A Dynamic Analysis Method for the Weights of Geological Hazard Susceptibility Evaluation Factors for Coal Mine Power Supply Lines Incorporating Attention Mechanism

Abstract

This invention belongs to the technical field of coal mine power supply safety and geological disaster assessment. It discloses a dynamic analysis method for the weighting of geological disaster susceptibility evaluation factors in coal mine power supply lines, incorporating an attention mechanism. The method includes: constructing an evaluation factor system comprising static factors, dynamic factors, and mining-related influence factors; constructing a weighted directed factor network based on the temporal causal co-occurrence relationship between dynamic factors in historical geological disaster cases; acquiring real-time monitoring data for each dynamic factor, calculating the self-state contribution, network propagation contribution, and global context contribution of each dynamic factor, and obtaining the dynamic attention weight after weighted summation and normalization; calculating the geological disaster susceptibility index based on the dynamic attention weight using a weighted information content model; and achieving dynamic adaptive adjustment of the weights through time triggering, event triggering, and network structure updates. This invention can integrate real-time monitoring data, significantly improving the timeliness and accuracy of geological disaster susceptibility assessment.

Description

Technical Field

[0001] This invention relates to the field of coal mine power supply safety and geological disaster assessment technology, specifically to a dynamic analysis method for the weight of geological disaster susceptibility evaluation factors for coal mine power supply lines that incorporates an attention mechanism. Background Technology

[0002] Coal mine power supply lines often traverse complex geological environments such as mountainous areas, river valleys, and mining subsidence areas, making them susceptible to geological hazards such as landslides, debris flows, ground subsidence, and tower foundation deformation. Currently, in geological hazard susceptibility assessments, the weights of assessment factors (such as slope, aspect, lithology, fault distance, river distance, vegetation cover index, rainfall, peak ground acceleration, and distance to mining subsidence areas) are mostly determined statically using the analytic hierarchy process (AHP), entropy weight method, or principal component analysis. Existing technologies, such as "A Landslide Susceptibility Assessment Method Integrating Convolutional Neural Networks and Self-Attention Models" (CN117094234A) and "A Geological Hazard Risk Assessment Method and System Integrating Random Forests and Attention" (CN116167617A), although incorporating attention mechanisms, still have the following shortcomings:

[0003] (1) The weight cannot dynamically respond to environmental changes: continuous heavy rainfall will significantly increase the importance of the "rainfall amount" factor, but the static weight cannot be automatically adjusted.

[0004] (2) Ignoring the causal propagation relationship between factors: Geological disasters are often the result of the coupling effect of multiple factors (such as "fault + rainfall + steep slope" jointly triggering landslides). Traditional weighted methods are difficult to quantify this nonlinear interaction effect, and existing attention methods only focus on feature importance and do not model the directional propagation relationship between factors.

[0005] (3) Difficulty in utilizing real-time monitoring data: Existing methods are mostly based on static geographic information and cannot integrate real-time data such as rain gauges and displacement sensors to correct the weights.

[0006] Therefore, a method is needed to dynamically and objectively update the weights of evaluation factors in order to improve the timeliness and accuracy of the assessment of the susceptibility of power supply lines to geological disasters. Summary of the Invention

[0007] To address the shortcomings of existing technologies, the present invention aims to provide a dynamic analysis method for the weights of evaluation factors for geological hazard susceptibility of coal mine power supply lines that integrates an attention mechanism. For areas along coal mine power supply lines, a weighted directed factor network that distinguishes between static and dynamic factors is constructed, and a dynamic weight update function based on causal propagation and self-attention is designed to achieve adaptive adjustment of the evaluation factor weights.

[0008] To achieve the above objectives, the following technical solution is adopted:

[0009] This invention provides a dynamic analysis method for the weights of geological hazard susceptibility evaluation factors for coal mine power supply lines, incorporating an attention mechanism, comprising the following steps:

[0010] S1: For the area along the power supply line of the coal mine, select environmental factors related to the occurrence of geological disasters to form an evaluation factor set. The environmental factors include static factors, dynamic factors and mining-related factors.

[0011] S2: Based on historical geological disaster cases, directed edges are established according to the temporal causal co-occurrence relationship between dynamic factors, and edge weights are calculated according to the frequency of occurrence of the directed edges in all historical cases to construct a weighted directed factor network among dynamic factors; wherein, static factors and mining impact factors are used as spatial constraints to modify the edge weights.

[0012] S3: Obtain real-time monitoring data for each dynamic factor and define a dynamic attention weight for each dynamic factor; the dynamic attention weight is obtained by normalizing the weighted sum of three factors: its own state contribution, network propagation contribution, and global context contribution.

[0013] S4: Based on the dynamic attention weight, the geological disaster susceptibility index of each evaluation unit along the power supply line is calculated using a weighted information volume model, and the susceptibility zone level is divided according to the susceptibility index;

[0014] S5: Based on a preset time period or when the monitoring value of a dynamic factor exceeds a preset threshold, trigger step S3 to update the dynamic attention weight of each dynamic factor; at the same time, update the sampling impact factor on a monthly or yearly basis, and the sampling impact factor does not trigger the execution of step S3.

[0015] Furthermore, in step S2, the directed edges and edge weights of the weighted directed factor network are determined as follows: a danger threshold is pre-set for each dynamic factor, and when the monitored value of a dynamic factor exceeds the threshold, it is determined to be in an abnormal state; if in historical geological disaster cases, the abnormal state of the first dynamic factor occurs before the abnormal state of the second dynamic factor, then a directed edge is established from the first dynamic factor to the second dynamic factor; the edge weight is equal to the number of times the second dynamic factor becomes abnormal after the first dynamic factor becomes abnormal in historical cases divided by the total number of times the first dynamic factor becomes abnormal.

