermal imaging data): Complete radiometric calibration, atmospheric correction, geometric correction and distortion removal; perform ground feature segmentation based on the improved U-Net model; extract key ground feature features such as power line towers, surface cracks, landslides, and vegetation cover; and generate standardized spatial feature maps.

[0097] 3. 3D point cloud and deformation data (InSAR deformation data, LiDAR point cloud, DSM / DEM data): Complete point cloud denoising, registration and simplification, extract surface temporal deformation rate data through differential InSAR technology, construct a global surface deformation 3D feature matrix, and uniformly convert it into a standardized 3D spatial feature tensor.

[0098] Step S3: Two-level deep fusion of multi-source heterogeneous data based on multimodal deep learning to mine deep complementary features.

[0099] This step employs a two-level deep fusion architecture of feature level + decision level, fully leveraging the complementary advantages of space-time and ground data to solve the problems of low accuracy and feature redundancy in traditional fusion methods. Specifically, it includes:

[0100] S3.1 Feature-level deep fusion based on improved multimodal Transformer:

[0101] 1. Construct a multimodal feature encoding branch, and use corresponding encoding networks to perform deep feature extraction for the standardized temporal feature vector, spatial feature map, and 3D feature tensor respectively:

[0102] 1.1 Temporal Feature Branch: A Bi-LSTM bidirectional long short-term memory network combined with an attention mechanism is used to extract the dynamic change features of temporal data and capture temporal mutation information related to geological disasters;

[0103] 1.2 Spatial Image Feature Branch: ResNet50 backbone convolutional neural network is used to extract spatial texture and ground feature change features of the image, and capture the spatial distribution features of surface deformation and ground feature cracks;

[0104] 1.3 Three-dimensional deformation feature branch: 3D-CNN three-dimensional convolutional neural network is used to extract the three-dimensional spatial features of three-dimensional point cloud / deformation matrix and capture the evolution features of three-dimensional deformation of the ground surface.

[0105] 2. Input the single-modal deep features extracted from the three branches into the improved multimodal Transformer fusion encoder; embed the spatiotemporal location encoding of the spatial topology information and time series information of the coal mine power supply line into the Transformer self-attention layer, adaptively calculate the association weights of different modal features through the cross-modal attention mechanism, fuse complementary features, eliminate feature redundancy, and finally generate a multi-source fusion global geological disaster feature tensor.

[0106] S3.2 Decision-level fusion based on adaptive weighting:

[0107] 1. For the global geological disaster feature tensor output by feature-level fusion, multiple geological disaster identification sub-models are constructed for landslide identification, collapse identification, mining subsidence identification, and ground subsidence identification. Each sub-model outputs the identification probability and risk level of the corresponding geological disaster type.

[0108] 2. Construct an adaptive weighted decision fusion module. Based on DS evidence theory, combine the historical recognition accuracy and data confidence of each single data source, adaptively allocate decision weights for each sub-model, and perform decision-level fusion of the output results of multiple sub-models to eliminate recognition errors of single-modal data. Finally, output accurate recognition results of geological disaster type, spatial distribution range, and occurrence probability.

[0109] 3. The multi-source fusion global geological disaster feature tensor output in this step is a global geological disaster feature matrix in the real domain. At the same time, it outputs the identification probability results of each geological disaster type, providing standardized attribute dimension data and prior information on geological disaster risk for the construction of the spatiotemporal cube in step S4.1.

[0110] Step S4: Multi-dimensional spatiotemporal correlation analysis of geological disasters in coal mine power supply lines based on quaternion spatiotemporal graph convolutional neural network.

[0111] This step is a key technical improvement of the present invention. It addresses the core characteristics of coal mine power supply lines, such as strong spatiotemporal coupling of multiple causes of geological disasters, multidimensional heterogeneous spatial correlation, and temporal evolution with both gradual and sudden changes. Based on spatiotemporal graph convolutional neural networks, it constructs an integrated spatiotemporal correlation analysis system of "point (tower) - line (power supply line) - surface (global geological environment) - volume (three-dimensional space of goaf)". This solves the core pain points of existing technologies that cannot simultaneously capture explicit / implicit spatial dependence, long-term and short-term temporal correlation, and nonlinear coupling of multiple causes.

[0112] The core improvement principle of this invention compared to the prior art is explained as follows: Traditional spatiotemporal analysis methods use feature vector concatenation and multi-graph linear weighting to process multi-source heterogeneous data, which has three major defects: First, vector concatenation will destroy the inherent structural relationship between multi-dimensional features, resulting in feature redundancy and information loss; second, linear weighting cannot capture the high-order nonlinear coupling effect between multiple spatial relationships and multiple causes; and third, it cannot simultaneously capture long-term and short-term time series dependencies, resulting in large cumulative errors in long-term time series prediction.

[0113] This invention proposes applying a quaternion spatiotemporal graph neural network to the geological disaster monitoring scenario of coal mine power supply lines. Utilizing the algebraic structure of quaternions (1 real part + 3 imaginary parts), it naturally adapts to the unified encoding of four core spatial associations in coal mine scenarios. Nonlinear interactive operations between features are achieved through Hamiltonian product, resulting in four core technical advantages compared to traditional methods:

[0114] 1. Complete preservation of structural information: Multi-dimensional features / multi-graph structures are encoded into the real and imaginary parts of quaternions, without destroying the inherent coupling relationship between features and avoiding information loss during vector concatenation;

[0115] 2. Precise capture of nonlinear coupling: High-order nonlinear interaction between features is achieved through Hamiltonian product, which can accurately characterize the complex coupling effect between multiple spatial associations and multiple causes, which cannot be achieved by traditional linear fusion.

[0116] 3. Simultaneous encoding of short-term and long-term time series: Through a cross-time feature aggregation strategy, the features of short-term adjacent and long-term distant time steps are preserved in a single quaternion matrix, and the suddenness and gradual change of the temporal dependence of geological disasters are captured simultaneously.

[0117] 4. Significantly reduced number of parameters: The number of parameters in a quaternion network is only 1 / 4 of that in a real-valued network of the same dimension, which greatly reduces computing power and is suitable for edge deployment in coal mines.

[0118] It should be noted that in the field of coal mine geological disaster monitoring, the spatial correlation of geological disaster evolution is not a single physical topological relationship, but includes multiple heterogeneous relationships such as "physical connection of towers, similar geological properties, temporal evolution semantics, and implicit influence of mining". These heterogeneous relationships are not simply linear superpositions, but have physical effects of mutual modulation and nonlinear excitation. For example, underground mining will continuously weaken the strength of the rock and soil along the line, and simultaneously amplify the triggering effect of heavy rainfall on landslides and collapses. Small changes in a single factor will produce nonlinear disaster amplification effects through the coupling of multiple factors.

[0119] In related technologies, single graph neural networks are usually used to process single graph structures, or multiple graph structures are fused through linear weighting. However, they have always been unable to break through the limitations of linear fusion and cannot characterize the nonlinear coupling effect of the above-mentioned heterogeneous relationships. At the same time, in existing technologies, quaternion neural networks are mostly used in fields such as three-dimensional pose rotation and computer vision, while spatiotemporal graph convolutional neural networks are only used in general scenarios such as traffic flow and general photovoltaic power prediction.