[0016] Furthermore, the method for correcting the edge weights of the static factor is as follows: for each directed edge, the proportion of the current static factor in a dangerous state in the historical events corresponding to each directed edge is counted, and the larger value between this proportion and the preset benchmark coefficient is taken as the correction coefficient; in the network propagation contribution calculation, the weight of the directed edge is multiplied by the correction coefficient to obtain the corrected propagation strength; the historical events refer to the events in which the second dynamic factor becomes abnormal after the first dynamic factor becomes abnormal.

[0017] Furthermore, in step S3, the calculation method for the self-state contribution is as follows: for each dynamic factor, the monitoring value within the most recent time window is collected, and the monitoring values ​​within the window are weighted and averaged, where the weight of recent data is greater than that of long-term data, to obtain the current comprehensive state; the current comprehensive state is input into a logistic regression model obtained by fitting historical disaster data, and the output value is the self-state contribution.

[0018] Furthermore, in step S3, the network propagation contribution is calculated as follows: for the current dynamic factor, obtain all the previous driving factor nodes pointing to it in the weighted directed factor network, multiply the dynamic attention weight, the corresponding directed edge weight, and the static factor correction coefficient of each previous driving factor at the previous moment, and then sum them up to obtain the network propagation contribution.

[0019] Furthermore, in step S3, the global context contribution is calculated as follows: the time window monitoring sequence of each dynamic factor is mapped to an embedding vector through a linear embedding layer; multi-head self-attention is calculated on the embedding vector sequences of all dynamic factors; the position vector corresponding to each dynamic factor in the multi-head self-attention output sequence is taken, and then converted into a scalar through a fully connected layer, which is the global context contribution of the dynamic factor.

[0020] Furthermore, in step S3, the normalized calculation method for the dynamic attention weights is as follows:

[0021]

[0022] in, For the i-th dynamic factor The contribution value of its own state at the current time t; For the i-th dynamic factor The network propagation contribution value at current time t; For the i-th dynamic factor The global context contribution value at the current time t; , , These are the preset coefficients for the corresponding contribution items, and their sum is 1. m is the total number of dynamic factors, and j is the index of the dynamic factor.

[0023] Furthermore, in step S4, the geological hazard susceptibility index... Calculated using a weighted information content model:

[0024]

[0025] in, Dynamic factors in historical disasters The number of cases in a dangerous state, where N is the total number of historical geological disaster cases. To evaluate dynamic factors within the region The number of grid cells in a dangerous state, where S is the total number of grid cells in the evaluation area.

[0026] Furthermore, step S5 specifically includes:

[0027] Time-triggered update: Real-time monitoring data of each dynamic factor is re-collected every preset number of days and step S3 is executed; Event-triggered update: Step S3 is executed immediately when the monitoring value of any dynamic factor exceeds a preset threshold; Network structure update: The directed relationship between dynamic factors is re-statistically calculated every preset time period or when the number of new geological disaster cases exceeds a preset threshold, and the adjacency matrix and edge weights of the weighted directed factor network are updated.

[0028] Furthermore, the method also includes: a parameter optimization step: based on historical geological disaster cases, the data is divided into a training set and a validation set; multiple sets of different preset coefficient combinations are selected, and the susceptibility index of each disaster point in the validation set before its occurrence is calculated; the accuracy of disaster point identification is used as an evaluation index, and the preset coefficient combination that maximizes this index is selected as the optimal parameter.

[0029] Compared with the prior art, the present invention has the following beneficial effects:

[0030] 1. Achieved dynamic adaptive adjustment of evaluation factor weights: In traditional geological hazard susceptibility assessments, factor weights are fixed once determined, failing to respond to changes in dynamic environmental factors such as rainfall and slope displacement. This invention constructs a dynamic weight update function that includes self-state contribution, network propagation contribution, and global self-attention contribution, enabling weights to automatically adjust with changes in real-time monitoring data, significantly improving the evaluation model's responsiveness under extreme weather and sudden conditions.

[0031] 2. Quantifying the causal propagation relationship between disaster-causing factors: Existing attention methods only focus on the independent importance of each factor, ignoring the directed coupling effect between factors. This invention constructs a weighted directed factor network to solidify the temporal causal co-occurrence relationship between dynamic factors in historical cases in the form of edge weights, and introduces a network propagation contribution term in the weight update. This allows dynamic weights to reflect the transmission effect of risk along the causal chain, improving the physical interpretability of the evaluation results.

[0032] 3. This invention resolves the logical contradiction in mixed modeling of static and dynamic factors: static factors such as slope and fault distance do not change significantly over time, making it impossible to establish edges with dynamic factors according to a unified time-series rule. This invention distinguishes between static and dynamic factors, using static factors as conditional probability correction coefficients for edge weights and setting a baseline value. Preventing the undue disruption of risk transmission aligns with engineering realities and ensures the logical consistency of the model.

[0033] 4. Using time window sequences as self-attention input overcomes the deficiency of insufficient single-point information: The recent monitoring sequence of each dynamic factor, rather than single-point values, is input into the multi-head self-attention network, enabling the model to learn the temporal evolution pattern of factor states. Experiments demonstrate that this design can still obtain meaningful global context contribution values ​​through a pre-training-transfer learning strategy even with only 18 historical cases, avoiding overfitting caused by data sparsity.

[0034] 5. Frequency-based dynamic update mechanism balances real-time performance and data availability: Update strategies are differentiated based on the rate of change of factors, distinguishing between high-frequency (rainfall, slope displacement) and low-frequency (mineralized area) updates, combining event-triggered and time-triggered mechanisms. This method ensures timely response in high-risk scenarios while reducing the data acquisition and computation costs for low-frequency factors, thus improving the feasibility of engineering deployment.

[0035] 6. A parameter optimization method based on AUC maximization avoids subjective parameter tuning: the coefficients in the attention weight function are determined through historical backtracking verification and grid search. Using the AUC value of disaster point identification as the optimization objective, the model parameters have an objective and reproducible basis for determination. The results of the embodiments show that the dynamic weight model optimized by this invention (AUC=0.81) has a significantly improved evaluation accuracy compared to the static weight model (AUC=0.72).

[0036] 7. Excellent data adaptability and transferability: When historical disaster samples from the target mining area are insufficient, the embedding and attention layer parameters of the multi-head self-attention network can use pre-trained weights from publicly available geological disaster datasets (such as the NASA Global Landslide Database), with only the linear layers fine-tuned or kept fixed. This transfer learning strategy effectively solves the training challenge of deep learning models in small-sample scenarios, enabling the method to be extended to newly established mining areas or regions with scarce historical data.

[0037] It should be understood that the description in the Summary of the Invention is not intended to limit the key or essential features of the embodiments of the present invention, nor is it intended to restrict the scope of the invention. Other features of the invention will become readily apparent from the following description. Attached Figure Description

[0038] The above and other features, advantages, and aspects of the various embodiments of the present invention 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 invention. In the drawings, the same or similar reference numerals denote the same or similar elements, wherein:

[0039] Figure 1 This is a flowchart illustrating the dynamic analysis method for evaluating the geological hazard susceptibility of coal mine power supply lines based on an incorporating attention mechanism, according to an embodiment of the present invention.

[0040] Figure 2 This is a schematic diagram of the dynamic attention weight update process according to an embodiment of the present invention;

[0041] Figure 3 This is a spatial distribution map of geological hazard susceptibility assessment of a 110kV power supply line in a mining area, generated using the method of this invention.

[0042] Figure 4 This is a comparison chart of the effects of dynamic weights and traditional static weights in an embodiment of the present invention. Detailed Implementation

[0043] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0044] 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.

[0045] Figure 1 This is a flowchart illustrating the dynamic analysis method for evaluating the geological hazard susceptibility of coal mine power supply lines based on an incorporating attention mechanism, according to an embodiment of the present invention. Figure 2 This is a schematic diagram of the dynamic attention weight update process according to an embodiment of the present invention. Specifically, see... Figure 1 and Figure 2A dynamic analysis method for weighting evaluation factors of geological hazard susceptibility of coal mine power supply lines, incorporating attention mechanisms, includes the following steps:

[0046] S1: For the area along the power supply line of the coal mine, select environmental factors related to the occurrence of geological disasters to form an evaluation factor set. The environmental factors include static factors, dynamic factors and mining-related factors.

[0047] Step S1 is used to construct a geological hazard evaluation factor system for coal mine power supply lines.

[0048] For the area along the target power supply line, environmental factors closely related to the occurrence of geological disasters were selected to form an evaluation factor set. Factors are divided into three categories:

[0049] Static factors: Topographic and geological properties with a rate of change lower than the interannual scale, including slope, aspect, elevation, lithology, fault distance, river distance, vegetation cover index, and peak ground acceleration.

[0050] Dynamic factors: Environmental variables that can be monitored in real time or at high frequency through sensors or remote sensing, including rainfall (cumulative over multiple days), soil moisture content (retrieved by ground sensors or remote sensing), and slope displacement rate;

[0051] Mining-induced influencing factors: distance to goaf, goaf area, and surface subsidence rate. Mining-induced influencing factors are obtained with a medium-to-low frequency update cycle (monthly or annually). They do not participate in the real-time dynamic attention weight calculation in step S3, but instead participate in the edge weight correction in step S2 as spatial constraints along with static factors. They also participate as fixed background information in the susceptibility index calculation in step S4.

[0052] S2: Based on historical geological disaster cases, directed edges are established according to the temporal causal co-occurrence relationship between dynamic factors, and edge weights are calculated according to the frequency of occurrence of the directed edges in all historical cases to construct a weighted directed factor network among dynamic factors; wherein, static factors and mining impact factors are used as spatial constraints to modify the edge weights.

[0053] Step S2 is used to construct a weighted directed factor network based on historical disaster cases, including: the method for determining the directed edges and edge weights of the weighted directed factor network is as follows: a danger state threshold is pre-set for each dynamic factor, and when the monitored value of a dynamic factor exceeds the threshold, it is determined to be in an abnormal state; if, in historical geological disaster cases, the abnormal state of the first dynamic factor occurs before the abnormal state of the second dynamic factor, then a directed edge is established from the first dynamic factor to the second dynamic factor. The specific process is as follows:

[0054] Historical geological hazard records (including landslides, collapses, debris flows, and ground subsidence) were collected from the target area and adjacent areas. Each record included the location and time of the hazard, as well as the state values ​​of each factor. Two types of relationships were established based on the different factor types:

[0055] (1) Causal triggering relationship (applicable to dynamic factors and dynamic factors → disaster)

[0056] Establish directed edges based on time series: If in a certain historical case, dynamic factors... Abnormal conditions (such as 7-day rainfall > 50 mm) appear in dynamic factors If an abnormal state (such as soil moisture content > 0.3) occurs before the occurrence of an anomaly, and the two are correlated in time and space, then a directed edge is established from i to j to represent the direction of risk propagation.

[0057] (2) Spatial coupling relationship (applicable between static factors and dynamic factors / disasters)

[0058] Static factors themselves do not change over time and do not serve as "precursor nodes" for causal propagation, but rather participate in modeling as spatial constraints. For each directed edge i→j (where i,j are dynamic factors or disaster nodes), the frequency of the dangerous states of each static factor simultaneously satisfying when the directed relationship occurs is statistically analyzed in all historical cases to form conditional probabilities, which are used as correction coefficients for edge weights.