[0120] This application proposes that the algebraic structure of quaternions (satisfying...) The Hamiltonian product operation rules can precisely simulate the nonlinear physical process in coal mine disaster scenarios where multiple factors interact and stimulate each other. Introducing quaternions into spatiotemporal graph convolution is a unique technical solution proposed to address the long-standing unsolved technical problem of "nonlinear coupling modeling of multi-source heterogeneous spatial relationships" in coal mine scenarios.

[0121] This step S4 specifically includes the following four progressively executed sub-steps:

[0122] S4.1 Constructing the disaster spatiotemporal cube and quaternion feature tensor of coal mine power supply lines:

[0123] This step lays a standardized data foundation for spatiotemporal correlation analysis. The core is to convert the multi-source fusion real-valued features output in step S3 into quaternion feature tensors through the cross-time feature aggregation strategy (CTFA). At the same time, it solves the problem of huge differences in time resolution between star-space-ground-well multi-source data and the difficulty in synchronously preserving long- and short-term time series information.

[0124] S4.1.1 Constructing a spatiotemporal cube model specific to coal mine geological disasters:

[0125] Using the multi-source fusion global geological disaster feature tensor output in step S3 as the core, a three-dimensional spatiotemporal cube model is constructed, with the following three dimensions:

[0126] 1. Spatial Dimension: Taking power line towers as the smallest spatial unit, it integrates the spatial topology information of the tower body, geological units along the line, and the impact range of mining subsidence areas, covering the entire spatial scale of "point-line-surface-volume";

[0127] 2. Time dimension: Align with the unified UTC time base in step S2, covering time series data with full time resolution at spaceborne (daily / weekly), airborne (monthly), ground (second / minute), and downhole (hourly) levels, preserving the full-cycle time series characteristics of geological disasters from gradual evolution to sudden instability;

[0128] 3. Attribute Dimension: Integrates five major categories of core attribute features: surface deformation, tower status, geological attributes, meteorological factors, and underground mining activities. It also includes the prior information on geological disaster identification probability output in step S3.2, realizing unified structured management of multi-source heterogeneous features.

[0129] S4.1.2, Quaternion Feature Tensor Construction Based on CTFA Strategy:

[0130] For the real-valued temporal feature matrix output by the spatiotemporal cube, the CTFA strategy is adopted to convert the real-valued features into a quaternion-domain feature tensor, thereby achieving synchronous encoding of long and short-term temporal information. The specific formula is as follows:

[0131] ;

[0132] in For the first Quaternion characteristic matrix of time step, For the real part, , The feature mapping rule between the real and imaginary parts is as follows: (The original text contains three imaginary parts.)

[0133] ;

[0134] ;

[0135] in, : No. The quaternion feature matrix of the time step consists of one real part and three imaginary parts, and is used to synchronously encode long-term and short-term time series information of geological disasters; The real part of the quaternion characteristic matrix, corresponding to the first... Geological disaster baseline characteristics at the time step; The first imaginary part of the quaternion characteristic matrix corresponds to the first imaginary part of the quaternion characteristic matrix. The associated temporal characteristics of time steps; The second imaginary part of the quaternion characteristic matrix, corresponding to the... The associated temporal characteristics of time steps; The third imaginary part of the quaternion characteristic matrix, corresponding to the... The associated temporal characteristics of time steps; The imaginary unit of a quaternion satisfies the rules of arithmetic. ; : No. The original real-valued feature matrix of the time step is used as the real part reference data of the quaternion feature matrix; The time step parameter defines the time interval between the imaginary and real parts of the feature, and performs zero-padding on feature values ​​that are outside the range of the time series. : The current time step, which is the cutoff time node for the time series data; : The total number of historical time steps used to construct the quaternion feature tensor, i.e., selecting past Encode the real-valued features of each continuous time step; Historical time step index, with a value range of: ,pass Traversing the past All data at each time step.

[0136] For different prediction periods and data source characteristics in coal mine scenarios, A multi-scale combination selection method is adopted, and the selection examples are shown in Table 1:

[0137] Table 1

[0138]

[0139] Through the aforementioned CTFA strategy, the real-valued observation sequence at time step T of history is ultimately transformed into a standardized feature tensor in the quaternion field, which serves as the input to the subsequent spatiotemporal graph convolutional neural network model.

[0140] S4.2 Construction of Multi-Dimensional Explicit-Implicit Spatiotemporal Relationship Graphs and Quaternion Adjacency Matrices for Coal Mine Scenarios.

[0141] S4.2.1 Quantification of Nonlinear Causal Relationships of Multidimensional Causes: A nonlinear spatiotemporal Granger causality test algorithm based on kernel functions is adopted to replace the traditional linear Granger causality test, solving the problem of insufficient accuracy of linear tests caused by the highly nonlinear evolution of coal mine geological disasters. The original time-series data is mapped to a high-dimensional feature space through a radial basis kernel function. In this high-dimensional space, the spatiotemporal causal relationship between multidimensional causes such as mining disturbance, rainfall, topography, and soil moisture content and geological disaster deformation evolution is examined. The influence weight and spatiotemporal lag period of each cause are quantified, providing prior support for the subsequent adaptive optimization of the quaternion adjacency matrix weights, and achieving accurate tracing of the core causes of geological disasters.

[0142] S4.2.2 Construction of a multi-dimensional association graph specific to coal mine scenarios.

[0143] Based on multi-dimensional data from spatiotemporal cubes and causal analysis results, one set of implicit correlation graphs and three sets of explicit correlation graphs are constructed to comprehensively cover the full-dimensional spatial correlations of coal mine geological disasters.

[0144] 1. Implicit Association Graph (Sampling-Deformation Implicit Adjacency Matrix) ): A unique correlation graph structure for coal mine scenarios is used. An adaptive graph learning strategy is adopted. Based on underground mining time series data and surface deformation and tower displacement time series data, the potential nonlinear correlation between mining disturbance and geological disaster evolution is automatically learned.

[0145] 2. Explicit Association Figure 1 (Spatial topological adjacency matrix) Based on the topological distance between power supply lines and the spatial Euclidean distance of geological units between towers, this method captures the proximity relationships in the geographical space along the line and depicts the direct transmission characteristics of geological disasters in adjacent spaces.

[0146] 3. Explicit Association Figure 2 (Temporal semantic adjacency matrix) Based on the Dynamic Time Warping (DTW) algorithm, the similarity of deformation time series, rainfall time series, and mining time series of different towers / geological units is calculated, and a temporal semantic association graph is constructed to capture semantic associations that are spatially distant but have highly similar geological disaster evolution patterns.

[0147] 4. Explicit Association Figure 3 (Geological attribute adjacency matrix) Based on the similarity of geological environmental attributes such as rock and soil type, terrain slope, degree of impact of mining subsidence, and vegetation coverage, this method captures the attribute associations of units with the same geological vulnerability, and solves the problem of characterizing the pattern of similar geological disasters occurring in clusters in similar geological environments.