[0059] Network construction results: Node set V includes dynamic factor nodes and disaster-occurring virtual nodes D; directed edge weights. Defined as:

[0060]

[0061] in The count represents the number of times dynamic factor j becomes abnormal after dynamic factor i becomes abnormal in historical cases. This represents the total number of times the dynamic factor i is in an abnormal state.

[0062] Similar to static factors, mining-induced impact factors (distance from goaf, goaf area, and surface subsidence rate) also participate in edge weight correction as spatial constraints due to their slow rate of change and lack of real-time time-series monitoring sequences. For each directed edge, the proportion of mining-induced impact factors in a dangerous state when the directed relationship occurs in historical cases is statistically analyzed, and the correction coefficient is calculated in the same way as for static factors.

[0063] S3: Obtain real-time monitoring data for each dynamic factor and define a dynamic attention weight for each dynamic factor; the dynamic attention weight is obtained by normalizing the weighted sum of three factors: its own state contribution, network propagation contribution, and global context contribution.

[0064] Step S3 is used to design the attention weight update function.

[0065] Let the time-series monitoring sequences of each dynamic factor for the current time period (e.g., the past 7 days) be the input. For each dynamic factor... Collect the monitoring values ​​of the most recent L days (e.g., L=7) to form a time window vector. All values ​​have been normalized to the [0,1] interval. Where L: time window length, representing the number of days since the most recent monitoring data for each dynamic factor was collected; in this embodiment, the value is 7. t: current time, representing the time point at which the dynamic weights are updated. Dynamic factors The normalized monitoring value at time t ranges from [0,1]. The dangerous state of the static factor serves as fixed background information and does not participate in real-time updates, but it is corrected by conditional probability when calculating the network propagation contribution.

[0066] Define dynamic factors Dynamic attention weights It consists of three parts:

[0067] 1. Contribution based on one's own status:

[0068] The calculation method for the self-state contribution is as follows: For each dynamic factor, the monitoring value within its most recent time window is collected, and a weighted average is calculated on the monitoring values ​​within the window, with the weight of recent data being greater than that of long-term data, to obtain the current comprehensive state; the current comprehensive state is then input into a logistic regression model fitted using historical disaster data, and the output value is the self-state contribution. Details are as follows:

[0069] The current overall state is obtained by taking a weighted average of the time series windows (with more recent data having higher weights). ,in This is the attenuation coefficient. Then:

[0070]

[0071] in, This is the sigmoid function. λ: The decay coefficient, used to assign higher weight to recent data when calculating the weighted average of monitored values ​​within the time window; in this embodiment, it is set to 0.8. k: The offset within the time window, ranging from 0 to L-1, where k=0 represents the current time and k=L-1 represents the earliest time within the window. Dynamic factors The weighted average of the monitored values ​​within the time window reflects the current overall status. The Sigmoid function maps the input to... Interval. The weight parameters of the Sigmoid function in the self-state contribution term are obtained by fitting historical disaster data. The bias parameter of the Sigmoid function in the self-state contribution term is obtained by fitting historical disaster data. : The i-th dynamic factor The self-state contribution value at the current time t is the output of the Sigmoid function.

[0072] parameter Fitting: Based on the monitoring values ​​of each dynamic factor before 18 historical disasters as positive samples, 36 spatiotemporal points (twice the number of disaster points) were randomly selected within the study area where no disasters occurred as negative samples. The label was whether a disaster occurred (positive samples were labeled 1, negative samples were labeled 0). Logistic regression was used to fit the data and obtain the fitting values ​​of each dynamic factor. and L2 regularization (regularization coefficient C=1.0) is used during training to prevent overfitting.

[0073] 2. Contribution to online dissemination:

[0074] The network propagation contribution is calculated as follows: For the current dynamic factor, all preceding driver factor nodes pointing to it in the weighted directed factor network are obtained. The dynamic attention weight, corresponding directed edge weight, and static factor correction coefficient of each preceding driver factor at the previous time step are multiplied and summed to obtain the network propagation contribution. The static factor edge weight correction is performed as follows: For each directed edge, the proportion of the current static factor in a dangerous state in the historical events corresponding to each directed edge (i.e., events where the second dynamic factor becomes abnormal after the first dynamic factor becomes abnormal) is calculated. The larger value between this proportion and a preset baseline coefficient is taken as the correction coefficient. In the network propagation contribution calculation, the weight of the directed edge is multiplied by the correction coefficient to obtain the corrected propagation strength. The details are as follows:

[0075] Propagation from other dynamic factors to The risk intensity is defined as:

[0076]

[0077] in, Let i be the set of pre-driving factor nodes. For network edge weights, This is the static factor correction coefficient.

[0078] Initial value setting: Initial time t=0 The results are assigned using either equal weighting or static weighting based on the analytic hierarchy process (AHP). The current weighting is saved after each update. As for the next update .

[0079] Static factor correction coefficient: , ,in, This represents the proportion of historical cases where the directed edge appears and the current static factor is in a dangerous state. =0.15 is the baseline coefficient.

[0080] : The i-th dynamic factor The network propagation contribution value at time t represents the propagation from other dynamic factors to... The intensity of the risk. : Points to the set of preceding driving factor nodes of dynamic factor node i. j: Pre-driving factor node, belonging to the set. . : The dynamic attention weight of the pre-driving factor j at the previous time t-1. From dynamic factors The weight of the directed edge pointing to the dynamic factor i is equal to the number of times i becomes abnormal after j becomes abnormal in historical cases, divided by the total number of times j becomes abnormal. : Static factor correction coefficient, used to correct the risk propagation intensity from j to i. t=0: Initial time, representing the starting point before the first calculation of dynamic weights. Dynamic factors The initial weights are determined using either equal weights or static weighting results based on the analytic hierarchy process (AHP). Current moment Dynamic attention weight vector for all dynamic factors. : The baseline value of the static factor correction coefficient, which is 0.15 in this embodiment, is used to ensure that the basic propagation capability is retained even when the static factor is in the safe range. In historical cases, for a directed edge j→i, the proportion of the current static factor that is in a dangerous state when the directed relation occurs.

[0081] 3. Global context contribution:

[0082] The global context contribution is calculated as follows: A multi-head self-attention mechanism is used to calculate the mutual influence between dynamic factors. This includes: mapping the time window monitoring sequence of each dynamic factor to an embedding vector through a linear embedding layer; performing multi-head self-attention calculation on the embedding vector sequences of all dynamic factors; taking the position vector corresponding to each dynamic factor in the multi-head self-attention output sequence, and then converting it to a scalar through a fully connected layer, which is the global context contribution of that dynamic factor. The specific implementation process is as follows:

[0083] Each dynamic factor Time window sequence Mapped to a d-dimensional vector through a learnable linear embedding layer. , where d is the embedding dimension, representing the dimension that maps the time window sequence to a vector; It is a dynamic factor The time window sequence obtained after being mapped by the embedding layer Dimensional embedding vector.

[0084] For embedded sequences Multi-head self-attention computation is performed, with an attention structure of 1 multi-head attention layer and h=4 heads.

[0085]

[0086] in, ;

[0087] Take the i-th position vector of the output sequence and pass it through a fully connected layer (containing ReLU activation) to obtain a scalar. . For the first One dynamic factor The global context contribution value at the current time t.

[0088] Parameter sources: The parameters of the embedding layer, attention layer and fully connected layer are obtained through supervised training when there is sufficient historical disaster data; when the amount of data is insufficient, the parameters adopt the weights of the public geological disaster dataset after pre-training, and remain fixed in the target mining area or only fine-tune the linear layer.

[0089] The publicly available dataset used for pre-training is the NASA Global Landslide Catalog (GLC). Monitoring records related to the dynamic factor types (rainfall, soil moisture content, slope displacement) of this invention were selected to construct time-series samples for self-supervised pre-training. The pre-training task employs masked time-series prediction: 20% of the time steps in the input sequence are randomly masked, and the model is trained to reconstruct the masked values. Parameters of the embedding layer, multi-head attention layer, and fully connected layer all participate in pre-training. When applied to the target mining area, these parameters are frozen, and only the logistic regression part is modified. , Make minor adjustments or keep it fixed.

[0090] Embed: A learnable linear embedding layer function used to map input sequences to a high-dimensional space. Current moment The sequence consisting of the embedding vectors of all dynamic factors, i.e. m: Total number of dynamic factors. h: Number of attention heads in the multi-head self-attention mechanism, which is set to 4 in this embodiment. : The output of the kth attention head. Q: The weight matrix of the multi-head self-attention output layer, used to map the concatenated multi-head outputs to the final output dimension. Q: The query matrix in multi-head self-attention, composed of the embedding sequence. Multiply by the weight matrix K is obtained: the key matrix in multi-head self-attention, derived from the embedding sequence. Multiply by the weight matrix V: The value matrix in multi-head self-attention, derived from the embedding sequence. Multiply by the weight matrix get. : Learnable weight matrices used to calculate query, key, and value in the multi-head self-attention layer.

[0091] The final attention weights are:

[0092]

[0093] in, , , These are the preset coefficients for the corresponding contribution items, and their sum is 1. m is the total number of dynamic factors, and j is the index of the dynamic factor, satisfying... Its initial value is set to [0.4, 0.3, 0.3], and it is subsequently optimized and adjusted in step S6. Dynamic factors The global context contribution value is obtained by passing the i-th position vector of the multi-head self-attention output sequence through a fully connected layer. : Contribution to one's own status Preset coefficients in attention weight calculation. Contribution to online dissemination Preset coefficients in attention weight calculation. Global context contribution item Preset coefficients in attention weight calculation. Dynamic factors The dynamic attention weight at time t is obtained by weighted summation of the three contributions followed by Softmax normalization.

[0094] S4: Based on the dynamic attention weight, the geological disaster susceptibility index of each evaluation unit along the power supply line is calculated using a weighted information volume model, and the susceptibility zone level is divided according to the susceptibility index;

[0095] Step S4 is used to implement susceptibility evaluation based on dynamic weights.

[0096] A weighted information content model is used to calculate the geological hazard susceptibility index for each evaluation unit (such as a grid or line segment) along the power supply line:

[0097]

[0098] In the formula: Dynamic factors in historical disasters Number of cases in a hazardous state. Method for determining the hazardous state threshold: For continuous factors (such as rainfall), the optimal cutoff value is determined by using ROC curves based on the distribution of the factor in historical disaster cases; non-disaster samples are constructed by randomly selecting raster cells that have not experienced disasters in the study area, with a positive to negative sample ratio of approximately 1:2; discrete factors (such as lithology) are directly classified by category. : Geological hazard susceptibility index at current time t, used to quantify the geological hazard susceptibility of each evaluation unit along the power supply line. m: Total number of dynamic factors. i: Index number of the dynamic factor. Dynamic factors Dynamic attention weights at time t. N: Total number of historical geological disaster cases. : Dynamic factors within the evaluation area Number of grid cells in a hazardous state. S: Total number of grid cells in the evaluation area. : The i-th dynamic factor.