[0148] Sampling-deformation implicit adjacency matrix The adaptive learning strategy is as follows:

[0149] 1. Randomly initialized learnable node embedding matrix Where N is the number of tower nodes and d is the embedding dimension (default is 40). These are learnable weight parameters;

[0150] 2. Node embedding nonlinear mapping: ,in This is a hyperparameter used to control the saturation rate of the activation function (default value is 0.1).

[0151] 3. Preliminary calculation of the implicit correlation matrix: The asymmetric transfer relationship between mining and deformation is captured by antisymmetric matrix constraints;

[0152] 4. Sparsity constraint: using Top-K mask matrix right Sparsity processing is performed, retaining only the top-20 edges with the strongest association strength for each node, filtering out noise interference, and the masking rule is as follows: The final sparsified implicit adjacency matrix is: For Hadamah accumulation;

[0153] 5. Loss Function Design: A joint loss function of "prediction loss + graph regularization loss" is adopted to achieve end-to-end optimization of the implicit graph. The specific formula is as follows: ;

[0154] in, Mining-deformation implicit adjacency matrix is ​​used to characterize the potential nonlinear relationship between downhole mining disturbances and surface deformation and tower displacement. A randomly initialized learnable node embedding matrix is ​​used to map tower nodes to a high-dimensional feature space with dimension 1. ; The total number of tower nodes in the coal mine power supply line serves as the basis for the dimension of the adjacency matrix. Node embedding dimension, with a default value of 40, used to balance feature expressiveness and computational cost; The data type and dimension identifier of the node embedding matrix, representing... OK The real-valued matrix of columns; The data type and dimension identifier of the weight parameter matrix, indicating... OK The real-valued matrix of columns; : A learnable weight parameter matrix used to perform a linear transformation on the node embedding matrix, with dimensions of ; The feature matrix after node embedding nonlinear mapping is the basis for subsequent implicit correlation matrix calculation; The hyperbolic tangent activation function is used to introduce non-linearity into the node embedding map, enhancing feature representation capabilities. Its output range is... ; : A hyperparameter that controls the saturation rate of the activation function. The default value is 0.1, and it is used to adjust the intensity of the nonlinear mapping. : Rectified linear unit activation function, used to filter negative values ​​in the implicit correlation matrix and retain valid correlation information; Matrix multiplication operation refers to... and Multiplying the transpose of the matrix is ​​used to capture the association between the two sets of node embeddings; Matrix multiplication operation refers to... and Multiply by the transpose of the matrix, and... Together they form an antisymmetric matrix constraint; Top-K: a filtering operation used to retain the top-ranked nodes in terms of their association strength for each tower. (here) The associated edges of the adjacency matrix are used to achieve sparsification of the adjacency matrix; M: Top-K mask matrix, used to modify the initial calculations. Perform sparsification to filter out noisy, correlated edges; : Mask matrix In the Line number The elements of the column, with values ​​of 1 or 0, are used to indicate whether the associated edges at the corresponding positions are retained; : Index return function, input is ( No. (All elements in a row), the output is the top elements in that row with the highest correlation strength. The column index corresponding to the element; : Dynamic-deformation implicit adjacency matrix No. All elements in the row, corresponding to the first The strength of the connection between each tower node and all other nodes; The Hadamard product operator is used to multiply corresponding elements of two matrices of the same dimension; here, it is used to... With mask matrix Perform sparse fusion; Mean absolute error loss for deformation prediction is the core loss term of the joint loss function, used to measure the deviation between the model's predicted deformation and the actual deformation. Regularization weight coefficients, each taking a value of [value]. This is used to adjust the weight ratio of the two types of regularization loss; F-norm regularization term, used to avoid Overfitting, expressed as: ; The square of the F-norm, where the F-norm is the square root of the sum of the squares of all elements of the matrix, is used to quantify the overall size of the matrix and impose regularization constraints. Node embedding regularization terms are used for constraints. To ensure smoothness and avoid overfitting of node embeddings, the expression is: ; yes The square of the F-norm is used for quantization. The overall size and the application of regularization constraints; yes The square of the F-norm is used for quantization. The overall size is determined and regularization constraints are applied.

[0155] S4.2.3 Quaternion Graph Adjacency Matrix Construction: Utilizing the inherent structural characteristic of quaternions ("1 real part + 3 imaginary parts"), the above four sets of association graphs are integrated into a unified quaternion graph adjacency matrix. In this process, a unified nonlinear encoding for multi-graph structures is implemented, using the following formula:

[0156] ;

[0157] in, Quaternion graph adjacency matrix is ​​used to integrate one set of implicit association graphs and three sets of explicit association graphs to achieve unified nonlinear encoding of multi-graph structures, providing a foundation for subsequent nonlinear interactions of Hamiltonian products. The mining-deformation implicit adjacency matrix, as the real part of the quaternion graph adjacency matrix (unique to the coal mine scenario), characterizes the potential nonlinear relationship between underground mining disturbances and surface deformation and tower displacement. The spatial topological explicit adjacency matrix, as the first imaginary component of the quaternion graph adjacency matrix, is constructed based on the topological distance between power supply lines between towers and the spatial Euclidean distance of geological units to capture geographic spatial proximity relationships. The temporal semantic explicit adjacency matrix, as the second imaginary component of the quaternion graph adjacency matrix, is constructed based on the dynamic time warping (DTW) algorithm to calculate temporal similarity and capture semantic associations that are spatially distant but have highly similar geological disaster evolution patterns. The explicit adjacency matrix of geological attributes, as the third imaginary component of the quaternion graph adjacency matrix, is constructed based on the similarity of geological environmental attributes such as rock and soil type and terrain slope, and captures the association of units with the same geological vulnerability. , , The imaginary unit of a quaternion satisfies the rules of arithmetic. , used to mark the imaginary parts of the quaternions corresponding to the three explicit adjacency matrices respectively; +: quaternion addition operator, used to add the real parts With the three imaginary components ( , , It is integrated into a unified quaternion graph adjacency matrix, preserving the inherent relationships of the multi-graph structure; : scalar multiplication operator, used to combine each explicit adjacency matrix with its corresponding imaginary unit to generate the imaginary part of the quaternion, thereby realizing the quaternion field encoding of the explicit association graph.

[0158] In this formula, the real part is the mining-deformation implicit adjacency matrix unique to the coal mine scenario, and the three imaginary parts correspond to three sets of explicit adjacency matrices. Through the quaternion Hamiltonian product operation, the high-order nonlinear coupling effect between the four sets of graph structures can be directly captured, replacing the traditional linear weighted fusion scheme.

[0159] S4.3 Deep Extraction and Evolutionary Modeling of Geological Disaster Spatiotemporal Features Based on the Collaborative Development of Spatiotemporal Graph Convolutional Neural Network: This step is the core execution link of spatiotemporal correlation analysis. Based on the network architecture of alternating collaboration between 1D quaternion convolutional module (1DQC) and quaternion graph convolutional module (QGC), the temporal evolution features and spatial correlation features of geological disasters are deeply extracted, and an end-to-end spatiotemporal evolution model of geological disasters is constructed.

[0160] S4.3.1 Spatiotemporal Graph Convolutional Neural Network Basic Network Architecture Design:

[0161] The network architecture adopts a stacked structure of "1×1 convolutional layer + 3 groups of 1DQC-QGC alternating units + residual connection + output layer":

[0162] 1. The 1×1 convolutional layer is responsible for mapping the input quaternion feature tensor to the latent space, thus completing the feature dimension adaptation;

[0163] 2. In each 1DQC-QGC alternating unit, the temporal features are first extracted through the 1DQC module, and then the spatial features are extracted through the QGC module, so as to realize the alternating deep mining of spatiotemporal features;

[0164] 3. Set a residual connection between the input of the 1DQC module and the output of the QGC module to avoid the gradient vanishing problem during model training and improve the convergence performance of deep networks.

[0165] 4. The output layer is a real-valued linear layer. The real and imaginary parts of the quaternion features are concatenated and mapped to the prediction dimension to output the quantitative prediction results of geological disaster evolution.

[0166] The model training uses multi-source fusion features from historical time periods as input, and uses the cumulative InSAR surface deformation, the geological disaster locations and risk levels confirmed by field surveys as supervision labels. The network parameters are optimized end-to-end through the Adam optimizer and backpropagation algorithm.

[0167] S4.3.2, 1D Quaternion Convolution Module (1DQC) Temporal Feature Extraction: The core of this module is to implement one-dimensional convolution operations based on quaternion Hamiltonian products, simultaneously capturing the long-term and short-term dependencies of geological disaster time series, thus overcoming the limitation of traditional 1D convolution which can only capture local short-term temporal features.

[0168] 1. The module consists of two 1D quaternion convolutional layers plus gated linear units (GLUs). The convolutional kernels are designed based on the quaternion field, and each convolution operation is implemented through a Hamiltonian product. The GLU (Gated Linear Unit) is a gated linear unit function that performs non-linear filtering on the output of the 1D quaternion convolution, preserving temporal characteristics strongly correlated with the evolution of coal mine geological hazards (such as landslides triggered by heavy rainfall and collapses indicated by minor needs) while suppressing irrelevant noise interference. The specific formula is as follows:

[0169] ;

[0170] in, Let be the quaternion feature tensor input to the I-th layer. The quaternion convolution kernel of layer I, For the quaternion Hamiltonian product, The kernel size is the convolution kernel size. The number of input feature channels; : No. The quaternion feature tensor output by the 1D quaternion convolution module is in The elements of the location are used to characterize the feature values ​​that integrate the long-term and short-term temporal dependencies of coal mine geological disasters at that location; : Network layer index of the 1D quaternion convolution module, used to distinguish convolution operations at different levels; Double summation operator, traversing input feature channels ( ) and kernel size ( ( ) dimension, to complete the cumulative calculation of quaternion convolution; Input feature channel index, with a value ranging from 1 to... This is used to iterate through all input feature channels; : No. The number of input feature channels in a 1D quaternion convolution layer represents the dimensionality of the input quaternion features; : Kernel size index, with a value ranging from 1 to This is used to iterate through all dimensions of the convolution kernel; : No. The size of the quaternion convolution kernel determines the length of the coal mine geological disaster time window captured by the convolution operation; : No. Layer quaternion convolution kernel in The location elements are designed based on the quaternion field and are specifically optimized for extracting the temporal features of coal mine geological disasters; : No. The quaternion feature tensor of the layer input is in The element at the given position is the input data for the 1D quaternion convolution operation; The quaternion Hamiltonian product operator is the core operation of 1D quaternion convolution and quaternion graph convolution. It is used to realize the nonlinear interaction between quaternion domain features / matrices and capture the high-order coupling effect of multi-dimensional features of coal mine geological disasters.

[0171] 2. For coal mine geological disaster scenarios, the module can simultaneously capture short-term sudden time-series dependencies (such as the triggering effect of 24-hour heavy rainfall on landslides, and the precursor indication of microseismic events on collapse) and long-term slowly changing time-series dependencies (such as the cumulative impact of months of mining operations on surface subsidence, and the continuous weakening of rock and soil strength by seasonal rainfall).

[0172] 3. Gated linear units perform nonlinear filtering on the convolution output, retaining temporal features that are strongly correlated with the evolution of geological disasters, suppressing irrelevant noise interference, and improving the accuracy of temporal feature extraction.

[0173] S4.3.3 Quaternion Graph Convolution Module (QGC) Spatial Feature Extraction: The core of this module is to implement graph convolution operations based on the quaternion Hamiltonian product, capturing high-order nonlinear dependencies between multi-dimensional explicit and implicit graph structures, thus overcoming the limitation of traditional graph convolution which can only handle single graph structures.

[0174] 1. The core of the module is a quaternion graph convolutional layer, using a quaternion adjacency matrix constructed with S4.2. Using the temporal features output by module 1 as input, graph convolution is performed through Hamiltonian product, as shown in the following formula: ;

[0175] in, The quaternion features are the input of the I-th layer. For learnable quaternion weight parameters, It is the ReLU activation function; : No. The quaternion features output by the layer quaternion graph convolution module integrate the high-order nonlinear dependencies of the multi-graph structure of coal mine geological disaster spatial dimension. : Time step identifier, corresponding to the time sequence node of coal mine geological disaster monitoring data (such as hourly or daily monitoring time). : Network layer index of quaternion graph convolutional layer This represents the output feature of the next layer after the current layer; The ReLU activation function performs a nonlinear transformation on the result of quaternion graph convolution, filters out negative values, and enhances the expressive power of spatial features of coal mine geological disasters. Quaternion graph adjacency matrix integrates the mining-deformation implicit adjacency matrix and three sets of explicit adjacency matrices, providing a spatial association structure specific to coal mine scenarios for quaternion graph convolution; : No. The quaternion features input to the layer are the temporal features output by the 1D quaternion convolution module, which are the initial feature inputs for the quaternion graph convolution. : No. The layer-learnable quaternion weight parameters are used to adaptively transform the input features to meet the spatial feature extraction requirements of coal mine geological disaster scenarios.

[0176] 2. For coal mine scenarios, the module can simultaneously model the nonlinear coupling effects of four types of spatial relationships: tower topology, geological attributes, temporal semantics, and mining-induced implicit correlations, accurately capturing the spatial transmission patterns of geological disasters along power supply lines.

[0177] 3. Lightweight Model Adaptation: Weight parameters are optimized through quaternion low-rank matrix decomposition. Compression optimization is performed to further reduce the number of parameters and computing power without sacrificing model accuracy, thus adapting to the deployment needs of edge terminals in coal mines.

[0178] The final output layer of the network consists of two branches: the main branch uses the Sigmoid activation function to output the probability of geological disaster occurrence for each tower node at a specified future time. The auxiliary regression branch uses a linear activation function to output the predicted value of the cumulative surface deformation at a specified future time. (unit: This provides a quantitative basis for subsequent risk classification and early warning.

[0179] S4.4 Multi-timescale Advanced Prediction and Risk Transmission Path Analysis of Geological Disasters: Based on the spatiotemporal graph convolutional neural network model of geological disasters trained in S4.3, this model enables multi-timescale advanced prediction of geological disasters, identification of risk transmission paths, and location of key risk nodes, providing core decision-making basis for subsequent graded early warning.