[0099] Normalization method: For continuous factors, Min-Max normalization is applied. For the interval, outliers are smoothed using a moving average filter (window length 3).

[0100] according to Values ​​are used to classify areas into low, medium, and high-risk zones, and the spatial distribution of high-risk areas is output. See details. Figure 3 . Figure 3 This paper presents a spatial distribution map of geological hazard susceptibility of a 110kV power supply line in a mining area generated using the method of this invention. The results show that this method can accurately identify the section from K3+200 to K4+800 as a high-risk section. This area highly overlaps with the boundary of the goaf, the F1 fault zone, and the steep slope area, and is completely matched with the on-site geological conditions and historical landslide hazards, thus verifying the accuracy of the risk positioning of this invention in the spatial dimension.

[0101] S5: Based on a preset time period or when the monitoring value of a dynamic factor exceeds a preset threshold, trigger step S3 to update the dynamic attention weight of each dynamic factor; at the same time, update the sampling impact factor on a monthly or yearly basis, and the sampling impact factor does not trigger the execution of step S3.

[0102] Step S5: Dynamic update mechanism, including: Time-triggered update: Real-time monitoring data of each dynamic factor is re-collected every preset number of days and step S3 is executed; Event-triggered update: Step S3 is executed immediately when the monitoring value of any dynamic factor exceeds a preset threshold; Network structure update: The directed relationships between dynamic factors are re-statistically analyzed every preset time period or when the number of newly added geological disaster cases exceeds a preset threshold, and the adjacency matrix and edge weights of the weighted directed factor network are updated. Details are as follows:

[0103] (1) Time-triggered update: Every T days (e.g., 7 days), real-time monitoring data of each dynamic factor is re-collected, and step S3 is executed to update. .

[0104] (2) Event-triggered update: When the monitoring value of any high-frequency dynamic factor exceeds the preset threshold (such as daily rainfall > 50 mm, slope displacement > 3 mm / day), step S3 is executed immediately to trigger weight update.

[0105] Low-frequency factor updates: The area of ​​mined-out areas, the rate of surface subsidence, etc. are updated on a monthly or yearly basis and do not participate in event triggering.

[0106] Network structure update: Every six months or whenever more than three new geological disaster cases are added, the directed relationships are recalculated, and the adjacency matrix and edge weights are updated.

[0107] In this embodiment of the invention, S6 is also included: a parameter optimization step, comprising: dividing historical geological disaster cases into a training set and a validation set; selecting multiple different preset coefficient combinations to calculate the susceptibility index of each disaster point in the validation set before its occurrence; and selecting the preset coefficient combination that maximizes the disaster point identification accuracy as the optimal parameter. Specifically, as follows:

[0108] To determine The optimal combination is determined by historical backtracking verification:

[0109] (1) Based on historical geological disaster cases, one part is used as the training set and the other part is used as the validation set;

[0110] (2) For each group of candidates ( (Using a grid search with a step size of 0.05), calculate the susceptibility index of each disaster point in the validation set before the occurrence of the disaster. ;

[0111] (3) The accuracy of disaster point identification (AUC value, i.e., area under the ROC curve) is used as the evaluation index;

[0112] (4) Select the parameter combination that maximizes the AUC value as .

[0113] Specific Example: Dynamic Evaluation of Landslide Susceptibility of a 110kV Power Supply Line in a Mining Area

[0114] 1.1 Overview of the Study Area

[0115] A 110kV power supply line from an open-pit coal mine in southwestern my country, with a total length of 12.5km, was selected. The terrain along the line is highly undulating, and there are 7 historical landslide records (2018–2024). At the same time, 11 landslide records from two adjacent mining areas with similar geological conditions were collected, for a total of 18 historical cases used for network construction.

[0116] 1.2 Factor System and Data Sources

[0117] Selected factors: (1) Static factors: slope (°), distance from fault (m), lithology (soft rock / hard rock); (2) Dynamic factors: 7-day cumulative rainfall (mm), soil moisture index, slope displacement rate (mm / day); (3) Mining impact factor (medium frequency): distance from mined-out area (m).

[0118] Data sources and update frequency: Slope, fault, lithology: Geological survey report (static, one-time input); 7-day cumulative rainfall: Automatic weather station in the mining area (daily, real-time); Soil moisture content: TDR soil moisture sensors deployed at key points (hourly, real-time), supplemented by Sentinel-1 SAR inversion (every 12 days, used for regional areal correction); Raw data were processed by moving average filtering (window length 3) to eliminate measurement noise; Slope displacement rate: GNSS displacement monitoring points deployed (daily, real-time); Distance to goaf: Mining engineering plan provided by the mine (updated monthly).

[0119] 1.3 Network Construction

[0120] Based on 18 historical landslide cases, time series of dynamic factor anomalies were extracted. For example, the preceding events of a certain landslide were: 7-day cumulative rainfall reached 120 mm (abnormal) → soil moisture content increased to 0.35 (abnormal) → slope displacement rate increased to 5 mm / day (abnormal) → landslide occurred. The directed co-occurrence frequency among dynamic factors was statistically analyzed to obtain directed edges and their weights.

[0121] (1) Rainfall → Soil moisture content: 12 times, weight 0.67;

[0122] (2) Soil moisture content → slope displacement: 10 times, weight 0.56;

[0123] (3) Slope displacement → disaster: 14 times, weight 0.78.

[0124] Static factor correction coefficient Taking a slope greater than 30° as an example, in all 12 cases of "rainfall → soil moisture content", 10 cases simultaneously met the condition of a slope greater than 30°, so the frequency is 0.83. .