[0180] S4.4.1 Multi-timescale advanced prediction of geological disaster evolution: Based on model output, accurate predictions are achieved across three time scales, covering the full-scenario needs of coal mine power supply line operation and maintenance.

[0181] 1. Short-term forecast (24h): Focusing on sudden geological disasters such as landslides and collapses triggered by sudden factors such as rainfall and micro-earthquakes, predicting the surface deformation rate and the probability of geological disaster occurrence in the next 24 hours, and achieving early warning of impending disasters;

[0182] 2. Mid-term forecast (7 days): Focusing on the evolution trend of geological disasters induced by continuous mining disturbances and continuous rainfall, predict the development trajectory of surface deformation and changes in risk level in the next 7 days, and reserve a window period for emergency response;

[0183] 3. Long-term forecast (30 days): Focusing on slow-change geological disasters such as mining subsidence and large-scale ground subsidence, predicting the evolution range of geological disasters and their impact on tower foundations in the next 30 days, providing advanced planning basis for line operation and maintenance and comprehensive geological disaster management.

[0184] S4.4.2 Geological Hazard Risk Transmission Path Analysis and Key Node Identification: Based on the spatial correlation weights output by the QGC module and the model prediction results, geological hazard risk transmission path analysis is conducted.

[0185] 1. Quantify the direction and intensity of risk transmission between power line towers and geological units, and identify the dominant transmission path of geological disaster evolution;

[0186] 2. Identify key risk nodes along the power line (such as towers in areas affected by mining subsidence, towers on steep slopes, and towers in geologically vulnerable areas), and quantify the impact of these risk nodes on the safe operation of the entire power supply line.

[0187] 3. Based on the results of the cause tracing, identify the dominant inducing factors at different risk nodes, and provide precise support for differentiated geological disaster prevention and control and line reinforcement solutions.

[0188] Step S5: Geological Disaster Risk Classification, Early Warning, and Full Life Cycle Management: Based on the spatiotemporal correlation analysis and prediction results of Step S4, and combined with the safety operation specifications for coal mine power supply lines, a geological disaster risk classification and evaluation system is constructed, such as... Figure 3 As shown, risks are divided into four levels: Level I (red, extremely high risk), Level II (orange, high risk), Level III (yellow, medium risk), and Level IV (blue, low risk). Corresponding early warning strategies are triggered for different risk levels, and the spatial location, impact range, risk level, evolution trend, cause analysis, and disposal suggestions of geological disasters are pushed to the coal mine operation and maintenance management platform. Based on the big data platform, the entire life cycle of monitoring data, fusion results, and analysis and early warning information is visualized and managed, supporting a single map display of geological disaster risks across the entire power line area, providing a basis for decision-making in the operation and maintenance of coal mine power supply lines and geological disaster prevention.

[0189] The implementation process and effects of the above method are further explained below with reference to several embodiments:

[0190] Example 1: Comprehensive monitoring scenario of 110kV coal mine power supply line.

[0191] This embodiment focuses on a 110kV power supply line for a coal mine. The line is 25km long and traverses three coal mine goaf areas and multiple geologically vulnerable steep slopes in mountainous areas. It carries risks of multiple geological disasters, including landslides and goaf collapses. The specific implementation process of the above method is as follows:

[0192] 1. Multi-source data acquisition and aggregation: Acquire Gaofen-3 InSAR data (12-day revisit period, 2mm deformation accuracy) and Gaofen-2 optical imagery to achieve full-line, full-area surface monitoring; conduct UAV LiDAR and oblique photogrammetry aerial surveys once a month to acquire centimeter-level DSM / DEM data; deploy GNSS monitoring stations and tilt sensors on all 32 towers along the line, and deploy crack gauges, rain gauges, soil moisture sensors, and microseismic monitors in key sections to achieve minute-level real-time dynamic monitoring; simultaneously collect mining progress, goaf distribution, and mine pressure monitoring data from two underground mining faces; aggregate all data into a data lake based on a Spark-based distributed big data platform.

[0193] 2. Spatiotemporal reference unification and preprocessing: Unify all data to the CGCS2000 coordinate system and UTC time reference, with spatial registration accuracy controlled within 5cm; achieve time axis alignment of multi-source data through the Flink streaming engine, with time alignment error controlled within 1 minute; perform denoising, outlier removal, GAN completion and standardization processing on time series data; perform correction, land cover segmentation and feature extraction on image data; and process InSAR and point cloud data to generate a temporal deformation feature matrix.

[0194] 3. Two-level deep fusion of multi-source data: Deep features of each modality are extracted by Bi-LSTM, ResNet50 and 3D-CNN respectively, and input into a multimodal Transformer with embedded spatiotemporal location coding to complete feature-level fusion. Then, decision-level fusion is completed by DS evidence theory, which finally realizes accurate identification of geological disasters such as landslides and mining subsidence, with an identification accuracy of 95.2%, which is 32% higher than traditional methods.

[0195] 4. Spatiotemporal correlation analysis and prediction using spatiotemporal graph convolutional neural networks:

[0196] 4.1 Construct a spatiotemporal cube model of the line, using a multi-scale combination of [1,3,7]. The parameters are used to generate quaternion feature tensors through the CTFA strategy;

[0197] 4.2 Construct a quaternion adjacency matrix, with the real part being the mining-deformation implicit adjacency matrix and the three imaginary parts corresponding to the spatial topology, temporal semantics, and geological attribute adjacency matrices, respectively. Through nonlinear spatiotemporal Granger causality test, mining disturbance and heavy rainfall were identified as the core causes of geological disasters along this line, with influence weights of 58% and 27%, respectively.

[0198] 4.3. Construct a spatiotemporal graph convolutional neural network model with three sets of alternating 1DQC-QGC units, complete model training and optimization, and realize spatiotemporal evolution modeling of geological disasters. In this embodiment, the 1DQC and QGC modules are based on the PyTorch deep learning framework and implemented through a custom quaternion multiplication operator. Model training and inference are completed on an NVIDIA RTX 3090 GPU. The full data inference time for a single line is less than 30 seconds, which can adapt to the real-time analysis needs of the edge of the coal mine.

[0199] 4.4 Based on the model output, the deformation trend and risk level of the next 24 hours, 7 days and 30 days are predicted, and two Class II high-risk goaf impact sections are identified, triggering an orange alert.

[0200] Step S5, Tiered Early Warning and Implementation: The early warning information is simultaneously pushed to the coal mine operation and maintenance management platform, along with the scope of the geological disaster impact, its evolution trend, the main causes, and suggestions for tower reinforcement. Based on the early warning information, operation and maintenance personnel conduct on-site verification and foundation reinforcement, effectively avoiding power outages caused by geological disasters and ensuring safe production in the coal mine.

[0201] Example 2: Special monitoring scenario for landslide-prone areas in mountainous regions.

[0202] This embodiment focuses on the K8-K12km section of the aforementioned 110 kV line. This section is characterized by steep mountain slopes and is free from mining subsidence. The main risks are rainfall-induced landslides and collapses. The core parameters for this embodiment are differentiated as follows:

[0203] 1. Parameter selection: using The single-scale parameters focus on capturing short-term temporal abrupt changes in minute-level rainfall and microseismic data, adapting to the needs of landslide disaster early warning;

[0204] 2. Quaternion adjacency matrix weight optimization: The real part implicit adjacency matrix is ​​replaced with the rainfall-deformation implicit correlation matrix to focus on learning the potential nonlinear correlation between rainfall and surface deformation; the imaginary part weight is tilted towards the temporal semantic adjacency matrix and the geological attribute adjacency matrix to focus on capturing the landslide clustering patterns of similar geological units;

[0205] 3. Implementation Results: In this scenario, the method of the present invention achieves a landslide identification accuracy rate of 100%. The average lead time for disaster warnings is 6 hours, and compared with the traditional spatiotemporal graph neural network method, the long-term prediction error is reduced by 28% and the false alarm rate is reduced by 42%.

[0206] Example 3: Special monitoring scenario for core image section of goaf collapse.

[0207] This embodiment focuses on the K18-K22km section of the aforementioned 110kV line. This section directly traverses the goaf of the first mining face in a coal mine, with mining-induced surface subsidence and uneven settlement as the main risks. The core parameters of this embodiment are differentiated as follows:

[0208] 1. Parameter selection: using With a single-scale parameter of 7, it focuses on capturing the long-term gradual trend of weekly InSAR deformation data and monthly mining progress data, which is suitable for the long-term evolution prediction of goaf collapse.

[0209] 2. Quaternion adjacency matrix weight optimization: The weight of the real part mining-deformation implicit adjacency matrix is ​​increased to 60%, focusing on learning the spatiotemporal transmission law of underground mining disturbance and surface subsidence; the weight of the imaginary part is tilted towards the spatial topological adjacency matrix, focusing on capturing the transmission path of collapse risk between adjacent towers;

[0210] 3. Implementation Results: In this scenario, the method of this invention achieved a prediction accuracy of 94.3% for the collapse of the goaf area, and the long-term prediction accuracy over 30 days was improved by 31% compared with the traditional method. It successfully predicted the risk of uneven settlement of a tower foundation 15 days in advance, leaving sufficient window period for operation and maintenance.

[0211] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods according to the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, or network device, etc.) to execute the methods of the various embodiments of this application.

[0212] Embodiments of this application also provide a system for fusing and spatiotemporal correlation analysis of multi-source heterogeneous geological disaster monitoring data for coal mine power supply lines, such as... Figure 4 As shown, the system includes:

[0213] The data acquisition module 42 is used to collect multi-source geological disaster data of coal mine power supply lines and to standardize the multi-source geological disaster data to obtain a standardized multi-source heterogeneous dataset. The multi-source geological disaster data includes: spaceborne remote sensing data, airborne aerial survey data, ground tower monitoring data and underground mining data.

[0214] Feature fusion module 44 is used to sequentially perform feature-level fusion and decision-level fusion on the standardized multi-source heterogeneous dataset to obtain the global geological disaster feature tensor in the real-valued domain.

[0215] The geological disaster prediction module 46 is used to analyze the global geological disaster feature tensor through a coal mine geological disaster spatiotemporal correlation analysis model constructed based on a quaternion algebra structure, and output the geological disaster prediction results for multiple areas of the coal mine power supply line. The output layer of the coal mine geological disaster spatiotemporal correlation analysis model is divided into two branches: the main branch uses the Sigmoid activation function to output the probability value of geological disaster occurrence for each tower node at a specified future time; the auxiliary regression branch uses the linear activation function to output the predicted value of cumulative surface deformation at a specified future time; and performs geological disaster risk classification and early warning based on multiple geological disaster prediction results. The multiple geological disaster prediction results correspond one-to-one with the multiple areas, and the geological disaster prediction results include: the probability of geological disaster occurrence and the predicted result of surface deformation.

[0216] The aforementioned system collects four types of geological disaster data: spaceborne remote sensing, airborne aerial surveying, ground tower monitoring, and underground mining activity data related to coal mine power supply lines. After spatiotemporal benchmark unification and standardization, a standardized multi-source heterogeneous dataset is constructed. Feature-level fusion and decision-level fusion are then performed sequentially on the standardized multi-source heterogeneous dataset to generate a real-valued global geological disaster feature tensor. This global geological disaster feature tensor is input into a spatiotemporal correlation analysis model based on a quaternion algebra structure. Quaternion encoding is used to simultaneously capture the nonlinear coupling effects of long-term and short-term temporal dependencies and multi-dimensional spatial correlations of geological disasters, outputting the probability of geological disaster occurrence and surface deformation for each region. The method predicts the magnitude of geological disasters and uses this to complete the geological disaster risk classification and early warning. This solution introduces a quaternion algebraic structure to realize high-order nonlinear fusion and spatiotemporal correlation modeling of multi-source heterogeneous data, which significantly improves the accuracy of geological disaster identification and the advance warning of coal mine power supply lines. It is particularly suitable for coal mine-specific scenarios such as mining-induced geological disasters. This solves the problem in related technologies that cannot simultaneously capture the high-order nonlinear coupling effect of multi-dimensional explicit / implicit spatial correlation and long-short-term time series dependence in the evolution of geological disasters in coal mine power supply lines, resulting in low geological disaster identification accuracy, delayed early warning and difficulty in adapting to mining-induced geological disaster scenarios unique to coal mines.

[0217] Optionally, the aforementioned geological disaster prediction module 46 is further configured to encode the global geological disaster feature tensor into a quaternion feature tensor using a cross-time feature aggregation strategy. Each element of the quaternion feature tensor consists of a real part and three imaginary parts. The real part represents the geological disaster baseline feature at the current time step, and the three imaginary parts are associated temporal features with preset cross-time step intervals. The module also constructs a quaternion graph adjacency matrix corresponding to the coal mine power supply line. This quaternion graph adjacency matrix uses the tower nodes of the coal mine power supply line as graph nodes. The structure of the quaternion graph adjacency matrix corresponds one-to-one with the quaternion feature tensor. The real part corresponds to the mining-deformation implicit adjacency matrix of the quaternion graph adjacency matrix. The three imaginary parts... The imaginary parts correspond to the spatial topological explicit adjacency matrix, the temporal semantic explicit adjacency matrix, and the geological attribute explicit adjacency matrix, respectively. A quaternion spatiotemporal graph neural network is constructed, wherein the quaternion spatiotemporal graph neural network includes alternating stacked one-dimensional quaternion convolution modules and quaternion graph convolution modules. The coal mine geological disaster spatiotemporal correlation analysis model includes the quaternion spatiotemporal graph neural network. The one-dimensional quaternion convolution module performs one-dimensional convolution operation based on the quaternion Hamiltonian product to extract long- and short-term temporal correlation features in the quaternion feature tensor. The quaternion graph convolution module performs graph convolution operation based on the quaternion Hamiltonian product to capture the high-order nonlinear coupling effect of multi-dimensional spatial correlation in the quaternion graph adjacency matrix.