[0125] 1.4 Example of Attention Weight Calculation

[0126] Taking a heavy rainfall event on June 15, 2023 as an example (data from actual measurements at the mining area meteorological station):

[0127] 1) Monitoring sequence (last 7 days, L=7, attenuation coefficient) =0.8):

[0128] Rainfall: → Normalized value after weighted average =0.92;

[0129] Soil moisture content: → Weighted average normalization =0.88;

[0130] Slope displacement: → Weighted average normalization =0.35.

[0131] 2) Initial weights Let the weight be equal to 1 / 3 = 0.333.

[0132] 3) Parameters Fitting results: Based on the monitoring values ​​of various dynamic factors before 18 historical disasters, and using whether a disaster occurred as the label, logistic regression was used for fitting, resulting in:

[0133] Rainfall: , Soil moisture content: , Slope displacement: , Substitute Calculation: Rainfall: Soil moisture content: Slope displacement: .

[0134] 4) Contribution to online dissemination (taken) ):

[0135] From the previous time point (for simplicity, assume the weight of each dynamic factor is 0.333 at time t-1): Rainfall =0 (no precursor); Soil moisture content =0.333×0.67×0.83=0.185; Slope displacement =0.333×0.56×0.85=0.158.

[0136] 5) Global Self-Attention Contribution: The temporal window sequences (each 7-dimensional) of the three dynamic factors are embedded into 64-dimensional vectors (the parameters of the embedding layer and attention layer use pre-trained weights from a public landslide dataset; in this embodiment, these parameters are frozen and not fine-tuned). After passing through a multi-head attention layer (1 layer, h=4), the output is passed through a fully connected layer to obtain a scalar: =0.32, =0.28, =0.22 (obtained by inference from a pre-trained model in a TensorFlow 2.4 environment).

[0137] 6) Calculate the weighted sum and softmax: Rainfall: Soil moisture content: Slope displacement: Normalized rainfall Soil moisture content Slope displacement .

[0138] 1.5 Susceptibility Assessment Results

[0139] Substituting the aforementioned dynamic weights into the weighted information content model, the susceptibility index of each grid cell along the line is calculated. Threat thresholds: rainfall > 50 mm, soil moisture content > 0.3%, slope displacement > 3 mm / day. Non-hazard samples: 200 grid cells that have not experienced landslides are randomly selected from the study area and used to construct ROC curves with 18 hazard points to determine the thresholds. Historical statistics. and Based on 18 case studies and grid statistics of the entire area, the calculation results show that the high-risk section is concentrated between K3+200 and K4+800, with a slope greater than 35° and a distance of less than 200m from the goaf. Based on the evaluation results, the mine reinforced the tower foundations in this section and added drainage ditches.

[0140] 1.6 Verification of Dynamic Update Effect

[0141] During the 2023 rainy season (June-August), the system automatically updated the weights every 7 days. Using a newly occurring landslide on August 12, 2023, as a validation point (this landslide was not included in the training), the susceptibility index predicted by the static weights (analytic hierarchy process weighting) ranked 12th at this point (out of 200 grid cells), while the dynamic weights (our method) ranked 4th. The AUC values ​​were calculated as follows: static weight model AUC = 0.72, dynamic weight model AUC = 0.81. See details... Figure 4 , Figure 4The comparison charts show the effects of dynamic weights and traditional static weights: Subplot (a) shows that the AUC of the dynamic weight model of this invention reaches 0.81, which is about 12.5% ​​higher than that of the traditional static AHP weight (0.72), significantly improving the identification accuracy of geological hazard susceptibility assessment; Subplot (b) shows the weight change curves, which intuitively demonstrate the environmental adaptability of the dynamic weights. During the heavy rainfall in August 2023, the weights of rainfall and soil moisture content factors continuously increased with the rainfall process, and the slope displacement rate weight responded synchronously, realizing the dynamic adjustment of the importance of disaster-causing factors under extreme weather conditions, effectively solving the technical defects of traditional static weights in adapting to environmental changes. The above results fully verify that the method of this invention has the advantages of spatial positioning accuracy, superior evaluation accuracy, and dynamic response timeliness, and can provide reliable technical support for geological hazard risk prevention and control and operation and maintenance decisions for coal mine power supply lines.

[0142] 1.7 Parameter Optimization Results

[0143] Grid search , , With a step size of 0.05, and using the AUC values ​​of historical cases as the optimization objective, the AUC values ​​under different combinations were calculated. The combination that ultimately maximizes the AUC was determined to be... , , The specific embodiment 1.4 used , , These are initial empirical values, which have been optimized and adjusted to the above combination. Subsequent actual deployments will use the optimized parameters.

[0144] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

[0145] It should also be noted that, in the embodiments of this application, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0146] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined in the embodiments of this application may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown in this application, but is to be accorded the widest scope consistent with the principles and novel features disclosed in the embodiments of this application.

Claims

1. A dynamic analysis method for the weights of geological hazard susceptibility evaluation factors in coal mine power supply lines, incorporating an attention mechanism, characterized in that... Includes the following steps: S1: For the area along the power supply line of the coal mine, select environmental factors related to the occurrence of geological disasters to form an evaluation factor set. The environmental factors include static factors, dynamic factors and mining-related factors. S2: Based on historical geological disaster cases, directed edges are established according to the temporal causal co-occurrence relationship between dynamic factors, and edge weights are calculated according to the frequency of occurrence of the directed edges in all historical cases to construct a weighted directed factor network among dynamic factors; wherein, static factors and mining impact factors are used as spatial constraints to modify the edge weights. S3: Obtain real-time monitoring data for each dynamic factor and define a dynamic attention weight for each dynamic factor; the dynamic attention weight is obtained by normalizing the weighted sum of three factors: its own state contribution, network propagation contribution, and global context contribution. S4: Based on the dynamic attention weight, the geological disaster susceptibility index of each evaluation unit along the power supply line is calculated using a weighted information volume model, and the susceptibility zone level is divided according to the susceptibility index; S5: Based on a preset time period or when the monitoring value of a dynamic factor exceeds a preset threshold, trigger step S3 to update the dynamic attention weight of each dynamic factor; at the same time, update the sampling impact factor on a monthly or yearly basis, and the sampling impact factor does not trigger the execution of step S3.