[0218] Optionally, the aforementioned geological disaster prediction module 46 is further configured to encode the global geological disaster feature tensor into a quaternion feature tensor using a cross-time feature aggregation strategy, including: the construction formula of the quaternion feature tensor is: ,in, Let be the quaternion characteristic matrix at time step (t-T+n). For the real part, , , Let i, j, and k be the imaginary units of the quaternion, satisfying the following condition: The feature mapping rule between the real part and the three imaginary parts is as follows: , , , , The geological disaster baseline characteristics at time step (t-T+n) are as follows. For the ( The associated temporal characteristics of time steps, For the ( The associated temporal characteristics of time steps, For the ( The associated temporal characteristics of time steps, This is a preset time step.

[0219] Optionally, the preset time step The prediction scenarios employ a multi-scale combination setting, including: for 24-hour short-term disaster early warning scenarios, A value of 1 is used to adapt to minute-level data from ground sensors; this is for scenarios involving 7-day mid-term trend prediction. A value of 3 is used to adapt to hourly downhole mining data; for 30-day long-term evolution prediction scenarios. A value of 7 is used to adapt to weekly InSAR satellite data; this is suitable for general scenarios across the entire data lifecycle. A multi-scale combination of [1,3,7] is adopted.

[0220] Optionally, the aforementioned geological disaster prediction module 46 is also used to predict disasters using formulas. Construct the adjacency matrix of the quaternion graph, where, The quaternion graph adjacency matrix is... This refers to the sampling-deformation implicit adjacency matrix. Let be the explicit adjacency matrix of the spatial topology. The temporal semantic explicit adjacency matrix is... Let i, j, k be the quaternion imaginary unit, satisfying the following conditions: .

[0221] Optionally, the aforementioned geological disaster prediction module 46 is also used to initialize a randomly learnable node embedding matrix. Where N is the number of tower nodes and d is the embedding dimension; the feature matrix is ​​obtained through nonlinear mapping. , ,in, To control the hyperparameters of the activation function saturation rate, The weight parameters are learnable; the implicit incidence matrix is ​​initially calculated. ; using Top-K mask matrix to Sparsity processing is performed, retaining the top K edges with the strongest association strength for each node, to obtain the final mining-deformation implicit adjacency matrix. This mining-deformation implicit adjacency matrix is ​​obtained through a joint loss function. Achieve end-to-end optimization The mean absolute error loss for deformation prediction. The F-norm regularization term is used to avoid... Overfitting Regularization terms are used to embed nodes to avoid overfitting of node embeddings. This is the regularization weight coefficient.

[0222] Optionally, the aforementioned geological disaster prediction module 46 is also used for the spatial topological explicit adjacency matrix. The temporal semantic explicit adjacency matrix is ​​constructed based on the topological distance of power supply lines between tower nodes and the spatial Euclidean distance of geological units; Based on the dynamic time warping algorithm, the similarity of deformation time series, rainfall time series, and mining time series of different tower nodes or geological units is constructed. The multiple regions include the different tower nodes or geological units. The geological attribute explicit adjacency matrix... The similarity is constructed based on geological environmental attributes, which include: rock and soil type, terrain slope, degree of impact of mining subsidence, and vegetation coverage.

[0223] Optionally, the convolution operation formula of the one-dimensional quaternion convolution module includes: ,in, Let be the quaternion feature tensor input to the I-th layer. The quaternion convolution kernel of layer I, For the quaternion Hamiltonian product, The kernel size is the convolution kernel size. Let f be the number of input feature channels, r be the spatial location index, s be the output channel index, t be the time step index, and k be the kernel size index. For the first The quaternion feature tensor output by the one-dimensional quaternion convolution module is in Positional elements, used to characterize Location-based fusion of long- and short-term temporal dependence features of coal mine geological disasters; : No. Layer quaternion convolution kernel in The location elements are designed based on the quaternion field and are specifically optimized for extracting the temporal features of coal mine geological disasters; : No. The quaternion feature tensor of the layer input is in The element at the given position is the input data for the 1D quaternion convolution operation.

[0224] Optionally, the graph convolution operation formula of the quaternion graph convolution module includes: ,in, The quaternion features are the input of the I-th layer. For learnable quaternion weight parameters, It is a quaternion graph adjacency matrix. It is the ReLU activation function. It is the Hamiltonian product of quaternions.

[0225] Embodiments of this application also provide a storage medium including a stored program, wherein the program, when executed, performs any of the methods described above. Optionally, in this embodiment, the storage medium may be configured to store program code for performing the steps of the methods described above.

[0226] Optionally, in this embodiment, the storage medium may include, but is not limited to, various media capable of storing program code, such as USB flash drives, read-only memory (ROM), random access memory (RAM), portable hard drives, magnetic disks, or optical disks.

[0227] Embodiments of this application also provide an electronic device, including a memory and a processor, wherein the memory stores a computer program and the processor is configured to run the computer program to perform the steps in any of the above method embodiments.

[0228] Optionally, the electronic device may further include a transmission device and an input / output device, wherein the transmission device is connected to the processor and the input / output device is connected to the processor.

[0229] Embodiments of this application also provide a computer program product, including a non-volatile computer-readable storage medium storing the computer program product, wherein the computer program, when executed by a processor, implements the steps of the methods described in various embodiments of this application.

[0230] Optionally, in this embodiment, the computer program described above can be configured to implement the steps in any of the above method embodiments when executed by a processor.

[0231] Optionally, specific examples in this embodiment can refer to the examples described in the above embodiments and optional implementations, and will not be repeated here.

Claims

1. A method for fusing and spatiotemporal correlation analysis of multi-source heterogeneous geological disaster monitoring data for coal mine power supply lines, characterized in that, include: Multi-source geological disaster data of coal mine power supply lines are collected and the multi-source geological disaster data are standardized to obtain a standardized multi-source heterogeneous dataset. The multi-source geological disaster data includes: spaceborne remote sensing data, airborne aerial survey data, ground tower monitoring data and underground mining data. The standardized multi-source heterogeneous dataset is subjected to feature-level fusion and decision-level fusion in sequence to obtain the global geological disaster feature tensor in the real-valued domain; A spatiotemporal correlation analysis model for coal mine geological disasters, constructed based on a quaternion algebraic structure, is used to analyze the global geological disaster feature tensor and output geological disaster prediction results for multiple areas of the coal mine power supply line. The output layer of the spatiotemporal correlation analysis model for coal mine geological disasters is divided into two branches: the main branch uses the Sigmoid activation function to output the probability value of geological disaster occurrence for each tower node at a specified future time; the auxiliary regression branch uses the linear activation function to output the predicted value of cumulative surface deformation at a specified future time. Based on the multiple geological disaster prediction results, geological disaster risk classification and early warning are performed. The multiple geological disaster prediction results correspond one-to-one with the multiple areas. The geological disaster prediction results include: probability of geological disaster occurrence and predicted surface deformation results.

2. The method according to claim 1, characterized in that, The spatiotemporal correlation analysis model of coal mine geological disasters, constructed based on quaternion algebraic structures, is used to analyze the global geological disaster feature tensor, including: A cross-time feature aggregation strategy is adopted to encode the global geological disaster feature tensor into a quaternion feature tensor. Each element of the quaternion feature tensor consists of a real part and three imaginary parts. The real part is the geological disaster baseline feature of the current time step, and the three imaginary parts are the associated temporal features with a preset cross-time step interval. A quaternion graph adjacency matrix is ​​constructed corresponding to the power supply line of the coal mine. The quaternion graph adjacency matrix uses the tower nodes of the power supply line of the coal mine as graph nodes. The structure of the quaternion graph adjacency matrix corresponds one-to-one with the quaternion feature tensor. The real part corresponds to the mining-deformation implicit adjacency matrix of the quaternion graph adjacency matrix. The three imaginary parts correspond to the spatial topology explicit adjacency matrix, the temporal semantic explicit adjacency matrix, and the geological attribute explicit adjacency matrix, respectively. A quaternion spatiotemporal graph neural network is constructed, wherein the quaternion spatiotemporal graph neural network includes alternating stacked one-dimensional quaternion convolution modules and quaternion graph convolution modules, and the coal mine geological disaster spatiotemporal correlation analysis model includes the quaternion spatiotemporal graph neural network; The one-dimensional quaternion convolution module performs one-dimensional convolution operations based on the quaternion Hamiltonian product to extract long-short-term temporal correlation features from the quaternion feature tensor. The quaternion graph convolution module performs graph convolution operations based on the quaternion Hamiltonian product to capture the high-order nonlinear coupling effect of multi-dimensional spatial correlation in the quaternion graph adjacency matrix.

3. The method according to claim 2, characterized in that, The global geological disaster feature tensor is encoded into a quaternion feature tensor using a cross-temporal feature aggregation strategy, including: The formula for constructing the quaternion feature tensor is: ,in, Let be the quaternion characteristic matrix at time step t-T+n. For the real part, , , Let i, j, and k be the imaginary units of the quaternion, satisfying the following condition: The feature mapping rule between the real part and the three imaginary parts is as follows: , , , , Let be the geological disaster baseline characteristics at time step t-T+n. For the first The associated temporal characteristics of time steps For the first The associated temporal characteristics of time steps For the first The associated temporal characteristics of time steps Preset time step; This is the time step index, representing the current time step; : The total number of historical time steps used to construct the quaternion feature tensor; Index for historical time steps; For spatial location index.

4. The method according to claim 3, characterized in that, The preset time step The prediction scenario employs a multi-scale combination setting, including: For 24-hour short-term disaster early warning scenarios, A value of 1 is used to adapt to minute-level data from ground sensors; For scenarios involving 7-day medium-term trend prediction, The value is set to 3, which is suitable for hourly downhole mining data. For the 30-day long-term evolution prediction scenario, A value of 7 is used to adapt to weekly InSAR satellite data. For general-purpose scenarios throughout the entire lifecycle, A multi-scale combination of [1,3,7] is adopted.

5. The method according to claim 2, characterized in that, Constructing the quaternion graph adjacency matrix corresponding to the coal mine power supply line includes: Through formula Construct the adjacency matrix of the quaternion graph, where, The quaternion graph adjacency matrix is... This refers to the sampling-deformation implicit adjacency matrix. Let be the explicit adjacency matrix of the spatial topology. This is the explicit adjacency matrix of the temporal semantics. Let i, j, k be the quaternion imaginary unit, satisfying the following conditions: .

6. The method according to claim 5, characterized in that, The mining-deformation implicit adjacency matrix Constructed through an adaptive graph learning strategy, including: Randomly initialized learnable node embedding matrix Where N is the number of tower nodes and d is the embedding dimension; The characteristic matrix is ​​obtained through nonlinear mapping. , ,in, To control the hyperparameter of the activation function saturation rate, These are learnable weight parameters; Preliminary calculation of implicit incidence matrix ; Using the Top-K mask matrix Sparsity processing is performed, retaining the top K edges with the strongest association strength for each node, to obtain the final mining-deformation implicit adjacency matrix. This mining-deformation implicit adjacency matrix is ​​obtained through a joint loss function. Achieve end-to-end optimization The mean absolute error loss for deformation prediction. The F-norm regularization term is used to avoid... Overfitting Regularization terms are used to embed nodes to avoid overfitting of node embeddings. This is the regularization weight coefficient.

7. The method according to claim 5, characterized in that, The method further includes: The spatial topological explicit adjacency matrix Constructed based on the topological distance of power supply lines between tower nodes and the spatial Euclidean distance of geological units; The temporal semantic explicit adjacency matrix Based on the dynamic time warping algorithm, the similarity of deformation time series, rainfall time series, and mining time series of different tower nodes or geological units is constructed, wherein the multiple regions include the different tower nodes or geological units; The geological attribute explicit adjacency matrix The similarity is constructed based on geological environmental attributes, which include: rock and soil type, terrain slope, degree of impact of mining subsidence, and vegetation coverage.

8. The method according to claim 2, characterized in that, The convolution operation formula of the one-dimensional quaternion convolution module includes: ,in, Let be the quaternion feature tensor input to the I-th layer. The quaternion convolution kernel of layer I, For the quaternion Hamiltonian product, The kernel size is the convolution kernel size. Let f be the number of input feature channels, r be the spatial location index, s be the output channel index, t be the time step index, and k be the kernel size index. For the first The quaternion feature tensor output by the one-dimensional quaternion convolution module is in Positional elements, used to characterize Location-based fusion of long- and short-term temporal dependence features of coal mine geological disasters; For the first Layer quaternion convolution kernel in The element at the specified position; For the first The quaternion feature tensor of the layer input is in The element at a given position.

9. The method according to claim 2, characterized in that, The graph convolution operation formula of the quaternion graph convolution module includes: ,in, For the first Quaternion features output by the layer quaternion graph convolution module; The quaternion features are the input of the I-th layer. For learnable quaternion weight parameters, It is a quaternion graph adjacency matrix. It is the ReLU activation function. It is the Hamiltonian product of quaternions.

10. A system for fusing and spatiotemporal correlation analysis of multi-source heterogeneous geological disaster monitoring data for coal mine power supply lines, characterized in that, include: The data acquisition module is used to collect multi-source geological disaster data of coal mine power supply lines and to standardize the multi-source geological disaster data to obtain a standardized multi-source heterogeneous dataset. The multi-source geological disaster data includes: spaceborne remote sensing data, airborne aerial survey data, ground tower monitoring data, and underground mining data. The feature fusion module is used to sequentially perform feature-level fusion and decision-level fusion on the standardized multi-source heterogeneous dataset to obtain the global geological disaster feature tensor in the real-valued domain. The geological disaster prediction module is used to analyze the global geological disaster feature tensor using a coal mine geological disaster spatiotemporal correlation analysis model constructed based on a quaternion algebra structure, and output geological disaster prediction results for multiple areas of the coal mine power supply line. The output layer of the coal mine geological disaster spatiotemporal correlation analysis model is divided into two branches: the main branch uses the Sigmoid activation function to output the probability value of geological disaster occurrence for each tower node at a specified future time; the auxiliary regression branch uses the linear activation function to output the predicted value of cumulative surface deformation at a specified future time. Based on the multiple geological disaster prediction results, geological disaster risk classification and early warning are performed. The multiple geological disaster prediction results correspond one-to-one with the multiple areas. The geological disaster prediction results include: probability of geological disaster occurrence and predicted surface deformation results.

Images