2. The method for dynamic analysis of the weights of geological hazard susceptibility evaluation factors for coal mine power supply lines based on the fusion attention mechanism as described in claim 1, characterized in that, In step S2, the directed edges and edge weights of the weighted directed factor network are determined as follows: a danger threshold is pre-set for each dynamic factor, and when the monitored value of a dynamic factor exceeds the threshold, it is determined to be in an abnormal state; if in historical geological disaster cases, the abnormal state of the first dynamic factor occurs before the abnormal state of the second dynamic factor, then a directed edge is established from the first dynamic factor to the second dynamic factor; the edge weight is equal to the number of times the second dynamic factor becomes abnormal after the first dynamic factor becomes abnormal in historical cases divided by the total number of times the first dynamic factor becomes abnormal.

3. The method for dynamic analysis of the weights of geological hazard susceptibility evaluation factors for coal mine power supply lines based on the attention mechanism as described in claim 2, is characterized in that... The method for correcting the edge weight of the static factor is as follows: for each directed edge, the proportion of the current static factor in a dangerous state in the historical events corresponding to each directed edge is counted, and the larger value between this proportion and the preset benchmark coefficient is taken as the correction coefficient. In the calculation of network propagation contribution, the weight of the directed edge is multiplied by the correction coefficient to obtain the corrected propagation strength; the historical event refers to the event in which the state of the second dynamic factor becomes abnormal after the state of the first dynamic factor becomes abnormal.

4. The method for dynamic analysis of the weights of geological hazard susceptibility evaluation factors for coal mine power supply lines based on the attention mechanism as described in claim 1, characterized in that, In step S3, the self-state contribution is calculated as follows: for each dynamic factor, the monitoring value within the most recent time window is collected, and the monitoring values ​​within the window are weighted and averaged, with the weight of recent data being greater than that of distant data, to obtain the current comprehensive state; The current comprehensive state is input into a logistic regression model obtained by fitting historical disaster data, and the output value is the contribution of its own state.

5. The method for dynamic analysis of the weights of geological hazard susceptibility evaluation factors for coal mine power supply lines based on the fusion attention mechanism according to claim 1, characterized in that, In step S3, the network propagation contribution is calculated as follows: For the current dynamic factor, obtain all the previous driving factor nodes pointing to it in the weighted directed factor network, multiply the dynamic attention weight, the corresponding directed edge weight, and the static factor correction coefficient of each previous driving factor at the previous time step, and then sum them up to obtain the network propagation contribution.

6. The method for dynamic analysis of the weights of geological hazard susceptibility evaluation factors for coal mine power supply lines based on the fusion attention mechanism as described in claim 1, characterized in that, In step S3, the global context contribution is calculated as follows: the time window monitoring sequence of each dynamic factor is mapped to an embedding vector through a linear embedding layer; multi-head self-attention is calculated on the embedding vector sequence of all dynamic factors; the position vector corresponding to each dynamic factor in the multi-head self-attention output sequence is taken, and then converted into a scalar through a fully connected layer, which is the global context contribution of the dynamic factor.

7. The method for dynamic analysis of the weights of geological hazard susceptibility evaluation factors for coal mine power supply lines based on the attention mechanism as described in claim 1, characterized in that, In step S3, the normalized calculation method for the dynamic attention weights is as follows: ； in, For the i-th dynamic factor The contribution value of its own state at the current time t; For the i-th dynamic factor The network propagation contribution value at current time t; For the i-th dynamic factor The global context contribution value at the current time t; , , These are the preset coefficients for the corresponding contribution items, and their sum is 1. m is the total number of dynamic factors, and j is the index of the dynamic factor.

8. The method for dynamic analysis of the weights of geological hazard susceptibility evaluation factors for coal mine power supply lines based on the fusion attention mechanism as described in claim 7, characterized in that, In step S4, the geological hazard susceptibility index Calculated using a weighted information content model: ； in, Dynamic factors in historical disasters The number of cases in a dangerous state, where N is the total number of historical geological disaster cases. To evaluate dynamic factors within the region The number of grid cells in a dangerous state, where S is the total number of grid cells in the evaluation area.

9. The method for dynamic analysis of the weights of geological hazard susceptibility evaluation factors for coal mine power supply lines based on the fusion attention mechanism according to claim 1, characterized in that, Step S5 specifically includes: Time-triggered update: Real-time monitoring data of each dynamic factor is re-collected every preset number of days and step S3 is executed; Event-triggered update: Step S3 is executed immediately when the monitoring value of any dynamic factor exceeds a preset threshold; Network structure update: The directed relationship between dynamic factors is re-statistically calculated every preset time period or when the number of new geological disaster cases exceeds a preset threshold, and the adjacency matrix and edge weights of the weighted directed factor network are updated.

10. The method for dynamic analysis of the weights of geological hazard susceptibility evaluation factors for coal mine power supply lines based on the fusion attention mechanism according to claim 1, characterized in that, The method further includes: a parameter optimization step: based on historical geological disaster cases, the data is divided into a training set and a validation set; multiple sets of different preset coefficient combinations are selected, and the susceptibility index of each disaster point in the validation set before its occurrence is calculated; the accuracy of disaster point identification is used as the evaluation index, and the preset coefficient combination that maximizes this index is selected as the optimal parameter.

Images